Daily Papers

Hallucination, where models generate fluent text unsupported by visual evidence, remains a major flaw in vision-language models and is particularly critical in sign language translation (SLT). In SLT, meaning depends on precise grounding in video, and gloss-free models are especially vulnerable because they map continuous signer movements directly into natural language without intermediate gloss supervision that serves as alignment. We argue that hallucinations arise when models rely on language priors rather than visual input. To capture this, we propose a token-level reliability measure that quantifies how much the decoder uses visual information. Our method combines feature-based sensitivity, which measures internal changes when video is masked, with counterfactual signals, which capture probability differences between clean and altered video inputs. These signals are aggregated into a sentence-level reliability score, providing a compact and interpretable measure of visual grounding. We evaluate the proposed measure on two SLT benchmarks (PHOENIX-2014T and CSL-Daily) with both gloss-based and gloss-free models. Our results show that reliability predicts hallucination rates, generalizes across datasets and architectures, and decreases under visual degradations. Beyond these quantitative trends, we also find that reliability distinguishes grounded tokens from guessed ones, allowing risk estimation without references; when combined with text-based signals (confidence, perplexity, or entropy), it further improves hallucination risk estimation. Qualitative analysis highlights why gloss-free models are more susceptible to hallucinations. Taken together, our findings establish reliability as a practical and reusable tool for diagnosing hallucinations in SLT, and lay the groundwork for more robust hallucination detection in multimodal generation.

Token-level reweighting is a simple yet effective mechanism for controlling supervised fine-tuning, but common indicators are largely one-dimensional: the ground-truth probability reflects downstream alignment, while token entropy reflects intrinsic uncertainty induced by the pre-training prior. Ignoring entropy can misidentify noisy or easily replaceable tokens as learning-critical, while ignoring probability fails to reflect target-specific alignment. RankTuner introduces a probability--entropy calibration signal, the Relative Rank Indicator, which compares the rank of the ground-truth token with its expected rank under the prediction distribution. The inverse indicator is used as a token-wise Relative Scale to reweight the fine-tuning objective, focusing updates on truly under-learned tokens without over-penalizing intrinsically uncertain positions. Experiments on multiple backbones show consistent improvements on mathematical reasoning benchmarks, transfer gains on out-of-distribution reasoning, and pre code generation performance over probability-only or entropy-only reweighting baselines.

Single-prompt accuracy is the dominant way to benchmark language models, but it can miss reliability failures that matter. We evaluate a 15-model open-weight corpus, with the main reliability analyses focused on 10 instruct models across five classification and reasoning benchmarks under five prompt variants each, measuring accuracy, token-probability calibration, verbal-confidence calibration, verbal parse rate, and prompt-perturbation spread for every (model x dataset x variant) cell. We find three broad results. First, evaluation design can materially change the conclusion. Switching Expected Calibration Error (ECE) token from a raw to a label-set-normalised definition changes per-cell calibration by a mean absolute 0.149. More strikingly, pairing a chain-of-thought prompt with a first-character evaluator on ARC-Challenge reduces apparent accuracy by 72-88% across all five primary models; two independent repair procedures recover 93.8% and 102.7% of the lost performance, indicating an evaluator-side rather than model-side failure. Second, confidence signals are fragile. On MMLU-Pro, every primary model verbally reports confidence substantially above both its accuracy and its token-probability confidence on the same rows, and verbal parse rate can collapse for a single model on a single prompt variant. Third, prompt robustness does not track parameter count reliably. Across 10 instruct models, the correlation between model size and prompt-perturbation spread ranges from -0.244 to 0.474 across benchmarks. Taken together, these results show that reliability conclusions for small language models depend not only on the model being evaluated, but also on the evaluation pipeline used to measure it. We argue that calibration definitions, evaluator logic, verbal parseability, and prompt robustness should be reported explicitly when making reliability claims.

The increasing capability and widespread usage of large language models (LLMs) highlight the desirability of automatic detection of LLM-generated text. Zero-shot detectors, due to their training-free nature, have received considerable attention and notable success. In this paper, we identify a new feature, token cohesiveness, that is useful for zero-shot detection, and we demonstrate that LLM-generated text tends to exhibit higher token cohesiveness than human-written text. Based on this observation, we devise TOCSIN, a generic dual-channel detection paradigm that uses token cohesiveness as a plug-and-play module to improve existing zero-shot detectors. To calculate token cohesiveness, TOCSIN only requires a few rounds of random token deletion and semantic difference measurement, making it particularly suitable for a practical black-box setting where the source model used for generation is not accessible. Extensive experiments with four state-of-the-art base detectors on various datasets, source models, and evaluation settings demonstrate the effectiveness and generality of the proposed approach. Code available at: https://github.com/Shixuan-Ma/TOCSIN.

Despite significant progress in text anomaly detection for web applications such as spam filtering and fake news detection, existing methods are fundamentally limited to document-level analysis, unable to identify which specific parts of a text are anomalous. We introduce token-level anomaly detection, a novel paradigm that enables fine-grained localization of anomalies within text. We formally define text anomalies at both document and token-levels, and propose a unified detection framework that operates across multiple levels. To facilitate research in this direction, we collect and annotate three benchmark datasets spanning spam, reviews and grammar errors with token-level labels. Experimental results demonstrate that our framework get better performance than other 6 baselines, opening new possibilities for precise anomaly localization in text. All the codes and data are publicly available on https://github.com/charles-cao/TokenCore.

Neural metrics for machine translation evaluation, such as COMET, exhibit significant improvements in their correlation with human judgments, as compared to traditional metrics based on lexical overlap, such as BLEU. Yet, neural metrics are, to a great extent, "black boxes" returning a single sentence-level score without transparency about the decision-making process. In this work, we develop and compare several neural explainability methods and demonstrate their effectiveness for interpreting state-of-the-art fine-tuned neural metrics. Our study reveals that these metrics leverage token-level information that can be directly attributed to translation errors, as assessed through comparison of token-level neural saliency maps with Multidimensional Quality Metrics (MQM) annotations and with synthetically-generated critical translation errors. To ease future research, we release our code at: https://github.com/Unbabel/COMET/tree/explainable-metrics.

In the development of Large Language Models (LLMs), considerable attention has been given to the quality of training datasets. However, the role of tokenizers in the LLM training pipeline, particularly for multilingual models, has received less focus. The quality of tokenization can significantly impact a model's ability to handle diverse languages effectively. We introduce Qtok, a tool designed to assess tokenizer quality with a specific emphasis on their performance in multilingual contexts. Our research proposes a set of metrics for evaluating tokenizer quality, including measures of language coverage, token completeness, and distribution across languages and linguistic categories. Qtok applies these metrics to evaluate 13 distinct tokenizers from 58 publicly available models, analyzing their output across different linguistic contexts. Our analysis revealed significant variations in token distribution across languages and categories, highlighting potential biases and areas for improvement in current tokenization strategies. This research contributes to the field of tokenizer evaluation within multilingual LLM development by providing a systematic approach to assessing tokenizer quality. Our findings highlight the critical role of tokenization in multilingual LLM capability. The Qtok tool and our analysis methodology offer practical means for researchers to evaluate and improve tokenization strategies for multilingual applications. We offer a method to compare tokenizer quality across these metrics, which may be useful when selecting or adjusting tokenizers for specific multilingual LLM applications.

Practicing conversations with large language models (LLMs) presents a promising alternative to traditional in-person language learning. However, most LLMs generate text at a near-native level of complexity, making them ill-suited for beginner learners (CEFR: A1-A2). In this paper, we investigate whether controllable generation techniques -- specifically modular methods that do not require model fine-tuning -- can adapt LLM outputs to better support absolute beginners. We evaluate these methods through both automatic metrics and a user study with university-level learners of Japanese. Our findings show that while prompting alone fails to control output difficulty, the use of future discriminators (Yang and Klein, 2021) significantly improves output comprehensibility (from 40.4\% to 84.3\%). We further introduce a novel token-level evaluation metric, Token Miss Rate (TMR), that quantifies the proportion of incomprehensible tokens per utterance and correlates strongly with human judgments. To support future research in AI-assisted language learning, we release our code, models, annotation tools, and dataset.

Safe deployment of large language models (LLMs) may benefit from a reliable method for assessing their generated content to determine when to abstain or to selectively generate. While likelihood-based metrics such as perplexity are widely employed, recent research has demonstrated the limitations of using sequence-level probability estimates given by LLMs as reliable indicators of generation quality. Conversely, LLMs have demonstrated strong calibration at the token level, particularly when it comes to choosing correct answers in multiple-choice questions or evaluating true/false statements. In this work, we reformulate open-ended generation tasks into token-level prediction tasks, and leverage LLMs' superior calibration at the token level. We instruct an LLM to self-evaluate its answers, employing either a multi-way comparison or a point-wise evaluation approach, with the option to include a ``None of the above'' option to express the model's uncertainty explicitly. We benchmark a range of scoring methods based on self-evaluation and evaluate their performance in selective generation using TruthfulQA and TL;DR. Through experiments with PaLM-2 and GPT-3, we demonstrate that self-evaluation based scores not only improve accuracy, but also correlate better with the overall quality of generated content.

Frontier LLMs are increasingly utilised across academia, society and industry. A commonly used unit for comparing models, their inputs and outputs, and estimating inference pricing is the token. In general, tokens are used as a stable currency, assumed to be broadly consistent across tokenizers and contexts, enabling direct comparisons. However, tokenization varies significantly across models and domains of text, making naive interpretation of token counts problematic. We quantify this variation by providing a comprehensive empirical analysis of tokenization, exploring the compression of sequences to tokens across different distributions of textual data. Our analysis challenges commonly held heuristics about token lengths, finding them to be overly simplistic. We hope the insights of our study add clarity and intuition toward tokenization in contemporary LLMs.

Large language models (LLMs) reach high accuracy in mathematical reasoning, but individual traces on the same problem diverge; some arrive at the correct answer while others fail. Prior work analyzes failure at the step, chunk, or sentence level, or at tokens where failure has already occurred. Neither identifies the precise token that triggers the shift toward failure. We introduce the cliff token, a token where the token-wise potential drops significantly under an adaptive threshold that scales with the local token-wise potential, based on a one-sided two-proportion z-test. Across seven models and three mathematical reasoning benchmarks (GSM1K, MATH500, AIME 2025), cliff tokens act as failure triggers; deleting the first cliff token and resampling recovers pass@64 to 1.0, while keeping it limits recovery to between 0.71 and 1.00. We further introduce a cliff taxonomy of deterministic, uncertain, and sampled-off cliffs, defined by greedy choice and token entropy. Each type has distinct probabilistic characteristics, and the taxonomy generalizes across model scales. Finally, we validate the taxonomy via single-token preference optimization at cliff positions (Cliff-DPO). Trained on GSM8K, Cliff-DPO improves accuracy across benchmarks by up to +6.6. Optimizing at uncertain and sampled-off cliffs improves reasoning, while deterministic cliffs do not.

Large Language Models (LLMs) have demonstrated impressive generative capabilities across diverse tasks but remain susceptible to hallucinations, confidently generated yet factually incorrect outputs. We introduce a reference-free, token-level hallucination detection framework that leverages the variance in token log-probabilities across multiple stochastic generations. Unlike prior methods that require ground-truth references or sentence-level verification, our approach is model-agnostic, interpretable, and suited for real-time or post-hoc analysis. We evaluate our method on unanswerable question prompts from the SQuAD v2 dataset and benchmark across three autoregressive models of varying scales: GPT-Neo 125M, Falcon 1B, and Mistral 7B. Through both quantitative metrics and visual diagnostics, we show that token-level variance reliably highlights instability in model outputs and correlates with hallucination patterns. Our framework is lightweight, reproducible, and adaptable to multiple domains, offering a valuable diagnostic tool for analyzing generative reliability in LLMs.

The prefill stage in long-context LLM inference remains a computational bottleneck. Recent token-ranking heuristics accelerate inference by selectively processing a subset of semantically relevant tokens. However, existing methods suffer from unstable token importance estimation, often varying between layers. Evaluating token-ranking quality independently from heuristic-specific architectures is challenging. To address this, we introduce an Answer-Informed Oracle, which defines ground-truth token importance by measuring attention from generated answers back to the prompt. This oracle reveals that existing heuristics exhibit high variance across layers: rankings can degrade sharply at specific layers, a failure mode invisible to end-to-end benchmarks. The diagnosis suggests a simple fix: aggregate scores across layers rather than relying on any single one. We implement this as Cross-Layer Attention Aggregation (CLAA), which closes the gap to the oracle upper bound and reduces Time-to-First-Token (TTFT) by up to 39\% compared to the Full KV Cache baseline.

Large language models (LLMs) are increasingly utilized for machine translation, yet their predictions often exhibit uncertainties that hinder interpretability and user trust. Effectively visualizing these uncertainties can enhance the usability of LLM outputs, particularly in contexts where translation accuracy is critical. This paper addresses two primary objectives: (1) providing users with token-level insights into model confidence and (2) developing a web-based visualization tool to quantify and represent translation uncertainties. To achieve these goals, we utilized the T5 model with the WMT19 dataset for translation tasks and evaluated translation quality using established metrics such as BLEU, METEOR, and ROUGE. We introduced three novel uncertainty quantification (UQ) metrics: (1) the geometric mean of token probabilities, (2) the arithmetic mean of token probabilities, and (3) the arithmetic mean of the kurtosis of token distributions. These metrics provide a simple yet effective framework for evaluating translation performance. Our analysis revealed a linear relationship between the traditional evaluation metrics and our UQ metrics, demonstrating the validity of our approach. Additionally, we developed an interactive web-based visualization that uses a color gradient to represent token confidence. This tool offers users a clear and intuitive understanding of translation quality while providing valuable insights into model performance. Overall, we show that our UQ metrics and visualization are both robust and interpretable, offering practical tools for evaluating and accessing machine translation systems.

A common approach to explain NLP models, is to use importance measures that express which tokens are important for a prediction. Unfortunately, such explanations are often wrong despite being persuasive. Therefore, it is essential to measure their faithfulness. One such metric is if tokens are truly important, then masking them should result in worse model performance. However, token masking introduces out-of-distribution issues and existing solutions are computationally expensive and employ proxy-models. Furthermore, other metrics are very limited in scope. In this work, we propose an inherently faithfulness measurable model that addresses these challenges. This is achieved by using a novel fine-tuning method that incorporates masking, such that masking tokens become in-distribution by design. This differs from existing approaches, which are completely model-agnostic but are inapplicable in practice. We demonstrate the generality of our approach by applying it to various tasks and validate it using statistical in-distribution tests. Additionally, because masking is in-distribution, importance measures which themselves use masking become more faithful, thus our model becomes more explainable.

LLMs are highly sensitive to prompt phrasing, yet standard benchmarks typically report performance using a single prompt, raising concerns about the reliability of such evaluations. In this work, we argue for a stochastic method of moments evaluation over the space of meaning-preserving prompt perturbations. We introduce a formal definition of reliable evaluation that accounts for prompt sensitivity, and suggest ReliableEval - a method for estimating the number of prompt resamplings needed to obtain meaningful results. Using our framework, we stochastically evaluate five frontier LLMs and find that even top-performing models like GPT-4o and Claude-3.7-Sonnet exhibit substantial prompt sensitivity. Our approach is model-, task-, and metric-agnostic, offering a recipe for meaningful and robust LLM evaluation.

Recent advancements in large language models (LLMs) have significantly enhanced their reasoning capabilities. However, they continue to struggle with basic character-level tasks, such as counting letters in words, a problem rooted in their tokenization process. While existing benchmarks have highlighted this weakness through basic character operations, such failures are often dismissed due to lacking practical relevance. Yet, many real-world applications, such as navigating text-based maps or interpreting structured tables, rely heavily on precise sub-token understanding. In this regard, we introduce SubTokenTest, a comprehensive benchmark that assesses sub-token understanding through practical, utility-driven tasks. Our benchmark includes ten tasks across four domains and isolates tokenization-related failures by decoupling performance from complex reasoning. We provide a comprehensive evaluation of nine advanced LLMs. Additionally, we investigate the impact of test-time scaling on sub-token reasoning and explore how character-level information is encoded within the hidden states.

Evaluating the reliability of news sources is a routine task for journalists and organizations committed to acquiring and disseminating accurate information. Recent research has shown that predicting sources' reliability represents an important first-prior step in addressing additional challenges such as fake news detection and fact-checking. In this paper, we introduce a novel approach for source reliability estimation that leverages reinforcement learning strategies for estimating the reliability degree of news sources. Contrary to previous research, our proposed approach models the problem as the estimation of a reliability degree, and not a reliability label, based on how all the news media sources interact with each other on the Web. We validated the effectiveness of our method on a news media reliability dataset that is an order of magnitude larger than comparable existing datasets. Results show that the estimated reliability degrees strongly correlates with journalists-provided scores (Spearman=0.80) and can effectively predict reliability labels (macro-avg. F_1 score=81.05). We release our implementation and dataset, aiming to provide a valuable resource for the NLP community working on information verification.

Advanced LLMs have achieved near-ceiling instruction-following accuracy on benchmarks such as IFEval. However, these impressive scores do not necessarily translate to reliable services in real-world use, where users often vary their phrasing, contextual framing, and task formulations. In this paper, we study nuance-oriented reliability: whether models exhibit consistent competence across cousin prompts that convey analogous user intents but with subtle nuances. To quantify this, we introduce a new metric, reliable@k, and develop an automated pipeline that generates high-quality cousin prompts via data augmentation. Building upon this, we construct IFEval++ for systematic evaluation. Across 20 proprietary and 26 open-source LLMs, we find that current models exhibit substantial insufficiency in nuance-oriented reliability -- their performance can drop by up to 61.8% with nuanced prompt modifications. What's more, we characterize it and explore three potential improvement recipes. Our findings highlight nuance-oriented reliability as a crucial yet underexplored next step toward more dependable and trustworthy LLM behavior. Our code and benchmark are accessible: https://github.com/jianshuod/IFEval-pp.

The problem of pre-training data detection for large language models (LLMs) has received growing attention due to its implications in critical issues like copyright violation and test data contamination. The current state-of-the-art approach, Min-K%, measures the raw token probability which we argue may not be the most informative signal. Instead, we propose Min-K%++ to normalize the token probability with statistics of the categorical distribution over the whole vocabulary, which accurately reflects the relative likelihood of the target token compared with other candidate tokens in the vocabulary. Theoretically, we back up our method by showing that the statistic it estimates is explicitly optimized during LLM training, thus serving as a reliable indicator for detecting training data. Empirically, on the WikiMIA benchmark, Min-K%++ outperforms the SOTA Min-K% by 6.2% to 10.5% in detection AUROC averaged over five models. On the more challenging MIMIR benchmark, Min-K%++ consistently improves upon Min-K% and performs on par with reference-based method, despite not requiring an extra reference model.

Safety-aligned language models often refuse prompts that are actually harmless. Current evaluations mostly report global rates such as false rejection or compliance. These scores treat each prompt alone and miss local inconsistency, where a model accepts one phrasing of an intent but rejects a close paraphrase. This gap limits diagnosis and tuning. We introduce "semantic confusion," a failure mode that captures such local inconsistency, and a framework to measure it. We build ParaGuard, a 10k-prompt corpus of controlled paraphrase clusters that hold intent fixed while varying surface form. We then propose three model-agnostic metrics at the token level: Confusion Index, Confusion Rate, and Confusion Depth. These metrics compare each refusal to its nearest accepted neighbors and use token embeddings, next-token probabilities, and perplexity signals. Experiments across diverse model families and deployment guards show that global false-rejection rate hides critical structure. Our metrics reveal globally unstable boundaries in some settings, localized pockets of inconsistency in others, and cases where stricter refusal does not increase inconsistency. We also show how confusion-aware auditing separates how often a system refuses from how sensibly it refuses. This gives developers a practical signal to reduce false refusals while preserving safety.

Question Answering (QA) in NLP is the task of finding answers to a query within a relevant context retrieved by a retrieval system. Yet, the mix of relevant and irrelevant information in these contexts can hinder performance enhancements in QA tasks. To address this, we introduce a context filtering approach that removes non-essential details, summarizing crucial content through Reward Modeling. This method emphasizes keeping vital data while omitting the extraneous during summarization model training. We offer a framework for developing efficient QA models by discerning useful information from dataset pairs, bypassing the need for costly human evaluation. Furthermore, we show that our approach can significantly outperform the baseline, as evidenced by a 6.8-fold increase in the EM Per Token (EPT) metric, which we propose as a measure of token efficiency, indicating a notable token-efficiency boost for low-resource settings.

Subword tokenization has become the de-facto standard for tokenization, although comparative evaluations of subword vocabulary quality across languages are scarce. Existing evaluation studies focus on the effect of a tokenization algorithm on the performance in downstream tasks, or on engineering criteria such as the compression rate. We present a new evaluation paradigm that focuses on the cognitive plausibility of subword tokenization. We analyze the correlation of the tokenizer output with the response time and accuracy of human performance on a lexical decision task. We compare three tokenization algorithms across several languages and vocabulary sizes. Our results indicate that the UnigramLM algorithm yields less cognitively plausible tokenization behavior and a worse coverage of derivational morphemes, in contrast with prior work.

Large language models (LLMs) have demonstrated impressive performance on several tasks and are increasingly deployed in real-world applications. However, especially in high-stakes settings, it becomes vital to know when the output of an LLM may be unreliable. Depending on whether an answer is trustworthy, a system can then choose to route the question to another expert, or otherwise fall back on a safe default behavior. In this work, we study the extent to which LLMs can reliably indicate confidence in their answers, and how this notion of confidence can translate into downstream accuracy gains. We propose Self-REF, a lightweight training strategy to teach LLMs to express confidence in whether their answers are correct in a reliable manner. Self-REF introduces confidence tokens into the LLM, from which a confidence score can be extracted. Compared to conventional approaches such as verbalizing confidence and examining token probabilities, we demonstrate empirically that confidence tokens show significant improvements in downstream routing and rejection learning tasks.

LLM judges are used to reduce the need for costly human labor in evaluating open-ended text generation. However, the reliability of these judges depends critically on their alignment with human raters -- a property that itself depends on costly human annotations. In this work, we develop a method (Metric Match) for estimating correlation-based reliability metrics of LLM judges from limited annotations. Metric Match selects a subset of samples for human annotation such that the subset matches the population reliability metric with respect to acquired synthetic labels. We empirically show that Metric Match achieves a win-rate of 0.838 against random subset selection across four different correlation metrics and 15 datasets, with an 18.7% decrease in average estimation error and reduces annotation needs by 32.5%. We provide a cost model and highlight a medical case study where our method saves $1,041.67 compared to random selection for expert annotation. Further, we shift our task from reliability estimation to reliability classification of whether a given judge is above a deployment threshold, outperforming random selection with Metric Match. All project code is publicly available, and we additionally provide an installable package for ease of use.

Text-to-image (T2I) models have significantly advanced in producing high-quality images. However, such models have the ability to generate images containing not-safe-for-work (NSFW) content, such as pornography, violence, political content, and discrimination. To mitigate the risk of generating NSFW content, refusal mechanisms, i.e., safety checkers, have been developed to check potential NSFW content. Adversarial prompting techniques have been developed to evaluate the robustness of the refusal mechanisms. The key challenge remains to subtly modify the prompt in a way that preserves its sensitive nature while bypassing the refusal mechanisms. In this paper, we introduce TokenProber, a method designed for sensitivity-aware differential testing, aimed at evaluating the robustness of the refusal mechanisms in T2I models by generating adversarial prompts. Our approach is based on the key observation that adversarial prompts often succeed by exploiting discrepancies in how T2I models and safety checkers interpret sensitive content. Thus, we conduct a fine-grained analysis of the impact of specific words within prompts, distinguishing between dirty words that are essential for NSFW content generation and discrepant words that highlight the different sensitivity assessments between T2I models and safety checkers. Through the sensitivity-aware mutation, TokenProber generates adversarial prompts, striking a balance between maintaining NSFW content generation and evading detection. Our evaluation of TokenProber against 5 safety checkers on 3 popular T2I models, using 324 NSFW prompts, demonstrates its superior effectiveness in bypassing safety filters compared to existing methods (e.g., 54%+ increase on average), highlighting TokenProber's ability to uncover robustness issues in the existing refusal mechanisms.

Recent advancements in the field of natural language generation have facilitated the use of large language models to assess the quality of generated text. Although these models have shown promising results in tasks such as machine translation and summarization, their applicability in code generation tasks remains limited without human involvement. The complexity of programming concepts required for such tasks makes it difficult to develop evaluation metrics that align with human judgment. Token-matching-based metrics, such as BLEU, have demonstrated weak correlations with human practitioners in code generation tasks. Moreover, the utilization of human-written test suites to evaluate functional correctness can be challenging in domains with low resources. To overcome these obstacles, we propose a new evaluation framework based on the GPT-3.5 (GPT-3.5-turbo), for code generation assessments. Our framework addresses the limitations of existing approaches by achieving superior correlations with functional correctness and human preferences, without the need for test oracles or references. We evaluate the efficacy of our framework on two different tasks and four programming languages, comparing its performance with the state-of-the-art CodeBERTScore metric, which relies on a pre-trained model. Our results demonstrate that our framework surpasses CodeBERTScore, delivering high levels of accuracy and consistency across various programming languages and tasks. We also make our evaluation framework and datasets available to the public at https://github.com/terryyz/llm-code-eval, encouraging further research in the evaluation of code generation.

Large Language Models (LLMs) are trained on extensive web corpora, which enable them to understand and generate human-like text. However, this training process also results in inherent biases within the models. These biases arise from web data's diverse and often uncurated nature, containing various stereotypes and prejudices. Previous works on debiasing models rely on benchmark datasets to measure their method's performance. However, these datasets suffer from several pitfalls due to the highly subjective understanding of bias, highlighting a critical need for contextual exploration. We propose understanding the context of inputs by considering the diverse situations in which they may arise. Our contribution is two-fold: (i) we augment 2,291 stereotyped statements from two existing bias-benchmark datasets with points for adding context; (ii) we develop the Context-Oriented Bias Indicator and Assessment Score (COBIAS) to assess a statement's contextual reliability in measuring bias. Our metric aligns with human judgment on contextual reliability of statements (Spearman's rho = 0.65, p = 3.4 * 10^{-60}) and can be used to create reliable datasets, which would assist bias mitigation works.

Reinforcement learning (RL) has been proven to be an effective and robust method for training neural machine translation systems, especially when paired with powerful reward models that accurately assess translation quality. However, most research has focused on RL methods that use sentence-level feedback, leading to inefficient learning signals due to the reward sparsity problem -- the model receives a single score for the entire sentence. To address this, we propose a novel approach that leverages fine-grained, token-level quality assessments along with error severity levels using RL methods. Specifically, we use xCOMET, a state-of-the-art quality estimation system, as our token-level reward model. We conduct experiments on small and large translation datasets with standard encoder-decoder and large language models-based machine translation systems, comparing the impact of sentence-level versus fine-grained reward signals on translation quality. Our results show that training with token-level rewards improves translation quality across language pairs over baselines according to both automatic and human evaluation. Furthermore, token-level reward optimization improves training stability, evidenced by a steady increase in mean rewards over training epochs.

Scaling laws in language modeling traditionally quantify training loss as a function of dataset size and model parameters, providing compute-optimal estimates but often neglecting the impact of data quality on model generalization. In this paper, we extend the conventional understanding of scaling law by offering a microscopic view of data quality within the original formulation -- effective training tokens -- which we posit to be a critical determinant of performance for parameter-constrained language models. Specifically, we formulate the proposed term of effective training tokens to be a combination of two readily-computed indicators of text: (i) text diversity and (ii) syntheticity as measured by a teacher model. We pretrained over 200 models of 25M to 1.5B parameters on a diverse set of sampled, synthetic data, and estimated the constants that relate text quality, model size, training tokens, and eight reasoning task accuracy scores. We demonstrated the estimated constants yield +0.83 Pearson correlation with true accuracies, and analyzed it in scenarios involving widely-used data techniques such as data sampling and synthesis which aim to improve data quality.

Estimating the difficulty of input questions as perceived by large language models (LLMs) is essential for accurate performance evaluation and adaptive inference. Existing methods typically rely on repeated response sampling, auxiliary models, or fine-tuning the target model itself, which may incur substantial computational costs or compromise generality. In this paper, we propose a novel approach for difficulty estimation that leverages only the hidden representations produced by the target LLM. We model the token-level generation process as a Markov chain and define a value function to estimate the expected output quality given any hidden state. This allows for efficient and accurate difficulty estimation based solely on the initial hidden state, without generating any output tokens. Extensive experiments across both textual and multimodal tasks demonstrate that our method consistently outperforms existing baselines in difficulty estimation. Moreover, we apply our difficulty estimates to guide adaptive reasoning strategies, including Self-Consistency, Best-of-N, and Self-Refine, achieving higher inference efficiency with fewer generated tokens.

While automatic metrics drive progress in Machine Translation (MT) and Text Summarization (TS), existing metrics have been developed and validated almost exclusively for English and other high-resource languages. This narrow focus leaves Indian languages, spoken by over 1.5 billion people, largely overlooked, casting doubt on the universality of current evaluation practices. To address this gap, we introduce ITEM, a large-scale benchmark that systematically evaluates the alignment of 26 automatic metrics with human judgments across six major Indian languages, enriched with fine-grained annotations. Our extensive evaluation, covering agreement with human judgments, sensitivity to outliers, language-specific reliability, inter-metric correlations, and resilience to controlled perturbations, reveals four central findings: (1) LLM-based evaluators show the strongest alignment with human judgments at both segment and system levels; (2) outliers exert a significant impact on metric-human agreement; (3) in TS, metrics are more effective at capturing content fidelity, whereas in MT, they better reflect fluency; and (4) metrics differ in their robustness and sensitivity when subjected to diverse perturbations. Collectively, these findings offer critical guidance for advancing metric design and evaluation in Indian languages.

Tokenization is a pivotal design choice for neural language modeling in morphologically rich languages (MRLs) such as Turkish, where productive agglutination challenges both vocabulary efficiency and morphological fidelity. Prior studies have explored tokenizer families and vocabulary sizes but typically (i) vary vocabulary without systematically controlling the tokenizer's training corpus, (ii) provide limited intrinsic diagnostics, and (iii) evaluate a narrow slice of downstream tasks. We present the first comprehensive, principled study of Turkish subword tokenization; a "subwords manifest", that jointly varies vocabulary size and tokenizer training corpus size (data and vocabulary coupling), compares multiple tokenizer families under matched parameter budgets (WordPiece, morphology level, and character baselines), and evaluates across semantic (NLI, STS, sentiment analysis, NER), syntactic (POS, dependency parsing), and morphology-sensitive probes. To explain why tokenizers succeed or fail, we introduce a morphology-aware diagnostic toolkit that goes beyond coarse aggregates to boundary-level micro/macro F1, decoupled lemma atomicity vs. surface boundary hits, over/under-segmentation indices, character/word edit distances (CER/WER), continuation rates, and affix-type coverage and token-level atomicity. Our contributions are fourfold: (i) a systematic investigation of the vocabulary-corpus-success triad; (ii) a unified, morphology-aware evaluation framework linking intrinsic diagnostics to extrinsic outcomes; (iii) controlled comparisons identifying when character-level and morphology-level tokenization pay off; and (iv) an open-source release of evaluation code, tokenizer pipelines, and models. As the first work of its kind, this "subwords manifest" delivers actionable guidance for building effective tokenizers in MRLs and establishes a reproducible foundation for future research.

Tokenization is a fundamental preprocessing step in Natural Language Processing (NLP), significantly impacting the capability of large language models (LLMs) to capture linguistic and semantic nuances. This study introduces a novel evaluation framework addressing tokenization challenges specific to morphologically-rich and low-resource languages such as Turkish. Utilizing the Turkish MMLU (TR-MMLU) dataset, comprising 6,200 multiple-choice questions from the Turkish education system, we assessed tokenizers based on vocabulary size, token count, processing time, language-specific token percentages (\%TR), and token purity (\%Pure). These newly proposed metrics measure how effectively tokenizers preserve linguistic structures. Our analysis reveals that language-specific token percentages exhibit a stronger correlation with downstream performance (e.g., MMLU scores) than token purity. Furthermore, increasing model parameters alone does not necessarily enhance linguistic performance, underscoring the importance of tailored, language-specific tokenization methods. The proposed framework establishes robust and practical tokenization standards for morphologically complex languages.

This paper reports on the ReINTEL Shared Task for Responsible Information Identification on social network sites, which is hosted at the seventh annual workshop on Vietnamese Language and Speech Processing (VLSP 2020). Given a piece of news with respective textual, visual content and metadata, participants are required to classify whether the news is `reliable' or `unreliable'. In order to generate a fair benchmark, we introduce a novel human-annotated dataset of over 10,000 news collected from a social network in Vietnam. All models will be evaluated in terms of AUC-ROC score, a typical evaluation metric for classification. The competition was run on the Codalab platform. Within two months, the challenge has attracted over 60 participants and recorded nearly 1,000 submission entries.

Despite its simplicity and efficacy, the high token expenditure of self-consistency can limit its practical utility. Here we investigate if self-consistency can be made more token-efficient for long chain-of-thought reasoning tasks, while preserving its parallelism, through early hypothesis pruning. Concretely, we generate all solutions in parallel, but periodically prune intermediate hypotheses that are deemed unnecessary based on two lightweight indicators: (a) the model's own confidence in individual hypotheses, and (b) lexical coverage of all current hypotheses by candidate subsets that are under consideration for continued retention. We design a fast weighted set cover algorithm that utilizes the two indicators; our evaluation of five LLMs on three math benchmarks shows that this method can improve token efficiency for all models, by 10-35% in many cases.

Recent progress in deep research systems has been impressive, but evaluation still lags behind real user needs. Existing benchmarks predominantly assess final reports using fixed rubrics, failing to evaluate the underlying research process. Most also offer limited multimodal coverage, rely on synthetic tasks that do not reflect real-world query complexity, and cannot be refreshed as knowledge evolves. To address these gaps, we introduce MiroEval, a benchmark and evaluation framework for deep research systems. The benchmark comprises 100 tasks (70 text-only, 30 multimodal), all grounded in real user needs and constructed via a dual-path pipeline that supports periodic updates, enabling a live and evolving setting. The proposed evaluation suite assesses deep research systems along three complementary dimensions: adaptive synthesis quality evaluation with task-specific rubrics, agentic factuality verification via active retrieval and reasoning over both web sources and multimodal attachments, and process-centric evaluation audits how the system searches, reasons, and refines throughout its investigation. Evaluation across 13 systems yields three principal findings: the three evaluation dimensions capture complementary aspects of system capability, with each revealing distinct strengths and weaknesses across systems; process quality serves as a reliable predictor of overall outcome while revealing weaknesses invisible to output-level metrics; and multimodal tasks pose substantially greater challenges, with most systems declining by 3 to 10 points. The MiroThinker series achieves the most balanced performance, with MiroThinker-H1 ranking the highest overall in both settings. Human verification and robustness results confirm the reliability of the benchmark and evaluation framework. MiroEval provides a holistic diagnostic tool for the next generation of deep research agents.

Developing large language models is expensive and involves making decisions with small experiments, typically by evaluating on large, multi-task evaluation suites. In this work, we analyze specific properties which make a benchmark more reliable for such decisions, and interventions to design higher-quality evaluation benchmarks. We introduce two key metrics that show differences in current benchmarks: signal, a benchmark's ability to separate better models from worse models, and noise, a benchmark's sensitivity to random variability between training steps. We demonstrate that benchmarks with a better signal-to-noise ratio are more reliable when making decisions at small scale, and those with less noise have lower scaling law prediction error. These results suggest that improving signal or noise will lead to more useful benchmarks, so we introduce three interventions designed to directly affect signal or noise. For example, we propose that switching to a metric that has better signal and noise (e.g., perplexity rather than accuracy) leads to better reliability and improved scaling law error. We also find that filtering noisy subtasks, to improve an aggregate signal-to-noise ratio, leads to more reliable multi-task evaluations. We also find that averaging the output of a model's intermediate checkpoints to reduce noise leads to consistent improvements. We conclude by recommending that those creating new benchmarks, or selecting which existing benchmarks to use, aim for high signal and low noise. We use 30 benchmarks for these experiments, and 375 open-weight language models from 60M to 32B parameters, resulting in a new, publicly available dataset of 900K evaluation benchmark results, totaling 200M instances.

While tokenization is a key step in language modeling, with effects on model training and performance, it remains unclear how to effectively evaluate tokenizer quality. One proposed dimension of tokenizer quality is the extent to which tokenizers preserve linguistically meaningful subwords, aligning token boundaries with morphological boundaries within a word. We expand MorphScore (Arnett & Bergen, 2025), which previously covered 22 languages, to support a total of 70 languages. The updated MorphScore offers more flexibility in evaluation and addresses some of the limitations of the original version. We then correlate our alignment scores with downstream task performance for five pre-trained languages models on seven tasks, with at least one task in each of the languages in our sample. We find that morphological alignment does not explain very much variance in model performance, suggesting that morphological alignment alone does not measure dimensions of tokenization quality relevant to model performance.

Verbal confidence elicitation is widely used to extract uncertainty estimates from LLMs. We tested whether seven instruction-tuned open-weight models (3-9B parameters, four families) produce verbalised confidence that meets minimal validity criteria for item-level Type-2 discrimination under minimal numeric elicitation with greedy decoding. In a pre-registered study (OSF: osf.io/azbvx), 524 TriviaQA items were administered under numeric (0-100) and categorical (10-class) elicitation to eight models at Q5_K_M quantisation on consumer hardware, yielding 8,384 deterministic trials. A psychometric validity screen was applied to each model-format cell. All seven instruct models were classified Invalid on numeric confidence (H2 confirmed, 7/7 vs. predicted >=4/7), with a mean ceiling rate of 91.7% (H1 confirmed). Categorical elicitation did not rescue validity. Instead, it disrupted task performance in six of seven models, producing accuracy below 5% (H4 not confirmed). Token-level logprobability did not usefully predict verbalised confidence under the observed variance regime (H5 confirmed, mean cross-validated R^2 < 0.01). Within the reasoning-distilled model, reasoning-trace length showed a strong negative partial correlation with confidence (rho = -0.36, p < .001), consistent with the Reasoning Contamination Effect. These results do not imply that internal uncertainty representations are absent. They show that minimal verbal elicitation fails to preserve internal signals at the output interface in this model-size regime. Psychometric screening should precede any downstream use of such signals.

Generating unbiased summaries in real-world settings such as political perspective summarization remains a crucial application of Large Language Models (LLMs). Yet, existing evaluation frameworks rely on traditional metrics for measuring key attributes such as coverage and faithfulness without verifying their applicability, and efforts to develop improved summarizers are still nascent. We address these gaps by (1) identifying reliable metrics for measuring perspective summary quality, and (2) investigating the efficacy of LLM-based methods beyond zero-shot inference. Namely, we build a test set for benchmarking metric reliability using human annotations and show that traditional metrics underperform compared to language model-based metrics, which prove to be strong evaluators. Using these metrics, we show that reranking-based methods yield strong results, and preference tuning with synthetically generated and reranking-labeled data further boosts performance. Our findings aim to contribute to the reliable evaluation and development of perspective summarization methods.

In Natural Language Processing (NLP), the Elo rating system, originally designed for ranking players in dynamic games such as chess, is increasingly being used to evaluate Large Language Models (LLMs) through "A vs B" paired comparisons. However, while popular, the system's suitability for assessing entities with constant skill levels, such as LLMs, remains relatively unexplored. We study two fundamental axioms that evaluation methods should adhere to: reliability and transitivity. We conduct extensive evaluation of Elo behaviour, illustrating that individual Elo computations exhibit volatility and delving into the impact of varying the Elo rating system's hyperparameters. We show that these axioms are not always satisfied raising questions about the reliability of current comparative evaluations of LLMs. If the current use of Elo scores is intended to substitute the costly head-to-head comparison of LLMs, it is crucial to ensure the ranking is as robust as possible. Guided by the axioms, our findings offer concrete guidelines for enhancing the reliability of LLM evaluation methods, suggesting a need for reassessment of existing comparative approaches.

Tasks that require character-level reasoning, such as counting or locating characters within words, remain challenging for contemporary language models. A common conjecture is that language models' reliance on subword units, rather than characters, contributes to their struggles with character-level tasks, yet recent studies offer conflicting conclusions about the role of tokenization, leaving its impact unclear. To address this gap, we introduce CharBench, a comprehensive benchmark of character-level tasks that is two orders of magnitude larger than existing alternatives. We evaluate a diverse range of leading open-weight and proprietary models on CharBench and find that it presents a significant challenge to modern LLMs, with an average accuracy of 43.6% and 32.3% on some tasks. We present an in-depth analysis of how intrinsic properties of words and their segmentations into tokens correspond to model performance. For counting tasks, we find that tokenization properties are weakly correlated with correctness, while the length of the queried word and the actual character count play a more significant part. In contrast, for tasks requiring intra-word positional understanding, performance is negatively correlated with the length of the token containing the queried character, suggesting that longer tokens obscure character position information for LLMs. We encourage future work to build on the benchmark and evaluation methodology introduced here as tools for improving model performance on such tasks.

Large language models (LLMs) are notorious for hallucinating, i.e., producing erroneous claims in their output. Such hallucinations can be dangerous, as occasional factual inaccuracies in the generated text might be obscured by the rest of the output being generally factual, making it extremely hard for the users to spot them. Current services that leverage LLMs usually do not provide any means for detecting unreliable generations. Here, we aim to bridge this gap. In particular, we propose a novel fact-checking and hallucination detection pipeline based on token-level uncertainty quantification. Uncertainty scores leverage information encapsulated in the output of a neural network or its layers to detect unreliable predictions, and we show that they can be used to fact-check the atomic claims in the LLM output. Moreover, we present a novel token-level uncertainty quantification method that removes the impact of uncertainty about what claim to generate on the current step and what surface form to use. Our method Claim Conditioned Probability (CCP) measures only the uncertainty of particular claim value expressed by the model. Experiments on the task of biography generation demonstrate strong improvements for CCP compared to the baselines for six different LLMs and three languages. Human evaluation reveals that the fact-checking pipeline based on uncertainty quantification is competitive with a fact-checking tool that leverages external knowledge.

In the age of open and free information, a concerning trend of reliance on AI is emerging. However, existing AI tools struggle to evaluate the credibility of information and to justify their assessments. Hence, there is a growing need for systems that can help users evaluate the trustworthiness of online information. Although major search engines incorporate AI features, they often lack clear reliability indicators. We present TrueGL, a model that makes trustworthy search results more accessible. The model is a fine-tuned version of IBM's Granite-1B, trained on the custom dataset and integrated into a search engine with a reliability scoring system. We evaluate the system using prompt engineering and assigning each statement a continuous reliability score from 0.1 to 1, then instructing the model to return a textual explanation alongside the score. Each model's predicted scores are measured against real scores using standard evaluation metrics. TrueGL consistently outperforms other small-scale LLMs and rule-based approaches across all experiments on key evaluation metrics, including MAE, RMSE, and R2. The model's high accuracy, broad content coverage, and ease of use make trustworthy information more accessible and help reduce the spread of false or misleading content online. Our code is publicly available at https://github.com/AlgazinovAleksandr/TrueGL, and our model is publicly released at https://huggingface.co/JoydeepC/trueGL.

Reinforcement-learning-based post-training has become a key approach for improving the reasoning ability of large language models, but its token-level learning signals remain poorly understood. This work studies their heterogeneity through attention entropy, which measures how concentrated or diffuse the contextual support is for each response token. We first show that token-level RL objectives are sparsely estimable: uniformly random 20 percent token subsets preserve much of the full-token held-out performance, suggesting substantial redundancy in token-level updates. However, entropy-structured subsets behave very differently. Low-attention-entropy tokens, which we call anchors, rely on concentrated support, produce stable gradients aligned with full-token updates, and provide a reliable optimization backbone, but tend to plateau on harder benchmarks. High-attention-entropy tokens, which we call explorers, aggregate more diffuse context and induce larger but more volatile gradients. Explorer-only training is unstable on average, though rare successful runs suggest that these tokens may contain useful hard-reasoning signals when optimization remains stable. We support this anchor-explorer spectrum with evidence-gathering analyses, entropy dynamics, gradient-geometry diagnostics, and controls showing that position, predictive entropy, and loss normalization do not explain the observed asymmetry. Finally, a dynamic entropy-aware soft-reweighting intervention improves Qwen3-8B-Base from 34.39 to 37.40 held-out average in the strongest setting. These findings suggest that attention entropy reveals optimization-relevant structure in token-level RL signals, and that uniform token averaging can obscure meaningful heterogeneity in reasoning post-training.

Human evaluation is the gold-standard for evaluating text generation models. It is also expensive, and to fit budgetary constraints, a random subset of the test data is often chosen in practice. The randomly selected data may not accurately represent test performance, making this approach economically inefficient for model comparison. Thus, in this work, we develop a suite of selectors to get the most informative datapoints for human evaluation while taking the evaluation costs into account. We show that selectors based on variance in automated metric scores, diversity in model outputs, or Item Response Theory outperform random selection. We further develop an approach to distill these selectors to the scenario where the model outputs are not yet available. In particular, we introduce source-based estimators, which predict item usefulness for human evaluation just based on the source texts. We demonstrate the efficacy of our selectors in two common NLG tasks, machine translation and summarization, and show that up to only ~50% of the test data is needed to produce the same evaluation result as the entire data. Our implementations are published in the subset2evaluate package.

Tokenization is a fundamental preprocessing step in NLP, directly impacting large language models' (LLMs) ability to capture syntactic, morphosyntactic, and semantic structures. This paper introduces a novel framework for systematically evaluating tokenization strategies, addressing challenges in morphologically rich and low-resource languages. Using a Turkish dataset of 6,200 multiple-choice questions from the Massive Multitask Language Understanding (MMLU) benchmark, the framework assesses tokenizers across five key metrics: vocabulary size, token count, processing time, language-specific token percentages (\%TR), and token purity. These metrics provide a structured approach to evaluating how well tokenizers preserve linguistic structures. While \%TR measures the proportion of valid words in the target language, \%Pure assesses the alignment of tokens with meaningful linguistic units, such as roots and valid morphemes, minimizing semantic fragmentation. The findings reveal that \%TR, introduced as a critical metric, exhibits a stronger correlation with downstream performance (e.g., MMLU scores) than token purity, emphasizing its role in improving model accuracy. Additionally, larger model parameters do not necessarily yield better tokenization quality or enhanced results, highlighting the importance of tailored tokenization strategies that prioritize linguistic alignment. This framework sets a new standard for developing robust tokenization methods optimized for morphologically complex and low-resource languages. Future work will refine morphological analysis, explore domain-specific customizations, and conduct cross-linguistic evaluations to further enhance tokenization practices.

Thousands of diverse benchmarks have been developed to measure the quality of large language models (LLMs). Yet prior work has demonstrated that LLM performance is often sufficiently explained by a small set of latent factors, or abilities. This suggests the potential for more efficient and principled benchmarking, but it remains difficult to compare the quality of different methods. Motivated by predictive validity, we argue that the quality of a benchmarking framework should be grounded in how efficiently it enables the prediction of model performance on unseen tasks. To analyze this objective, we collect the "Wide-scale Item Level Dataset" (WILD), a dataset of item-model response pairs, comprising evaluations of 65 models on 109,564 unique items spanning 163 tasks drawn from 27 datasets. This dataset enables the first analysis of how different techniques can predict a model's performance on a large, diverse collection of unseen tasks under different budget constraints. We demonstrate that combining a modified multidimensional item response theory (IRT) model with adaptive item selection driven by optimal experimental design can predict performance on 112 held-out benchmark tasks with a mean absolute error (MAE) of less than 7%, and can do so after observing only 16 items. We further demonstrate that incorporating cost-aware discount factors into our selection criteria can reduce the total tokens needed to reach 7% MAE from 141,000 tokens to only 22,000, an 85% reduction in evaluation cost.

Large Language Models (LLMs) are widely deployed in real-world applications, yet little is known about their training dynamics at the token level. Evaluation typically relies on aggregated training loss, measured at the batch level, which overlooks subtle per-token biases arising from (i) varying token-level dynamics and (ii) structural biases introduced by hyperparameters. While weight decay is commonly used to stabilize training, we reveal that it silently introduces performance biases detectable only at the token level. In fact, we empirically show across different dataset sizes, model architectures and sizes ranging from 270M to 3B parameters that as weight decay increases, low-frequency tokens are disproportionately depreciated. This is particularly concerning, as these neglected low-frequency tokens represent the vast majority of the token distribution in most languages, calling for novel regularization techniques that ensure fairness across all available tokens.

We study reliability in autonomous language-model agents that translate user mandates into validated tool actions under real capital. The setting is DX Terminal Pro, a 21-day deployment in which 3,505 user-funded agents traded real ETH in a bounded onchain market. Users configured vaults through structured controls and natural-language strategies, but only agents could choose normal buy/sell trades. The system produced 7.5M agent invocations, roughly 300K onchain actions, about $20M in volume, more than 5,000 ETH deployed, roughly 70B inference tokens, and 99.9% settlement success for policy-valid submitted transactions. Long-running agents accumulated thousands of sequential decisions, including 6,000+ prompt-state-action cycles for continuously active agents, yielding a large-scale trace from user mandate to rendered prompt, reasoning, validation, portfolio state, and settlement. Reliability did not come from the base model alone; it emerged from the operating layer around the model: prompt compilation, typed controls, policy validation, execution guards, memory design, and trace-level observability. Pre-launch testing exposed failures that text-only benchmarks rarely measure, including fabricated trading rules, fee paralysis, numeric anchoring, cadence trading, and misread tokenomics. Targeted harness changes reduced fabricated sell rules from 57% to 3%, reduced fee-led observations from 32.5% to below 10%, and increased capital deployment from 42.9% to 78.0% in an affected test population. We show that capital-managing agents should be evaluated across the full path from user mandate to prompt, validated action, and settlement.

Generative spoken language models pretrained on large-scale raw audio can continue a speech prompt with appropriate content while preserving attributes like speaker and emotion, serving as foundation models for spoken dialogue. In prior literature, these models are often evaluated using ``global token perplexity'', which directly applies the text perplexity formulation to speech tokens. However, this practice overlooks fundamental differences between speech and text modalities, possibly leading to an underestimation of the speech characteristics. In this work, we propose a variety of likelihood- and generative-based evaluation methods that serve in place of naive global token perplexity. We demonstrate that the proposed evaluations more faithfully reflect perceived generation quality, as evidenced by stronger correlations with human-rated mean opinion scores (MOS). When assessed under the new metrics, the relative performance landscape of spoken language models is reshaped, revealing a significantly reduced gap between the best-performing model and the human topline. Together, these results suggest that appropriate evaluation is critical for accurately assessing progress in spoken language modeling.

In recent years, general visual foundation models (VFMs) have witnessed increasing adoption, particularly as image encoders for popular multi-modal large language models (MLLMs). However, without semantically fine-grained supervision, these models still encounter fundamental prediction errors in the context of downstream text-image-related tasks, i.e., perception, understanding and reasoning with images containing small and dense texts. To bridge this gap, we develop TokenOCR, the first token-level visual foundation model specifically tailored for text-image-related tasks, designed to support a variety of traditional downstream applications. To facilitate the pretraining of TokenOCR, we also devise a high-quality data production pipeline that constructs the first token-level image text dataset, TokenIT, comprising 20 million images and 1.8 billion token-mask pairs. Furthermore, leveraging this foundation with exceptional image-as-text capability, we seamlessly replace previous VFMs with TokenOCR to construct a document-level MLLM, TokenVL, for VQA-based document understanding tasks. Finally, extensive experiments demonstrate the effectiveness of TokenOCR and TokenVL. Code, datasets, and weights will be available at https://token-family.github.io/TokenOCR_project.

We introduce a corpus of 7,032 sentences rated by human annotators for formality, informativeness, and implicature on a 1-7 scale. The corpus was annotated using Amazon Mechanical Turk. Reliability in the obtained judgments was examined by comparing mean ratings across two MTurk experiments, and correlation with pilot annotations (on sentence formality) conducted in a more controlled setting. Despite the subjectivity and inherent difficulty of the annotation task, correlations between mean ratings were quite encouraging, especially on formality and informativeness. We further explored correlation between the three linguistic variables, genre-wise variation of ratings and correlations within genres, compatibility with automatic stylistic scoring, and sentential make-up of a document in terms of style. To date, our corpus is the largest sentence-level annotated corpus released for formality, informativeness, and implicature.

When deploying large language models (LLMs), it is important to ensure that these models are not only capable, but also reliable. Many benchmarks have been created to track LLMs' growing capabilities, however there has been no similar focus on measuring their reliability. To understand the potential ramifications of this gap, we investigate how well current benchmarks quantify model reliability. We find that pervasive label errors can compromise these evaluations, obscuring lingering model failures and hiding unreliable behavior. Motivated by this gap in the evaluation of reliability, we then propose the concept of so-called platinum benchmarks, i.e., benchmarks carefully curated to minimize label errors and ambiguity. As a first attempt at constructing such benchmarks, we revise examples from fifteen existing popular benchmarks. We evaluate a wide range of models on these platinum benchmarks and find that, indeed, frontier LLMs still exhibit failures on simple tasks such as elementary-level math word problems. Analyzing these failures further reveals previously unidentified patterns of problems on which frontier models consistently struggle. We provide code at https://github.com/MadryLab/platinum-benchmarks

The predictions of question answering (QA)systems are typically evaluated against manually annotated finite sets of one or more answers. This leads to a coverage limitation that results in underestimating the true performance of systems, and is typically addressed by extending over exact match (EM) with pre-defined rules or with the token-level F1 measure. In this paper, we present the first systematic conceptual and data-driven analysis to examine the shortcomings of token-level equivalence measures. To this end, we define the asymmetric notion of answer equivalence (AE), accepting answers that are equivalent to or improve over the reference, and publish over 23k human judgments for candidates produced by multiple QA systems on SQuAD. Through a careful analysis of this data, we reveal and quantify several concrete limitations of the F1 measure, such as a false impression of graduality, or missing dependence on the question. Since collecting AE annotations for each evaluated model is expensive, we learn a BERT matching (BEM) measure to approximate this task. Being a simpler task than QA, we find BEM to provide significantly better AE approximations than F1, and to more accurately reflect the performance of systems. Finally, we demonstrate the practical utility of AE and BEM on the concrete application of minimal accurate prediction sets, reducing the number of required answers by up to x2.6.

Representing speech as discrete tokens provides a framework for transforming speech into a format that closely resembles text, thus enabling the use of speech as an input to the widely successful large language models (LLMs). Currently, while several speech tokenizers have been proposed, there is ambiguity regarding the properties that are desired from a tokenizer for specific downstream tasks and its overall generalizability. Evaluating the performance of tokenizers across different downstream tasks is a computationally intensive effort that poses challenges for scalability. To circumvent this requirement, we present STAB (Speech Tokenizer Assessment Benchmark), a systematic evaluation framework designed to assess speech tokenizers comprehensively and shed light on their inherent characteristics. This framework provides a deeper understanding of the underlying mechanisms of speech tokenization, thereby offering a valuable resource for expediting the advancement of future tokenizer models and enabling comparative analysis using a standardized benchmark. We evaluate the STAB metrics and correlate this with downstream task performance across a range of speech tasks and tokenizer choices.

A unique aspect of ColBERT is its use of [MASK] tokens in queries to score documents (query augmentation). Prior work shows [MASK] tokens weighting non-[MASK] query terms, emphasizing certain tokens over others , rather than introducing whole new terms as initially proposed. We begin by demonstrating that a term weighting behavior previously reported for [MASK] tokens in ColBERTv1 holds for ColBERTv2. We then examine the effect of changing the number of [MASK] tokens from zero to up to four times past the query input length used in training, both for first stage retrieval, and for scoring candidates, observing an initial decrease in performance with few [MASK]s, a large increase when enough [MASK]s are added to pad queries to an average length of 32, then a plateau in performance afterwards. Additionally, we compare baseline performance to performance when the query length is extended to 128 tokens, and find that differences are small (e.g., within 1% on various metrics) and generally statistically insignificant, indicating performance does not collapse if ColBERT is presented with more [MASK] tokens than expected.

Large language models (LLMs) have demonstrated significant utility in real-world applications, exhibiting impressive capabilities in natural language processing and understanding. Benchmark evaluations are crucial for assessing the capabilities of LLMs as they can provide a comprehensive assessment of their strengths and weaknesses. However, current evaluation methods often overlook the inherent randomness of LLMs by employing deterministic generation strategies or relying on a single random sample, resulting in unaccounted sampling variance and unreliable benchmark score estimates. In this paper, we propose a hierarchical statistical model that provides a more comprehensive representation of the benchmarking process by incorporating both benchmark characteristics and LLM randomness. We show that leveraging multiple generations improves the accuracy of estimating the benchmark score and reduces variance. Multiple generations also allow us to define mathbb Pleft(correctright), a prompt-level difficulty score based on correct ratios, providing fine-grained insights into individual prompts. Additionally, we create a data map that visualizes difficulty and semantics of prompts, enabling error detection and quality control in benchmark construction.

We investigate the relationship between the geometry of token embeddings and their role in the next token prediction within transformer models. An important aspect of this connection uses the notion of empirical measure, which encodes the distribution of token point clouds across transformer layers and drives the evolution of token representations in the mean-field interacting picture. We use metrics such as intrinsic dimension, neighborhood overlap, and cosine similarity to observationally probe these empirical measures across layers. To validate our approach, we compare these metrics to a dataset where the tokens are shuffled, which disrupts the syntactic and semantic structure. Our findings reveal a correlation between the geometric properties of token embeddings and the cross-entropy loss of next token predictions, implying that prompts with higher loss values have tokens represented in higher-dimensional spaces.

Prevalent semantic speech tokenizers, designed to capture linguistic content, are surprisingly fragile. We find they are not robust to meaning-irrelevant acoustic perturbations; even at high Signal-to-Noise Ratios (SNRs) where speech is perfectly intelligible, their output token sequences can change drastically, increasing the learning burden for downstream LLMs. This instability stems from two flaws: a brittle single-path quantization architecture and a distant training signal indifferent to intermediate token stability. To address this, we introduce StableToken, a tokenizer that achieves stability through a consensus-driven mechanism. Its multi-branch architecture processes audio in parallel, and these representations are merged via a powerful bit-wise voting mechanism to form a single, stable token sequence. StableToken sets a new state-of-the-art in token stability, drastically reducing Unit Edit Distance (UED) under diverse noise conditions. This foundational stability translates directly to downstream benefits, significantly improving the robustness of SpeechLLMs on a variety of tasks.

Vision-language models (VLMs) are increasingly augmented with continuous or latent non-textual tokens intended to support "visual thinking." Despite improved task accuracy, this alone does not show that models actually use these tokens for reasoning -- gains may arise from confounds such as added context length, special-token anchoring, or training-time regularization. We formalize a diagnostic principle, Ablate-to-Validate, for testing whether latent-token content is genuinely utilized, and instantiate it as the Token Replacement Test (TRT), a standardized suite of content-replacement ablations. TRT holds the prompt, image, token budget, and decoding fixed while replacing intermediate tokens with zero, random, first-repeat, or oracle alternatives, isolating whether performance depends on token content or merely on token presence. As a controlled testbed, we study relative depth reasoning with LLaVA-13B and Qwen2.5-VL-3B, training models to predict and consume continuous or discrete depth spans across multiple frozen encoders (SigLIP2, CLIP, DINOv2) and token budgets. We additionally apply TRT to three off-the-shelf visual-thinking systems (Mirage, Mull-Tokens, CoVT) on BLINK, VSP, and CV-Bench. Across all settings, accuracy gains are a misleading proxy for latent-token reasoning: VLMs retain most improvement even when token content is corrupted or replaced, revealing a persistent gap between having a latent channel and using it as an information bottleneck. We recommend TRT as a standard diagnostic alongside accuracy for any method introducing continuous thought tokens.

The rapid development of large language models (LLMs) has shown promising practical results. However, their low interpretability often leads to errors in unforeseen circumstances, limiting their utility. Many works have focused on creating comprehensive evaluation systems, but previous benchmarks have primarily assessed problem-solving abilities while neglecting the response's uncertainty, which may result in unreliability. Recent methods for measuring LLM reliability are resource-intensive and unable to test black-box models. To address this, we propose UBENCH, a comprehensive benchmark for evaluating LLM reliability. UBENCH includes 3,978 multiple-choice questions covering knowledge, language, understanding, and reasoning abilities. Experimental results show that UBENCH has achieved state-of-the-art performance, while its single-sampling method significantly saves computational resources compared to baseline methods that require multiple samplings. Additionally, based on UBENCH, we evaluate the reliability of 15 popular LLMs, finding GLM4 to be the most outstanding, closely followed by GPT-4. We also explore the impact of Chain-of-Thought prompts, role-playing prompts, option order, and temperature on LLM reliability, analyzing the varying effects on different LLMs.

Uncertainty Quantification (UQ) research has primarily focused on closed-book factual question answering (QA), while contextual QA remains unexplored, despite its importance in real-world applications. In this work, we focus on UQ for the contextual QA task and propose a theoretically grounded approach to quantify epistemic uncertainty. We begin by introducing a task-agnostic, token-level uncertainty measure defined as the cross-entropy between the predictive distribution of the given model and the unknown true distribution. By decomposing this measure, we isolate the epistemic component and approximate the true distribution by a perfectly prompted, idealized model. We then derive an upper bound for epistemic uncertainty and show that it can be interpreted as semantic feature gaps in the given model's hidden representations relative to the ideal model. We further apply this generic framework to the contextual QA task and hypothesize that three features approximate this gap: context-reliance (using the provided context rather than parametric knowledge), context comprehension (extracting relevant information from context), and honesty (avoiding intentional lies). Using a top-down interpretability approach, we extract these features by using only a small number of labeled samples and ensemble them to form a robust uncertainty score. Experiments on multiple QA benchmarks in both in-distribution and out-of-distribution settings show that our method substantially outperforms state-of-the-art unsupervised (sampling-free and sampling-based) and supervised UQ methods, achieving up to a 13-point PRR improvement while incurring a negligible inference overhead.

With the development of large language models (LLMs), the ability to handle longer contexts has become a key capability for Web applications such as cross-document understanding and LLM-powered search systems. However, this progress faces two major challenges: performance degradation due to sequence lengths out-of-distribution, and excessively long inference times caused by the quadratic computational complexity of attention. These issues hinder the application of LLMs in long-context scenarios. In this paper, we propose Dynamic Token-Level KV Cache Selection (TokenSelect), a model-agnostic, training-free method for efficient and accurate long-context inference. TokenSelect builds upon the observation of non-contiguous attention sparsity, using Query-Key dot products to measure per-head KV Cache criticality at token-level. By per-head soft voting mechanism, TokenSelect selectively involves a small number of critical KV cache tokens in the attention calculation without sacrificing accuracy. To further accelerate TokenSelect, we designed the Selection Cache based on observations of consecutive Query similarity and implemented efficient dot product kernel, significantly reducing the overhead of token selection. A comprehensive evaluation of TokenSelect demonstrates up to 23.84x speedup in attention computation and up to 2.28x acceleration in end-to-end latency, while providing superior performance compared to state-of-the-art long-context inference methods.

Multi-vector retrieval models such as ColBERT [Khattab and Zaharia, 2020] allow token-level interactions between queries and documents, and hence achieve state of the art on many information retrieval benchmarks. However, their non-linear scoring function cannot be scaled to millions of documents, necessitating a three-stage process for inference: retrieving initial candidates via token retrieval, accessing all token vectors, and scoring the initial candidate documents. The non-linear scoring function is applied over all token vectors of each candidate document, making the inference process complicated and slow. In this paper, we aim to simplify the multi-vector retrieval by rethinking the role of token retrieval. We present XTR, ConteXtualized Token Retriever, which introduces a simple, yet novel, objective function that encourages the model to retrieve the most important document tokens first. The improvement to token retrieval allows XTR to rank candidates only using the retrieved tokens rather than all tokens in the document, and enables a newly designed scoring stage that is two-to-three orders of magnitude cheaper than that of ColBERT. On the popular BEIR benchmark, XTR advances the state-of-the-art by 2.8 nDCG@10 without any distillation. Detailed analysis confirms our decision to revisit the token retrieval stage, as XTR demonstrates much better recall of the token retrieval stage compared to ColBERT.

Moral sentiments expressed in natural language significantly influence both online and offline environments, shaping behavioral styles and interaction patterns, including social media selfpresentation, cyberbullying, adherence to social norms, and ethical decision-making. To effectively measure moral sentiments in natural language processing texts, it is crucial to utilize large, annotated datasets that provide nuanced understanding for accurate analysis and modeltraining. However, existing corpora, while valuable, often face linguistic limitations. To address this gap in the Chinese language domain,we introduce the Moral Foundation Weibo Corpus. This corpus consists of 25,671 Chinese comments on Weibo, encompassing six diverse topic areas. Each comment is manually annotated by at least three systematically trained annotators based on ten moral categories derived from a grounded theory of morality. To assess annotator reliability, we present the kappa testresults, a gold standard for measuring consistency. Additionally, we apply several the latest large language models to supplement the manual annotations, conducting analytical experiments to compare their performance and report baseline results for moral sentiment classification.

Large language models (LLMs) have demonstrated impressive reasoning capabilities by scaling test-time compute via long Chain-of-Thought (CoT). However, recent findings suggest that raw token counts are unreliable proxies for reasoning quality: increased generation length does not consistently correlate with accuracy and may instead signal "overthinking," leading to performance degradation. In this work, we quantify inference-time effort by identifying deep-thinking tokens -- tokens where internal predictions undergo significant revisions in deeper model layers prior to convergence. Across four challenging mathematical and scientific benchmarks (AIME 24/25, HMMT 25, and GPQA-diamond) and a diverse set of reasoning-focused models (GPT-OSS, DeepSeek-R1, and Qwen3), we show that deep-thinking ratio (the proportion of deep-thinking tokens in a generated sequence) exhibits a robust and consistently positive correlation with accuracy, substantially outperforming both length-based and confidence-based baselines. Leveraging this insight, we introduce Think@n, a test-time scaling strategy that prioritizes samples with high deep-thinking ratios. We demonstrate that Think@n matches or exceeds standard self-consistency performance while significantly reducing inference costs by enabling the early rejection of unpromising generations based on short prefixes.

Cross-tokenizer knowledge distillation allows a student model to learn from teachers with incompatible vocabularies. Prior work operates on hidden states or logits; the latter is preferred as a drop-in replacement requiring no auxiliary components. Logit-based methods either use only the correct-token probability, missing the full 'dark knowledge' in the teacher's distribution, or operate on the full output distribution, relying on strict token partitioning and/or unprincipled heuristic ranking. We identify two key shortcomings of full-distribution, logit-based methods: (i) an uncommon-token failure, where critical tokens fall into the unmatched subset (e.g., Llama's 1100 multi-digit numerals under digit-splitting Qwen supervision) and are suppressed during training, reducing GSM8k from 12.89 to 2.56 compared to same-tokenizer KD from a weaker teacher; and (ii) over-conservative matching, where strict 1-to-1 matching excludes near-equivalent tokens across surface forms. These failures require distinct remedies: eliminating the partition when critical tokens are misaligned, and refining it when alignment is reliable. We propose X-Token, an approach with two complementary loss formulations targeting these issues. P-KL removes partitioning and aligns the student's distribution with the teacher's via a sparse projection matrix W (initialized from tokenizer-level string rules) to address the uncommon-token failure. H-KL retains the hybrid form while relaxing matching to align each student token with its top-ranked teacher mapping under W. Both objectives share W and extend naturally to multiple teachers. Empirically, on Llama-3.2-1B, X-Token outperforms the current state of the art GOLD by +3.82 average points with a Qwen3-4B teacher and by +0.5 with a Phi-4-Mini teacher. Further, a two-teacher setup (Phi-4-mini + Llama-3B) improves over single-teacher distillation by +1.3 points.

Data quality plays a critical role in enhancing supervised fine-tuning (SFT) for large language models (LLMs), and token-level data selection has emerged as a promising direction for its fine-grained nature. Despite their strong empirical performance, existing token-level selection methods share two key limitations: (1) requiring training or accessing an additional reference model, and (2) relying solely on loss information for token selection, which cannot well preserve semantically important tokens that are not favored by loss-based metrics. To address these challenges, we propose ssToken, a Self-modulated and Semantic-aware Token Selection approach. ssToken leverages readily accessible history models to compute the per-token loss difference with the current model, which serves as a self-modulated signal that enables the model to adaptively select tokens along its optimization trajectory, rather than relying on excess loss from an offline-trained reference model as in prior works. We further introduce a semantic-aware, attention-based token importance estimation metric, orthogonal to loss-based selection and providing complementary semantic information for more effective filtering. Extensive experiments across different model families and scales demonstrate that both self-modulated selection and semantic-aware selection alone outperform full-data fine-tuning, while their integration--ssToken--achieves synergistic gains and further surpasses prior token-level selection methods, delivering performance improvements while maintaining training efficiency.

When evaluating large language models (LLMs) with multiple-choice question answering (MCQA), it is common to end the prompt with the string "Answer:" to facilitate automated answer extraction via next-token probabilities. However, there is no consensus on how to tokenize the space following the colon, often overlooked as a trivial choice. In this paper, we uncover accuracy differences of up to 11% due to this (seemingly irrelevant) tokenization variation as well as reshuffled model rankings, raising concerns about the reliability of LLM comparisons in prior work. Surprisingly, we are able to recommend one specific strategy -- tokenizing the space together with the answer letter -- as we observe consistent and statistically significant performance improvements. Additionally, it improves model calibration, enhancing the reliability of the model's confidence estimates. Our findings underscore the importance of careful evaluation design and highlight the need for standardized, transparent evaluation protocols to ensure reliable and comparable results.

The way users acquire information is undergoing a paradigm shift with the advent of ChatGPT. Unlike conventional search engines, ChatGPT retrieves knowledge from the model itself and generates answers for users. ChatGPT's impressive question-answering (QA) capability has attracted more than 100 million users within a short period of time but has also raised concerns regarding its reliability. In this paper, we perform the first large-scale measurement of ChatGPT's reliability in the generic QA scenario with a carefully curated set of 5,695 questions across ten datasets and eight domains. We find that ChatGPT's reliability varies across different domains, especially underperforming in law and science questions. We also demonstrate that system roles, originally designed by OpenAI to allow users to steer ChatGPT's behavior, can impact ChatGPT's reliability. We further show that ChatGPT is vulnerable to adversarial examples, and even a single character change can negatively affect its reliability in certain cases. We believe that our study provides valuable insights into ChatGPT's reliability and underscores the need for strengthening the reliability and security of large language models (LLMs).

Critic-free methods like GRPO reduce memory demands by estimating advantages from multiple rollouts but tend to converge slowly, as critical learning signals are diluted by an abundance of uninformative samples and tokens. To tackle this challenge, we propose the Dynamic Dual-Level Down-Sampling (D^3S) framework that prioritizes the most informative samples and tokens across groups to improve the efficient of policy optimization. D^3S operates along two levels: (1) the sample-level, which selects a subset of rollouts to maximize advantage variance (Var(A)). We theoretically proven that this selection is positively correlated with the upper bound of the policy gradient norms, yielding higher policy gradients. (2) the token-level, which prioritizes tokens with a high product of advantage magnitude and policy entropy (|A_{i,t}|times H_{i,t}), focusing updates on tokens where the policy is both uncertain and impactful. Moreover, to prevent overfitting to high-signal data, D^3S employs a dynamic down-sampling schedule inspired by curriculum learning. This schedule starts with aggressive down-sampling to accelerate early learning and gradually relaxes to promote robust generalization. Extensive experiments on Qwen2.5 and Llama3.1 demonstrate that integrating D^3S into advanced RL algorithms achieves state-of-the-art performance and generalization while requiring fewer samples and tokens across diverse reasoning benchmarks. Our code is added in the supplementary materials and will be made publicly available.

Latent tokens are gaining attention for enhancing reasoning in large language models (LLMs), yet their internal mechanisms remain unclear. This paper examines the problem from a reliability perspective, uncovering fundamental weaknesses: latent tokens function as uninterpretable placeholders rather than encoding faithful reasoning. While resistant to perturbation, they promote shortcut usage over genuine reasoning. We focus on Chain-of-Continuous-Thought (COCONUT), which claims better efficiency and stability than explicit Chain-of-Thought (CoT) while maintaining performance. We investigate this through two complementary approaches. First, steering experiments perturb specific token subsets, namely COCONUT and explicit CoT. Unlike CoT tokens, COCONUT tokens show minimal sensitivity to steering and lack reasoning-critical information. Second, shortcut experiments evaluate models under biased and out-of-distribution settings. Results on MMLU and HotpotQA demonstrate that COCONUT consistently exploits dataset artifacts, inflating benchmark performance without true reasoning. These findings reposition COCONUT as a pseudo-reasoning mechanism: it generates plausible traces that conceal shortcut dependence rather than faithfully representing reasoning processes.

Recent machine translation (MT) metrics calibrate their effectiveness by correlating with human judgement but without any insights about their behaviour across different error types. Challenge sets are used to probe specific dimensions of metric behaviour but there are very few such datasets and they either focus on a limited number of phenomena or a limited number of language pairs. We introduce ACES, a contrastive challenge set spanning 146 language pairs, aimed at discovering whether metrics can identify 68 translation accuracy errors. These phenomena range from simple alterations at the word/character level to more complex errors based on discourse and real-world knowledge. We conduct a large-scale study by benchmarking ACES on 50 metrics submitted to the WMT 2022 and 2023 metrics shared tasks. We benchmark metric performance, assess their incremental performance over successive campaigns, and measure their sensitivity to a range of linguistic phenomena. We also investigate claims that Large Language Models (LLMs) are effective as MT evaluators by evaluating on ACES. Our results demonstrate that different metric families struggle with different phenomena and that LLM-based methods fail to demonstrate reliable performance. Our analyses indicate that most metrics ignore the source sentence, tend to prefer surface-level overlap and end up incorporating properties of base models which are not always beneficial. We expand ACES to include error span annotations, denoted as SPAN-ACES and we use this dataset to evaluate span-based error metrics showing these metrics also need considerable improvement. Finally, we provide a set of recommendations for building better MT metrics, including focusing on error labels instead of scores, ensembling, designing strategies to explicitly focus on the source sentence, focusing on semantic content and choosing the right base model for representations.

Reasoning capability has advanced rapidly in large language models (LLMs), leading to an increasing size of key-value (KV) cache in both prefilling and decoding stages. Existing KV cache compression methods mainly rely on attention weights to estimate token importance. While attention effectively captures contextual relevance, it overlooks complementary information-theoretic signals related to predictive uncertainty and token informativeness. In this paper, we revisit token importance from a forward-looking perspective and introduce Forward Influence, a metric that measures how compressed tokens affect future contexts. Our analysis reveals that tokens selected by attention scores mainly influence nearby contexts, whereas tokens associated with high predictive uncertainty exhibit substantially stronger influence on distant future contexts. Based on the observation, we propose InfoKV, an entropy-aware KV cache compression framework that incorporates information-theoretic signals. It combines token-level predictive uncertainty with layer-wise representation evolution and integrates the resulting entropy scores with attention scores during reasoning. Experiments on long-context reasoning benchmarks with Llama-3.1, Llama-3.2, and DeepSeek-R1 demonstrate that InfoKV consistently outperforms existing attention-based KV compression methods in both long prefilling and decoding scenarios.

One of the most striking findings in modern research on large language models (LLMs) is that scaling up compute during training leads to better results. However, less attention has been given to the benefits of scaling compute during inference. This survey focuses on these inference-time approaches. We explore three areas under a unified mathematical formalism: token-level generation algorithms, meta-generation algorithms, and efficient generation. Token-level generation algorithms, often called decoding algorithms, operate by sampling a single token at a time or constructing a token-level search space and then selecting an output. These methods typically assume access to a language model's logits, next-token distributions, or probability scores. Meta-generation algorithms work on partial or full sequences, incorporating domain knowledge, enabling backtracking, and integrating external information. Efficient generation methods aim to reduce token costs and improve the speed of generation. Our survey unifies perspectives from three research communities: traditional natural language processing, modern LLMs, and machine learning systems.

Recent advances have demonstrated that integrating reinforcement learning with rule-based rewards can significantly enhance the reasoning capabilities of large language models, even without supervised fine-tuning. However, prevalent reinforcement learning algorithms such as GRPO and its variants like DAPO, suffer from a coarse granularity issue when computing the advantage. Specifically, they compute rollout-level advantages that assign identical values to every token within a sequence, failing to capture token-specific contributions and hindering effective learning. To address this limitation, we propose Key-token Advantage Estimation (KTAE) - a novel algorithm that estimates fine-grained, token-level advantages without introducing additional models. KTAE leverages the correctness of sampled rollouts and applies statistical analysis to quantify the importance of individual tokens within a sequence to the final outcome. This quantified token-level importance is then combined with the rollout-level advantage to obtain a more fine-grained token-level advantage estimation. Empirical results show that models trained with GRPO+KTAE and DAPO+KTAE outperform baseline methods across five mathematical reasoning benchmarks. Notably, they achieve higher accuracy with shorter responses and even surpass R1-Distill-Qwen-1.5B using the same base model.

Recent studies show that in supervised fine-tuning (SFT) of large language models (LLMs), data quality matters more than quantity. While most data cleaning methods concentrate on filtering entire samples, the quality of individual tokens within a sample can vary significantly. After pre-training, even in high-quality samples, patterns or phrases that are not task-related can be redundant, uninformative, or even harmful. Continuing to fine-tune on these patterns may offer limited benefit and even degrade downstream task performance. In this paper, we investigate token quality from a noisy-label perspective and propose a generic token cleaning pipeline for SFT tasks. Our method filters out uninformative tokens while preserving those carrying key task-specific information. Specifically, we first evaluate token quality by examining the influence of model updates on each token, then apply a threshold-based separation. The token influence can be measured in a single pass with a fixed reference model or iteratively with self-evolving reference models. The benefits and limitations of both methods are analyzed theoretically by error upper bounds. Extensive experiments show that our framework consistently improves downstream performance. Code is available at https://github.com/UCSC-REAL/TokenCleaning.

Current LLM safety benchmarks are predominantly English-centric and often rely on translation, failing to capture country-specific harms. Moreover, they rarely evaluate a model's ability to detect culturally embedded sensitivities as distinct from universal harms. We introduce XL-SafetyBench. a suite of 5,500 test cases across 10 country-language pairs, comprising a Jailbreak Benchmark of country-grounded adversarial prompts and a Cultural Benchmark where local sensitivities are embedded within innocuous requests. Each item is constructed via a multi-stage pipeline that combines LLM-assisted discovery, automated validation gates, and dual independent native-speaker annotators per country. To distinguish principled refusal from comprehension failure, we evaluate Attack Success Rate (ASR) alongside two complementary metrics we introduce: Neutral-Safe Rate (NSR) and Cultural Sensitivity Rate (CSR). Evaluating 10 frontier and 27 local LLMs reveals two key findings. First, jailbreak robustness and cultural awareness do not show a coupled relationship among frontier models, so a composite safety score obscures per-axis variation. Second, local models exhibit a near-linear ASR-NSR trade-off (r = -0.81), indicating that their apparent safety reflects generation failure rather than genuine alignment. XL-SafetyBench enables more nuanced, cross-cultural safety evaluation in the multilingual era.

Comprehensive evaluations of language models (LM) during both development and deployment phases are necessary because these models possess numerous capabilities (e.g., mathematical reasoning, legal support, or medical diagnostic) as well as safety risks (e.g., racial bias, toxicity, or misinformation). The average score across a wide range of benchmarks provides a signal that helps guide the use of these LMs in practice. Currently, holistic evaluations are costly due to the large volume of benchmark questions, making frequent evaluations impractical. A popular attempt to lower the cost is to compute the average score on a subset of the benchmark. This approach, unfortunately, often renders an unreliable measure of LM performance because the average score is often confounded with the difficulty of the questions in the benchmark subset. Item response theory (IRT) was designed to address this challenge, providing a reliable measurement by careful controlling for question difficulty. Unfortunately, question difficulty is expensive to estimate. Facing this challenge, we train a model that predicts question difficulty from its content, enabling a reliable measurement at a fraction of the cost. In addition, we leverage this difficulty predictor to further improve the evaluation efficiency through training a question generator given a difficulty level. This question generator is essential in adaptive testing, where, instead of using a random subset of the benchmark questions, informative questions are adaptively chosen based on the current estimation of LLM performance. Experiments on 22 common natural language benchmarks and 172 LMs show that this approach is more reliable and efficient compared to current common practice.

Large language models (LLMs) are increasingly deployed as customer-facing agents, yet evaluating their reliability remains challenging due to stochastic, multi-turn interactions. Current evaluation protocols rely on linear Monte Carlo rollouts of complete agent-user conversations to estimate success. However, this approach is computationally inefficient, repeatedly regenerating identical early prefixes, and often fails to uncover deep failure modes that arise from rare user behaviors. We introduce DIVERT (Diversity-Induced Evaluation via Branching of Trajectories), an efficient, snapshot-based, coverage-guided user simulation framework for systematic exploration of agent-user interactions. DIVERT captures the full agent-environment state at critical decision points and resumes execution from these snapshots, enabling reuse of shared conversation prefixes and reducing redundant computation. From each junction, the framework branches using targeted, diversity-inducing user responses, allowing directed exploration of alternative interaction paths. By focusing evaluation on semantically diverse and underexplored trajectories, DIVERT improves both efficiency and coverage. Empirical results show that it discovers more failures per token compared to standard linear rollout protocols, while expanding the set of tasks on which failures are identified.

Accurate and consistent evaluation is crucial for decision-making across numerous fields, yet it remains a challenging task due to inherent subjectivity, variability, and scale. Large Language Models (LLMs) have achieved remarkable success across diverse domains, leading to the emergence of "LLM-as-a-Judge," where LLMs are employed as evaluators for complex tasks. With their ability to process diverse data types and provide scalable, cost-effective, and consistent assessments, LLMs present a compelling alternative to traditional expert-driven evaluations. However, ensuring the reliability of LLM-as-a-Judge systems remains a significant challenge that requires careful design and standardization. This paper provides a comprehensive survey of LLM-as-a-Judge, addressing the core question: How can reliable LLM-as-a-Judge systems be built? We explore strategies to enhance reliability, including improving consistency, mitigating biases, and adapting to diverse assessment scenarios. Additionally, we propose methodologies for evaluating the reliability of LLM-as-a-Judge systems, supported by a novel benchmark designed for this purpose. To advance the development and real-world deployment of LLM-as-a-Judge systems, we also discussed practical applications, challenges, and future directions. This survey serves as a foundational reference for researchers and practitioners in this rapidly evolving field.

Most evaluations of generative models rely on feature-distribution metrics such as FID, which operate on continuous recognition features that are explicitly trained to be invariant to appearance variations, and thus discard cues critical for perceptual quality. We instead evaluate models in the space of discrete visual tokens, where modern 1D image tokenizers compactly encode both semantic and perceptual information and quality manifests as predictable token statistics. We introduce Codebook Histogram Distance (CHD), a training-free distribution metric in token space, and Code Mixture Model Score (CMMS), a no-reference quality metric learned from synthetic degradations of token sequences. To stress-test metrics under broad distribution shifts, we further propose VisForm, a benchmark of 210K images spanning 62 visual forms and 12 generative models with expert annotations. Across AGIQA, HPDv2/3, and VisForm, our token-based metrics achieve state-of-the-art correlation with human judgments, and we will release all code and datasets to facilitate future research.

Although neural-based machine translation evaluation metrics, such as COMET or BLEURT, have achieved strong correlations with human judgements, they are sometimes unreliable in detecting certain phenomena that can be considered as critical errors, such as deviations in entities and numbers. In contrast, traditional evaluation metrics, such as BLEU or chrF, which measure lexical or character overlap between translation hypotheses and human references, have lower correlations with human judgements but are sensitive to such deviations. In this paper, we investigate several ways of combining the two approaches in order to increase robustness of state-of-the-art evaluation methods to translations with critical errors. We show that by using additional information during training, such as sentence-level features and word-level tags, the trained metrics improve their capability to penalize translations with specific troublesome phenomena, which leads to gains in correlation with human judgments and on recent challenge sets on several language pairs.

We present a large-scale human evaluation benchmark for assessing cultural localisation in machine translation produced by state-of-the-art multilingual large language models (LLMs). Existing MT benchmarks emphasise token-level and grammatical accuracy, but of ten overlook pragmatic and culturally grounded competencies required for real-world localisation. Building on a pilot study of 87 translations across 20 languages, we evaluate 7 multilingual LLMs across 15 target languages with 5 native-speaker raters per language. Raters scored both full-text translations and segment-level instances of culturally nuanced language (idioms, puns, holidays, and culturally embedded concepts) on an ordinal 0-3 quality scale; segment ratings additionally included an NA option for untranslated segments. Across full-text evaluations, mean overall quality is modest (1.68/3): GPT-5 (2.10/3), Claude Sonnet 3.7 (1.97/3), and Mistral Medium 3.1 (1.84/3) form the strongest tier with fewer catastrophic failures. Segment-level results show sharp category effects: holidays (2.20/3) and cultural concepts (2.19/3) translate substantially better than idioms (1.65/3) and puns (1.45/3), and idioms are most likely to be left untranslated. These findings demonstrate a persistent gap between grammatical adequacy and cultural resonance. To our knowledge, this is the first multilingual, human-annotated benchmark focused explicitly on cultural nuance in translation and localisation, highlighting the need for culturally informed training data, improved cross-lingual pragmatics, and evaluation paradigms that better reflect real-world communicative competence.

Large language models (LLMs) reason over discrete token ID sequences, yet modern subword tokenizers routinely produce non-unique encodings: multiple token ID sequences can detokenize to identical surface strings. This representational mismatch creates an unmeasured fragility wherein reasoning processes can fail. LLMs may treat two internal representations as distinct "words" even when they are semantically identical at the text level. In this work, we show that tokenization can betray LLM reasoning through one-to-many token ID mappings. We introduce a tokenization-consistency probe that requires models to replace designated target words in context while leaving all other content unchanged. The task is intentionally simple at the surface level, enabling us to attribute failures to tokenizer-detokenizer artifacts rather than to knowledge gaps or parameter limitations. Through analysis of over 11000 replacement trials across state-of-the-art open-source LLMs, we find a non-trivial rate of outputs exhibit phantom edits: cases where models operate under the illusion of correct reasoning, a phenomenon arising from tokenizer-induced representational defects. We further analyze these cases and provide a taxonomy of eight systematic tokenizer artifacts, including whitespace-boundary shifts and intra-word resegmentation. These findings indicate that part of apparent reasoning deficiency originates in the tokenizer layer, motivating tokenizer-level remedies before incurring the cost of training ever-larger models on ever-larger corpora.

Many commercial Large Language Models (LLMs) are often closed-source, limiting developers to prompt tuning for aligning content generation with specific applications. While these models currently do not provide access to token logits, we argue that if such access were available, it would enable more powerful adaptation techniques beyond prompt engineering. In this paper, we propose a token-level probability reweighting framework that, given access to logits and a small amount of task-specific data, can effectively steer black-box LLMs toward application-specific content generation. Our approach views next-token prediction through the lens of supervised classification. We show that aligning black-box LLMs with task-specific data can be formulated as a label noise correction problem, leading to Plugin model -- an autoregressive probability reweighting model that operates solely on logits. We provide theoretical justification for why reweighting logits alone is sufficient for task adaptation. Extensive experiments with multiple datasets, LLMs, and reweighting models demonstrate the effectiveness of our method, advocating for broader access to token logits in closed-source models.

Academic survey writing, which distills vast literature into a coherent and insightful narrative, remains a labor-intensive and intellectually demanding task. While recent approaches, such as general DeepResearch agents and survey-specialized methods, can generate surveys automatically (a.k.a. LLM4Survey), their outputs often fall short of human standards and there lacks a rigorous, reader-aligned benchmark for thoroughly revealing their deficiencies. To fill the gap, we propose a fine-grained, quiz-driven evaluation framework SurveyBench, featuring (1) typical survey topics source from recent 11,343 arXiv papers and corresponding 4,947 high-quality surveys; (2) a multifaceted metric hierarchy that assesses the outline quality (e.g., coverage breadth, logical coherence), content quality (e.g., synthesis granularity, clarity of insights), and non-textual richness; and (3) a dual-mode evaluation protocol that includes content-based and quiz-based answerability tests, explicitly aligned with readers' informational needs. Results show SurveyBench effectively challenges existing LLM4Survey approaches (e.g., on average 21% lower than human in content-based evaluation).

Reliable application of machine learning is of primary importance to the practical deployment of deep learning methods. A fundamental challenge is that models are often unreliable due to overconfidence. In this paper, we estimate a model's reliability by measuring the agreement between its latent space, and the latent space of a foundation model. However, it is challenging to measure the agreement between two different latent spaces due to their incoherence, \eg, arbitrary rotations and different dimensionality. To overcome this incoherence issue, we design a neighborhood agreement measure between latent spaces and find that this agreement is surprisingly well-correlated with the reliability of a model's predictions. Further, we show that fusing neighborhood agreement into a model's predictive confidence in a post-hoc way significantly improves its reliability. Theoretical analysis and extensive experiments on failure detection across various datasets verify the effectiveness of our method on both in-distribution and out-of-distribution settings.

Existing benchmarks measure capability -- whether a model succeeds on a single attempt -- but production deployments require reliability -- consistent success across repeated attempts on tasks of varying duration. We show these properties diverge systematically as task duration grows, and that pass@1 on short tasks is structurally blind to this divergence. We introduce a reliability science framework for long-horizon LLM agents with four metrics: Reliability Decay Curve (RDC), Variance Amplification Factor (VAF), Graceful Degradation Score (GDS), and Meltdown Onset Point (MOP). We evaluate 10 models across 23,392 episodes on a 396-task benchmark spanning four duration buckets and three domains. Key findings: (1) reliability decay is domain-stratified -- SE GDS drops from 0.90 to 0.44 while document processing is nearly flat (0.74 to 0.71); (2) VAF bifurcates by capability tier -- high VAF is a capability signature, not an instability signal; (3) capability and reliability rankings diverge substantially, with multi-rank inversions at long horizons; (4) frontier models have the highest meltdown rates (up to 19%) because they attempt ambitious multi-step strategies that sometimes spiral; and (5) memory scaffolds universally hurt long-horizon performance across all 10 models. These results motivate reliability as a first-class evaluation dimension alongside capability.

We consider multi-draft speculative sampling, where the proposal sequences are sampled independently from different draft models. At each step, a token-level draft selection scheme takes a list of valid tokens as input and produces an output token whose distribution matches that of the target model. Previous works have demonstrated that the optimal scheme (which maximizes the probability of accepting one of the input tokens) can be cast as a solution to a linear program. In this work we show that the optimal scheme can be decomposed into a two-step solution: in the first step an importance sampling (IS) type scheme is used to select one intermediate token; in the second step (single-draft) speculative sampling is applied to generate the output token. For the case of two identical draft models we further 1) establish a necessary and sufficient condition on the distributions of the target and draft models for the acceptance probability to equal one and 2) provide an explicit expression for the optimal acceptance probability. Our theoretical analysis also motives a new class of token-level selection scheme based on weighted importance sampling. Our experimental results demonstrate consistent improvements in the achievable block efficiency and token rates over baseline schemes in a number of scenarios.

To explain NLP models a popular approach is to use importance measures, such as attention, which inform input tokens are important for making a prediction. However, an open question is how well these explanations accurately reflect a model's logic, a property called faithfulness. To answer this question, we propose Recursive ROAR, a new faithfulness metric. This works by recursively masking allegedly important tokens and then retraining the model. The principle is that this should result in worse model performance compared to masking random tokens. The result is a performance curve given a masking-ratio. Furthermore, we propose a summarizing metric using relative area-between-curves (RACU), which allows for easy comparison across papers, models, and tasks. We evaluate 4 different importance measures on 8 different datasets, using both LSTM-attention models and RoBERTa models. We find that the faithfulness of importance measures is both model-dependent and task-dependent. This conclusion contradicts previous evaluations in both computer vision and faithfulness of attention literature.

Retrieval-Augmented Generation (RAG) is an effective approach to enhance the factual accuracy of large language models (LLMs) by retrieving information from external databases, which are typically composed of diverse sources, to supplement the limited internal knowledge of LLMs. However, the standard RAG often risks retrieving incorrect information, as it relies solely on relevance between a query and a document, overlooking the heterogeneous reliability of these sources. To address this issue, we propose Reliability-Aware RAG (RA-RAG), a new multi-source RAG framework that estimates the reliability of sources and leverages this information to prioritize highly reliable and relevant documents, ensuring more robust and accurate response generation. Specifically, RA-RAG first estimates source reliability by cross-checking information across multiple sources. It then retrieves documents from the top-kappa reliable and relevant sources and aggregates their information using weighted majority voting (WMV), where the selective retrieval ensures scalability while not compromising the performance. Comprehensive experiments show that RA-RAG consistently outperforms baselines in scenarios with heterogeneous source reliability while scaling efficiently as the number of sources increases. Furthermore, we demonstrate the ability of RA-RAG to estimate real-world sources' reliability, highlighting its practical applicability. Our code and data are available at \href{https://github.com/ml-postech/RA-RAG{RA-RAG}.}

Multi-agent debates have been introduced to improve the accuracy of Large Language Models (LLMs) by having multiple agents discuss solutions to a problem over several rounds of debate. However, models often generate incorrect yet confident-sounding responses, which can mislead others. This issue arises partly because agents do not consider how confident their peers are. To address this, we propose DebUnc, a debate framework that uses uncertainty metrics to assess agent confidence. Confidence is then conveyed through a modified attention mechanism that adjusts token weights, or through textual prompts. Evaluations across benchmarks show that attention-based methods are particularly effective and that performance continues to improve as uncertainty estimation becomes more reliable. The code is available at https://github.com/lukeyoffe/debunc.

Scaling the context size of large language models (LLMs) enables them to perform various new tasks, e.g., book summarization. However, the memory cost of the Key and Value (KV) cache in attention significantly limits the practical applications of LLMs. Recent works have explored token pruning for KV cache reduction in LLMs, relying solely on attention scores as a token importance indicator. However, our investigation into value vector norms revealed a notably non-uniform pattern questioning their reliance only on attention scores. Inspired by this, we propose a new method: Value-Aware Token Pruning (VATP) which uses both attention scores and the ell_{1} norm of value vectors to evaluate token importance. Extensive experiments on LLaMA2-7B-chat and Vicuna-v1.5-7B across 16 LongBench tasks demonstrate VATP's superior performance.

Existing benchmarks for tool-using LLM agents primarily report single-run success rates and miss reliability properties required in production. We introduce ReliabilityBench, a benchmark for evaluating agent reliability across three dimensions: (i) consistency under repeated execution using pass^k, (ii) robustness to semantically equivalent task perturbations at intensity ε, and (iii) fault tolerance under controlled tool/API failures at intensity λ. ReliabilityBench contributes a unified reliability surface R(k,ε,λ), action metamorphic relations that define correctness via end-state equivalence rather than text similarity, and a chaos-engineering-style fault injection framework (timeouts, rate limits, partial responses, schema drift). We evaluate two models (Gemini 2.0 Flash, GPT-4o) and two agent architectures (ReAct, Reflexion) across four domains (scheduling, travel, customer support, e-commerce) over 1,280 episodes. Perturbations alone reduce success from 96.9% at ε=0 to 88.1% at ε=0.2. Rate limiting is the most damaging fault in ablations. ReAct is more robust than Reflexion under combined stress, and Gemini 2.0 Flash achieves comparable reliability to GPT-4o at much lower cost. ReliabilityBench provides a systematic framework for assessing production readiness of LLM agents.

Reliable evaluation is essential for developing and deploying large language models, yet in practice it often requires substantial manual effort: practitioners must identify appropriate benchmarks, reproduce heterogeneous evaluation codebases, configure dataset schema mappings, and interpret aggregated metrics. To address these challenges, we present One-Eval, an agentic evaluation system that converts natural-language evaluation requests into executable, traceable, and customizable evaluation workflows. One-Eval integrates (i) NL2Bench for intent structuring and personalized benchmark planning, (ii) BenchResolve for benchmark resolution, automatic dataset acquisition, and schema normalization to ensure executability, and (iii) Metrics \& Reporting for task-aware metric selection and decision-oriented reporting beyond scalar scores. The system further incorporates human-in-the-loop checkpoints for review, editing, and rollback, while preserving sample evidence trails for debugging and auditability. Experiments show that One-Eval can execute end-to-end evaluations from diverse natural-language requests with minimal user effort, supporting more efficient and reproducible evaluation in industrial settings. Our framework is publicly available at https://github.com/OpenDCAI/One-Eval.

Generative recommender systems have recently attracted attention by formulating next-item prediction as an autoregressive sequence generation task. However, most existing methods optimize standard next-token likelihood and implicitly treat all tokens as equally informative, which is misaligned with semantic-ID-based generation. Accordingly, we propose two complementary information-gain-based token-weighting strategies tailored to generative recommendation with semantic IDs. Front-Greater Weighting captures conditional semantic information gain by prioritizing early tokens that most effectively reduce candidate-item uncertainty given their prefixes and encode coarse semantics. Frequency Weighting models marginal information gain under long-tailed item and token distributions, upweighting rare tokens to counteract popularity bias. Beyond individual strategies, we introduce a multi-target learning framework with curriculum learning that jointly optimizes the two token-weighted objectives alongside standard likelihood, enabling stable optimization and adaptive emphasis across training stages. Extensive experiments on benchmark datasets show that our method consistently outperforms strong baselines and existing token-weighting approaches, with improved robustness, strong generalization across different semantic-ID constructions, and substantial gains on both head and tail items. Code is available at https://github.com/CHIUWEINING/Token-Weighted-Multi-Target-Learning-for-Generative-Recommenders-with-Curriculum-Learning.

Reinforcement learning with verifiable rewards (RLVR) has become a central paradigm for improving reasoning and code generation in large language models, and GRPO-style training is widely adopted for its simplicity and effectiveness. However, an important design choice remains underexplored: how token-level policy gradient terms are aggregated within each sampled group. Standard GRPO uses sequence aggregation, while recent work has advocated token aggregation as a better alternative. We show that these two rules induce different optimization biases: token aggregation introduces sign-length coupling, while sequence aggregation implicitly downweights longer responses through sequence-level equal weighting. To address this tension, we propose Balanced Aggregation (BA), a simple drop-in replacement that computes token-level means separately within the positive and negative subsets and then combines them with sequence-count-based weights. Experiments with Qwen2.5-Math-7B and Qwen3-1.7B on DAPO-17k and Polaris, evaluated on six reasoning and coding benchmarks, show that BA consistently improves training stability and final performance over standard token and sequence aggregation. Our analysis further shows that the relative effectiveness of token and sequence aggregation is largely governed by response-length variation and the positive-negative length gap, highlighting aggregation as a critical design dimension in GRPO-style RLVR.

Existing benchmarks do not test language agents on their interaction with human users or ability to follow domain-specific rules, both of which are vital for deploying them in real world applications. We propose tau-bench, a benchmark emulating dynamic conversations between a user (simulated by language models) and a language agent provided with domain-specific API tools and policy guidelines. We employ an efficient and faithful evaluation process that compares the database state at the end of a conversation with the annotated goal state. We also propose a new metric (pass^k) to evaluate the reliability of agent behavior over multiple trials. Our experiments show that even state-of-the-art function calling agents (like gpt-4o) succeed on <50% of the tasks, and are quite inconsistent (pass^8 <25% in retail). Our findings point to the need for methods that can improve the ability of agents to act consistently and follow rules reliably.

Multiple choice questions (MCQs) are commonly used to evaluate the capabilities of large language models (LLMs). One common way to evaluate the model response is to rank the candidate answers based on the log probability of the first token prediction. An alternative way is to examine the text output. Prior work has shown that first token probabilities lack robustness to changes in MCQ phrasing, and that first token probabilities do not match text answers for instruction-tuned models. Therefore, in this paper, we investigate the robustness of text answers. We show that the text answers are more robust to question perturbations than the first token probabilities, when the first token answers mismatch the text answers. The difference in robustness increases as the mismatch rate becomes greater. As the mismatch reaches over 50\%, the text answer is more robust to option order changes than the debiased first token probabilities using state-of-the-art debiasing methods such as PriDe. Our findings provide further evidence for the benefits of text answer evaluation over first token probability evaluation.

Generative Recommendation (GR) has emerged as a transformative paradigm that reformulates the traditional cascade ranking system into a sequence-to-item generation task, facilitated by the use of discrete Semantic IDs (SIDs). However, current SIDs are suboptimal as the indexing objectives (Stage 1) are misaligned with the actual recommendation goals (Stage 2). Since these identifiers remain static (Stage 2), the backbone model lacks the flexibility to adapt them to the evolving complexities of user interactions. Furthermore, the prevailing strategy of flattening hierarchical SIDs into token sequences leads to sequence length inflation, resulting in prohibitive computational overhead and inference latency. To address these challenges, we propose IntRR, a novel framework that integrates objective-aligned SID Redistribution and structural Length Reduction. By leveraging item-specific Unique IDs (UIDs) as collaborative anchors, this approach dynamically redistributes semantic weights across hierarchical codebook layers. Concurrently, IntRR handles the SID hierarchy recursively, eliminating the need to flatten sequences. This ensures a fixed cost of one token per item. Extensive experiments on benchmark datasets demonstrate that IntRR yields substantial improvements over representative generative baselines, achieving superior performance in both recommendation accuracy and efficiency.

Estimating uncertainty in Large Language Models (LLMs) is important for properly evaluating LLMs, and ensuring safety for users. However, prior approaches to uncertainty estimation focus on the final answer in generated text, ignoring intermediate steps that might dramatically impact the outcome. We hypothesize that there exist key forking tokens, such that re-sampling the system at those specific tokens, but not others, leads to very different outcomes. To test this empirically, we develop a novel approach to representing uncertainty dynamics across individual tokens of text generation, and applying statistical models to test our hypothesis. Our approach is highly flexible: it can be applied to any dataset and any LLM, without fine tuning or accessing model weights. We use our method to analyze LLM responses on 7 different tasks across 4 domains, spanning a wide range of typical use cases. We find many examples of forking tokens, including surprising ones such as punctuation marks, suggesting that LLMs are often just a single token away from saying something very different.

Chains of thought (CoTs) have become central in interpreting and auditing behaviors of large language models. Yet growing evidence suggests that these traces often fail to faithfully represent the computations behind a model's predictions. Several faithfulness metrics have been proposed, but whether they indeed measure faithfulness remains unknown. Answering this requires ground-truth labels, which are hard to obtain since internal computations are not directly observable. Consequently, most works proposing metrics report only absolute scores or comparisons to prior metrics, and the few existing benchmarks rely on proxies like plausibility or importance, properties orthogonal to faithfulness that can mislead about whether a CoT can be trusted. We address this challenge by constructing tasks whose outputs reveal which intermediate computations must have produced them, and developing an automated labeling pipeline that yields ground-truth faithfulness labels at both the step and CoT level. Building on this methodology, we present BonaFide, a benchmark of 3,066 labeled CoTs across 13 tasks and 10 models, and use it to conduct the first systematic evaluation of prominent faithfulness metrics. Our experiments show that most metrics perform near chance, exhibit strong prediction biases and degrade on longer CoTs. The best metric reaches only 0.70 AUROC at the CoT level while another reaches 0.59 at the step level, with neither transferring across settings, while entailing prohibitively high computational cost. Our results expose fundamental gaps in current faithfulness evaluation and call for the development of more reliable and efficient metrics.

Language models typically tokenize raw text into sequences of subword identifiers from a predefined vocabulary, a process inherently sensitive to typographical errors, length variations, and largely oblivious to the internal structure of tokens-issues we term the curse of tokenization. In this study, we delve into these drawbacks and demonstrate that large language models (LLMs) remain susceptible to these problems. This study systematically investigates these challenges and their impact on LLMs through three critical research questions: (1) complex problem solving, (2) token structure probing, and (3) resilience to typographical variation. Our findings reveal that scaling model parameters can mitigate the issue of tokenization; however, LLMs still suffer from biases induced by typos and other text format variations. Our experiments show that subword regularization such as BPE-dropout can mitigate this issue. We will release our code and data to facilitate further research.

Large language models (LLMs) owe much of their stellar performance to expansive input contexts, yet such verbosity inflates monetary costs, carbon footprint, and inference-time latency. Much of this overhead manifests from the redundant low-utility tokens present in typical prompts, as only a fraction of tokens typically carries the majority of the semantic weight. We address this inefficiency by introducing FrugalPrompt, a novel prompt compression framework for LLMs, which retains only the most semantically significant tokens. Leveraging two state-of-the-art token attribution methods, GlobEnc and DecompX, we assign salience scores to every token in an input sequence, rank them to preserve the top-k% tokens in their original order, and obtain a sparse frugalized prompt. We evaluate the approach across four NLP tasks: Sentiment Analysis, Commonsense QA, Summarization, and Mathematical Reasoning, using a suite of frontier LLMs. For the first three tasks, a 20% prompt reduction incurs only a marginal loss in task performance, demonstrating that contemporary LLMs can reconstruct elided context from high-salience cues. In contrast, performance on mathematical reasoning deteriorates sharply, reflecting a stronger dependence on complete token continuity. Further analysis with bottom-k% and random-k% tokens reveals asymmetric performance patterns that may suggest potential task contamination effects, wherein models may resort to shallow memorized patterns from pretraining exposure for conventional NLP tasks. We posit that our work contributes to a more nuanced understanding of LLM behavior in performance-efficiency trade-offs, and delineate the boundary between tasks tolerant to contextual sparsity and those requiring exhaustive context. Our source code and models are available at: https://github.com/Starscream-11813/Frugal-ICL.

Claim-level Uncertainty Quantification (UQ) is a promising approach to mitigate the lack of reliability in Large Language Models (LLMs). We introduce MUCH, the first claim-level UQ benchmark designed for fair and reproducible evaluation of future methods under realistic conditions. It includes 4,873 samples across four European languages (English, French, Spanish, and German) and four instruction-tuned open-weight LLMs. Unlike prior claim-level benchmarks, we release 24 generation logits per token, facilitating the development of future white-box methods without re-generating data. Moreover, in contrast to previous benchmarks that rely on manual or LLM-based segmentation, we propose a new deterministic algorithm capable of segmenting claims using as little as 0.2% of the LLM generation time. This makes our segmentation approach suitable for real-time monitoring of LLM outputs, ensuring that MUCH evaluates UQ methods under realistic deployment constraints. Finally, our evaluations show that current methods still have substantial room for improvement in both performance and efficiency.

Language models have graduated from being research prototypes to commercialized products offered as web APIs, and recent works have highlighted the multilingual capabilities of these products. The API vendors charge their users based on usage, more specifically on the number of ``tokens'' processed or generated by the underlying language models. What constitutes a token, however, is training data and model dependent with a large variance in the number of tokens required to convey the same information in different languages. In this work, we analyze the effect of this non-uniformity on the fairness of an API's pricing policy across languages. We conduct a systematic analysis of the cost and utility of OpenAI's language model API on multilingual benchmarks in 22 typologically diverse languages. We show evidence that speakers of a large number of the supported languages are overcharged while obtaining poorer results. These speakers tend to also come from regions where the APIs are less affordable to begin with. Through these analyses, we aim to increase transparency around language model APIs' pricing policies and encourage the vendors to make them more equitable.

Tokenization is a key component of autoregressive (AR) generative models, converting raw data into more manageable units for modeling. Commonly, tokens describe local information, such as regions of pixels in images or word pieces in text, and AR generation predicts these tokens in a fixed order. A worthwhile question is whether token structures affect the ability to steer the generation through test-time search, where multiple candidate generations are explored and evaluated by a verifier. Using image generation as our testbed, we hypothesize that recent 1D ordered tokenizers with coarse-to-fine structure can be more amenable to search than classical 2D grid structures. This is rooted in the fact that the intermediate states in coarse-to-fine sequences carry semantic meaning that verifiers can reliably evaluate, enabling effective steering during generation. Through controlled experiments, we find that AR models trained on coarse-to-fine ordered tokens exhibit improved test-time scaling behavior compared to grid-based counterparts. Moreover, we demonstrate that, thanks to the ordered structure, pure test-time search over token sequences (i.e., without training an AR model) can perform training-free text-to-image generation when guided by an image-text verifier. Beyond this, we systematically study how classical search algorithms (best-of-N, beam search, lookahead search) interact with different token structures, as well as the role of different verifiers and AR priors. Our results highlight the impact of token structure on inference-time scalability and provide practical guidance for test-time scaling in AR models.

While Large Language Models (LLMs) have demonstrated remarkable potential in natural language generation and instruction following, a persistent challenge lies in their susceptibility to "hallucinations", which erodes trust in their outputs. Although Uncertainty Quantification (UQ) presents a promising solution, its accurate implementation within the context of LLMs remains a significant hurdle. To address this critical roadblock, our research originates from a fundamental heuristic insight: tokens within auto-regressive LLM-generated text do not equally reflect the underlying meaning. Some tokens carry greater relevance and representativeness than others, owing to the phenomenon of "linguistic redundancy", wherein a select few keywords suffice to convey the essence of lengthy sentences. Regrettably, existing methodologies treat all tokens with equal importance when estimating uncertainty, disregarding these inherent generative inequalities. Our analysis reveals a significant issue with state-of-the-art: numerous tokens (and sentences) of limited semantic significance receive equal or even excessive weighting during uncertainty estimation. To rectify this bias, we propose to jointly Shifting Attention to more Relevant (SAR) components, at both the token- and the sentence-levels for accurate uncertainty estimation. We conduct extensive experiments involving a range of popular "off-the-shelf" LLMs, including instruction-tuned LLMs such as Vicuna, WizardLM, and LLaMA-2-chat, as well as pretrained LLMs like OPT and LLaMA, with model sizes extending up to 33B parameters. We carry out evaluation across various free-form question-answering tasks, encompassing domains such as reading comprehension, science Q&A, and medical Q&A. Our experimental results demonstrate the superior performance of SAR in addressing the challenges of uncertainty estimation within the realm of LLMs.

Agentic systems are evaluated on benchmarks where agents interact with environments to solve tasks. Most papers report a pass@1 score computed from a single run per task, assuming this gives a reliable performance estimate. We test this assumption by collecting 60,000 agentic trajectories on SWE-Bench-Verified, spanning three models and two scaffolds. We find substantial variance: single-run pass@1 estimates vary by 2.2 to 6.0 percentage points depending on which run is selected, with standard deviations exceeding 1.5 percentage points even at temperature 0. This variance has critical implications: reported improvements of 2--3 percentage points may reflect evaluation noise rather than genuine algorithmic progress. Through token-level analysis, we show that trajectories diverge early, often within the first few percent of tokens, and that these small differences cascade into different solution strategies. To enable reliable evaluation of agentic systems, we recommend three concrete practices: (1) estimate pass@1 from multiple independent runs per task, especially when measuring small improvements, (2) use statistical power analysis to determine the number of runs needed to detect expected effect sizes, and (3) consider metrics like pass@k (optimistic bound) and pass^k (pessimistic bound) with k>1 to better characterize the full performance envelope. While these practices increase evaluation cost, they are essential for distinguishing genuine scientific progress from statistical noise.

Large language models frequently encounter conflicts between their parametric knowledge and contextual input, often resulting in factual inconsistencies or hallucinations. We propose Self-Reflective Debate for Contextual Reliability (SR-DCR), a lightweight framework that integrates token-level self-confidence with an asymmetric multi-agent debate to adjudicate such conflicts. A critic, deprived of context, challenges a defender who argues from the given passage; a judge model evaluates the debate and determines the context's reliability. The final answer is selected by combining the verdict with model confidence. Experiments on the ClashEval benchmark demonstrate that SR-DCR consistently enhances robustness to misleading context while maintaining accuracy on trustworthy inputs, outperforming both classical debate and confidence-only baselines with minimal computational overhead. The code is available at https://github.com/smiles724/Self-Reflective-Debates.

Reinforcement learning from human feedback (RLHF) has been crucial in aligning large language models (LLMs) with human values. Traditionally, RLHF involves generating responses to a query and using a reward model to assign a reward to the entire response. However, this approach faces challenges due to its reliance on a single, sparse reward, which makes it challenging for the model to identify which parts of the sequence contribute most significantly to the final reward. Recent methods have attempted to address this limitation by introducing token-level rewards. However, these methods often rely on either a trained credit assignment model or AI annotators, raising concerns about the quality and reliability of the rewards. In this paper, we propose token-level reward regularization (T-REG), a novel approach that leverages both sequence-level and token-level rewards for preference optimization. Harnessing the self-refinement capabilities of LLMs, our method uses contrastive prompting to enable LLMs to self-generate token-level rewards. These self-generated rewards then act as reward regularization, guiding the model to more effectively distribute sequence-level rewards across tokens. This facilitates better token-level credit assignment and enhances alignment performance. Experiments on the instruction following benchmarks, including Alpaca Eval 2 and Arena-Hard, show that our method consistently outperforms baseline methods by up to 3.8% and 4.4%, respectively. We will release the code and models at https://github.com/wzhouad/T-REG.

Autoregressive Model (AR) has shown remarkable success in conditional image generation. However, these approaches for multiple reference generation struggle with decoupling different reference identities. In this work, we propose the TokenAR framework, specifically focused on a simple but effective token-level enhancement mechanism to address reference identity confusion problem. Such token-level enhancement consists of three parts, 1). Token Index Embedding clusters the tokens index for better representing the same reference images; 2). Instruct Token Injection plays as a role of extra visual feature container to inject detailed and complementary priors for reference tokens; 3). The identity-token disentanglement strategy (ITD) explicitly guides the token representations toward independently representing the features of each identity.This token-enhancement framework significantly augments the capabilities of existing AR based methods in conditional image generation, enabling good identity consistency while preserving high quality background reconstruction. Driven by the goal of high-quality and high-diversity in multi-subject generation, we introduce the InstructAR Dataset, the first open-source, large-scale, multi-reference input, open domain image generation dataset that includes 28K training pairs, each example has two reference subjects, a relative prompt and a background with mask annotation, curated for multiple reference image generation training and evaluating. Comprehensive experiments validate that our approach surpasses current state-of-the-art models in multiple reference image generation task. The implementation code and datasets will be made publicly. Codes are available, see https://github.com/lyrig/TokenAR

Multi-modal generative document parsing systems challenge traditional evaluation: unlike deterministic OCR or layout models, they often produce semantically correct yet structurally divergent outputs. Conventional metrics-CER, WER, IoU, or TEDS-misclassify such diversity as error, penalizing valid interpretations and obscuring system behavior. We introduce SCORE (Structural and COntent Robust Evaluation), an interpretation-agnostic framework that integrates (i) adjusted edit distance for robust content fidelity, (ii) token-level diagnostics to distinguish hallucinations from omissions, (iii) table evaluation with spatial tolerance and semantic alignment, and (iv) hierarchy-aware consistency checks. Together, these dimensions enable evaluation that embraces representational diversity while enforcing semantic rigor. Across 1,114 pages spanning a holistic benchmark and a field dataset, SCORE consistently revealed cross-dataset performance patterns missed by standard metrics. In 2-5% of pages with ambiguous table structures, traditional metrics penalized systems by 12-25% on average, leading to distorted rankings. SCORE corrected these cases, recovering equivalence between alternative but valid interpretations. Moreover, by normalizing generative outputs into a format-agnostic representation, SCORE reproduces traditional scores (e.g., table F1 up to 0.93) without requiring object-detection pipelines, demonstrating that generative parsing alone suffices for comprehensive evaluation. By exposing how interpretive diversity impacts evaluation outcomes and providing multi-dimensional, interpretable diagnostics, SCORE establishes foundational principles for semantically grounded, fair, and practical benchmarking of modern document parsing systems.

Leveraging Large Language Models (LLMs) for generative recommendation has attracted significant research interest, where item tokenization is a critical step. It involves assigning item identifiers for LLMs to encode user history and generate the next item. Existing approaches leverage either token-sequence identifiers, representing items as discrete token sequences, or single-token identifiers, using ID or semantic embeddings. Token-sequence identifiers face issues such as the local optima problem in beam search and low generation efficiency due to step-by-step generation. In contrast, single-token identifiers fail to capture rich semantics or encode Collaborative Filtering (CF) information, resulting in suboptimal performance. To address these issues, we propose two fundamental principles for item identifier design: 1) integrating both CF and semantic information to fully capture multi-dimensional item information, and 2) designing order-agnostic identifiers without token dependency, mitigating the local optima issue and achieving simultaneous generation for generation efficiency. Accordingly, we introduce a novel set identifier paradigm for LLM-based generative recommendation, representing each item as a set of order-agnostic tokens. To implement this paradigm, we propose SETRec, which leverages CF and semantic tokenizers to obtain order-agnostic multi-dimensional tokens. To eliminate token dependency, SETRec uses a sparse attention mask for user history encoding and a query-guided generation mechanism for simultaneous token generation. We instantiate SETRec on T5 and Qwen (from 1.5B to 7B). Extensive experiments demonstrate its effectiveness under various scenarios (e.g., full ranking, warm- and cold-start ranking, and various item popularity groups). Moreover, results validate SETRec's superior efficiency and show promising scalability on cold-start items as model sizes increase.

Uncertainty quantification (UQ) for open-ended language generation remains a critical yet underexplored challenge, especially under black-box constraints where internal model signals are inaccessible. In this paper, we introduce Token-Entropy Conformal Prediction (TECP), a novel framework that leverages token-level entropy as a logit-free, reference-free uncertainty measure and integrates it into a split conformal prediction (CP) pipeline to construct prediction sets with formal coverage guarantees. Unlike existing approaches that rely on semantic consistency heuristics or white-box features, TECP directly estimates epistemic uncertainty from the token entropy structure of sampled generations and calibrates uncertainty thresholds via CP quantiles to ensure provable error control. Empirical evaluations across six large language models and two benchmarks (CoQA and TriviaQA) demonstrate that TECP consistently achieves reliable coverage and compact prediction sets, outperforming prior self-consistency-based UQ methods. Our method provides a principled and efficient solution for trustworthy generation in black-box LLM settings.

Rubric-based Reinforcement Learning (RL) has emerged as a promising approach for aligning Large Language Models (LLMs) with complex, open-domain instruction following tasks. However, existing methods predominantly rely on response-level rewards, introducing severe reward sparsity and reward ambiguity problems. To address these issues, we propose Rubrics to Tokens (RTT), a novel rubric-based RL framework that bridges coarse response-level scores and fine-grained token-level credit assignment. RTT introduces a Token-Level Relevance Discriminator to predict which tokens in the response are responsible for a specific constraint, and optimizes the policy model via RTT-GRPO, which integrates response-level and token-level advantages within a unified framework. Furthermore, when transitioning from one-dimensional, outcome-level reward to three-dimensional reward space in the token-level rubric-based RL, we propose a novel group normalization method, called Intra-sample Token Group Normalization, to accommodate this shift. Extensive experiments and benchmarks demonstrate that RTT consistently outperforms other baselines in both instruction- and rubric-level accuracy across different models.

Different large language models (LLMs) exhibit diverse strengths and weaknesses, and LLM ensemble serves as a promising approach to integrate their complementary capabilities. Despite substantial progress in improving ensemble quality, limited attention has been paid to the robustness of ensembles against potential erroneous signals, which often arise from heterogeneous tokenization schemes and varying model expertise. Our analysis shows that ensemble failures typically arise from both the token level and the model level: the former reflects severe disagreement in token predictions, while the latter involves low confidence and pronounced disparities among models. In light of this, we propose CoRE, a plug-and-play technique that harnesses model consistency for robust LLM ensemble, which can be seamlessly integrated with diverse ensemble methods. Token-level consistency captures fine-grained disagreements by applying a low-pass filter to downweight uncertain tokens with high inconsistency, often due to token misalignment, thereby improving robustness at a granular level. Model-level consistency models global agreement by promoting model outputs with high self-confidence and minimal divergence from others, enhancing robustness at a coarser level. Extensive experiments across diverse benchmarks, model combinations, and ensemble strategies demonstrate that CoRE consistently improves ensemble performance and robustness.