Human evaluation of machine translation is in an arms race with translation model quality: as our models get better, our evaluation methods need to be improved to ensure that quality gains are not lost in evaluation noise. To improve annotation quality, we experiment with a two-stage version of the current state-of-the-art translation evaluation paradigm (MQM), which we call MQM re-annotation. In this setup, an annotator reviews and edits a set of prior MQM annotations that may have come from themselves, another human annotator, or an automatic system. We demonstrate that rater behavior in re-annotation aligns with our goals, and that re-annotation results in higher-quality annotations, mostly due to finding errors that were missed during the first pass.
论文检索
输入标题、作者或关键词，从 100,903 篇学术成果中精准定位
We propose VC-Inspector, a lightweight, open-source large multimodal model (LMM) for reference-free evaluation of video captions, with a focus on factual accuracy. Unlike existing metrics that suffer from limited context handling, weak factuality assessment, or reliance on proprietary services, VC-Inspector offers a reproducible, fact-aware alternative that aligns closely with human judgments. To enable robust training and interpretable evaluation, we introduce a systematic approach for generating captions with controllable errors, paired with graded quality scores and explanatory annotations. Experiments show that VC-Inspector achieves state-of-the-art correlation with human judgments, generalizing across diverse domains (e.g., VATEX-Eval, Flickr8K-Expert, and Flickr8K-CF benchmarks) and revealing the potential for caption improvement.
As large language models (LLMs) evolve from conversational assistants into agents capable of handling complex tasks, they are increasingly deployed in high-risk domains. However, existing benchmarks largely rely on mixed queries and posterior evaluation, output-level scoring, which quantifies hallucination severity but offers limited insight into where and why hallucinations arise in the generation pipeline. We therefore reformulate hallucination evaluation as a diagnostic problem and propose PRISM, a controlled benchmark that disentangles hallucinations into four dimensions: knowledge missing, knowledge errors, reasoning errors, and instruction-following errors, grounded in three stages of generation (memory, instruction, and reasoning). PRISM contains 9,448 instances across 65 tasks and supports fine-grained, stage-aware diagnostic evaluation. Evaluating 24 mainstream open-source and proprietary LLMs, we uncover consistent trade-offs across instruction following, memory retrieval, and logical reasoning, showing that mitigation strategies often improve specific dimensions at the expense of others.We hope PRISM provides a framework for understanding the specific mechanisms behind LLMs hallucinations, ultimately accelerating the development of trustworthy large language models.
There is a growing consensus that, in order to serve as models of human language processing, language models (LMs) need to be constrained in their use of memory for context, the analogue to human working memory (WM). Here we take a novel yet simple approach to constraining WM in language models, in a way that reflects models of human cognition where memory is treated as a limited resource and deployed strategically. In order to capture this constraint on memory encoding, we inject noise into the hidden representations of Transformer-based LMs at tunable rates. Then we train the models with a hybrid objective, such that they learn to maximize the performance of next-word prediction subject to explicit constraints on the total encoding precision. We find that explicit WM constraints improve the model’s alignment with human reading times. More importantly, we find that the need to manage encoding precision reshapes the nature of the models’ context representations, making them more compressed and categorical. Our results show how resource-rational models of WM allocation can be implemented in neural models simply and successfully, and point to a dissociation between WM retrieval mechanisms and the underlying memory representations in models of human sentence processing.
AI Clones aim to simulate an individual’s thoughts and behaviors to enable long-term, personalized interaction, placing stringent demands on memory systems to model experiences, emotions, and opinions over time. Existing memory benchmarks primarily rely on user–agent conversational histories, which are temporally fragmented and insufficient for capturing continuous life trajectories. We introduce CloneMem, a benchmark for evaluating long-term memory in AI Clone scenarios grounded in non-conversational digital traces, including diaries, social media posts, and emails, spanning one to three years. CloneMem adopts a top-down data construction framework to ensure longitudinal coherence and defines tasks that assess an agent’s ability to track evolving personal states. Experiments show that current memory mechanisms struggle in this setting, highlighting open challenges for life-grounded personalized AI. Code and dataset are available at https://github.com/AvatarMemory/CloneMemBench
With the generative capabilities of large language models (LLMs) reshaping the information ecosystem, the concern with the sociological validity of claim detection benchmarks is increasing. Current claim detection benchmarks predominantly treat claims as static textual artifacts, overlooking the sociological etiology of how information naturally emerges and mutates. In this paper, we propose an evolutionary paradigm that models claims as socially evolving entities. In specific, we introduce a socially generative framework for synthetic claim generation, a multi-agent simulation grounded in the Open Claims Model. By decomposing claims into context, utterance, and proposition, our approach enables the precise simulation of unmitigated propagation to capture truth decay, and intervened propagation with multi-auditor oversight for targeted generation. Furthermore, we propose the background-user-perspective (BUP) framework, which reformulates check-worthiness as a condition-dependent probability rooted in social environment. Experiments on our datasets verify the data quality and reveal how network topology and user attributes systematically shape veracity drift.
Overlapping calendar invitations force busy professionals to repeatedly decide which meetings to attend, reschedule, or decline. We refer to this preference-driven decision process as calendar conflict resolution. Automating this decision process is crucial yet challenging. Scheduling logistics can drain hours, and human delegation often fails at scale, which motivates us to ask: Can we trust large language models (LLMs) or language agents to manage time? To enable a systematic study of this question, we introduce CalConflictBench, a benchmark for long-horizon calendar conflict resolution. In CalConflictBench, conflicts are presented to agents round-by-round over a calendar year, requiring them to infer and adapt to user preferences progressively. Our experiments show that current LLM agents perform poorly with high error rates, e.g., Qwen-3-30B-Think has an average error rate of 35%. To address this gap, we propose PEARL, a reinforcement-learning framework that (i) augments the language agent with an external preference memory that stores and updates inferred strategies (e.g., attendee priorities, topic importance, time/location preferences), and (ii) optimizes the agent with round-wise rewards that directly supervise decision correctness, ranking quality, and memory usage across rounds. Experiments on CalConflictBench show that PEARL achieves an error reduction rate of 0.76 and a 55% improvement in average error rate compared to the strongest baseline.
CL^2GEC: A Multi-Discipline Benchmark for Continual Learning in Chinese Literature Grammatical Error Correction
PDF ↗The growing demand for automated writing assistance in diverse academic domains highlights the need for robust Chinese Grammatical Error Correction (CGEC) systems that can adapt across disciplines. However, existing CGEC research largely lacks dedicated benchmarks for multi-disciplinary academic writing, overlooking continual learning (CL) as a promising solution to handle domain-specific linguistic variation and prevent catastrophic forgetting. To fill this crucial gap, we introduce CL^2GEC, the first Continual Learning benchmark for Chinese Literature Grammatical Error Correction, designed to evaluate adaptive CGEC across multiple academic fields. Our benchmark includes 10,000 human-annotated sentences spanning 10 disciplines, each exhibiting distinct linguistic styles and error patterns. CL^2GEC focuses on evaluating grammatical error correction in a continual learning setting, simulating sequential exposure to diverse academic disciplines to reflect real-world editorial dynamics. We evaluate large language models under sequential tuning, parameter-efficient adaptation, and four representative CL algorithms, using both standard GEC metrics and continual learning metrics adapted to task-level variation. Experimental results reveal that regularization-based methods mitigate forgetting more effectively than replay-based or naive sequential approaches. Our benchmark provides a rigorous foundation for future research in adaptive grammatical error correction across diverse academic domains.
Quantifying and Improving the Robustness of Retrieval-Augmented Language Models Against Spurious Features in Grounding Data
PDF ↗Robustness has become a critical attribute for the deployment of RAG systems in real-world applications. Existing research focuses on robustness to explicit noise (e.g., document semantics) but overlooks implicit noise (spurious features). Moreover, previous studies on spurious features in LLMs are limited to specific types (e.g., formats) and narrow scenarios (e.g., ICL). In this work, we identify and study spurious features in the RAG paradigm, a robustness issue caused by the sensitivity of LLMs to semantic-agnostic features. We then propose a novel framework,SURE, to empirically quantify the robustness of RALMs against spurious features. Beyond providing a comprehensive taxonomy and metrics for evaluation, the framework’s data synthesis pipeline facilitates training-based strategies to improve robustness. Further analysis suggests that spurious features are a widespread and challenging problem in the field of RAG. Our code is available at https://anonymous.4open.science/r/RAG-SpuriousFeatures-62B3.
Test-time scaling evaluates reasoning LLMs by sampling multiple outputs per prompt, but ranking models in this regime remains underexplored. We formalize dense benchmark ranking under test-time scaling and introduce Scorio, a library that implements statistical ranking methods such as paired-comparison models, item response theory (IRT) models, voting rules, and graph- and spectral-based methods. Across 20 reasoning models on four Olympiad-style math benchmarks (AIME’24, AIME’25, HMMT’25, and BrUMO’25; up to N = 80 trials), most full-trial rankings agree closely with the Bayesian gold standard Bayes_𝒰@80 (mean Kendall’s τ_b = 0.93–0.95), and 19–34 methods recover exactly the same ordering. In the single-trial regime, the best methods reach τ_b ≈ 0.86.Using greedy decoding as an empirical prior (Bayes_R₀@N) reduces variance at N = 1 by 16–52%, but can bias rankings when greedy and stochastic sampling disagree. These results identify reliable ranking methods for both high- and low-budget test-time scaling. We release Scorio as an open-source library at https://github.com/mohsenhariri/scorio.
LLM-powered multi-agent systems (MAS) have demonstrated strong performance on complex tasks. However, most existing approaches still rely on hand-crafted communication protocols or automatically designed communication topologies, which generalize poorly across tasks. We introduce NeuralFSM, a state-driven framework that formulates multi-agent problem solving as a finite-state execution process. NeuralFSM learns both the state transition distribution and inter-agent communication weights from interaction traces using a Temporal Coordination Controller. Rather than prioritizing explicit structure generation, the proposed framework uses task context to modulate transition and routing decisions, enabling flexible coordination without manual protocol design. To improve robustness against noisy or adversarial agents, we incorporate graph regularization during training and apply trust-aware message attenuation at runtime. Experiments on diverse benchmarks show that NeuralFSM consistently outperforms prior baselines by an average margin of 6.74%–19.39%, while substantially reducing token consumption. Moreover, NeuralFSM exhibits strong inherent robustness, which is further enhanced by the protection layer, resulting in only a 1.82% performance drop under attack.
Anchoring the Cache: Mitigating Contextual Hallucination in KV-Compressed Long-Context Summarization
PDF ↗Key-Value (KV) cache compression techniques have improved the efficiency of long-context summarization in Large Language Models (LLMs), but their impact on model hallucination remains underexplored. In this paper, we present the first systematic study of how KV cache compression affects hallucination in long-context summarization, demonstrating that aggressive compression can increase hallucination scores by up to 3.36× compared to the baseline. To mitigate this issue, we propose HalluKV, a decoding-phase strategy that selectively removes generated KV pairs from retrieval heads responsible for retrieving critical information from source context, thereby anchoring their attention on the preserved source information. Our approach maintains computational efficiency while significantly reducing hallucination across multiple models and datasets, achieving up to 5.48 average point reductions on Llama-3-8B-Instruct, enabling more trustworthy long-context summarization.
Most studies on Arabic Named Entity Recognition (NER) have focused on news texts and social media posts, while the large and rich corpus of literary Arabic books has been underrepresented. We introduce AdabNER, the first large-scale nested NER dataset for Modern Standard Arabic (MSA) literary texts, comprising the first 6,000 words annotated from each of 138 books spanning ten literary genres, including history, biography, literary criticism, and travel literature, and covering works from the 1880s to the 2020s. The corpus comprises about 876K tokens, manually annotated using a nested 21 entity tag annotation scheme, yielding 78,530 entity mentions, 18.96% of which are nested. We fine-tuned five pre-trained Arabic BERT encoders in two settings: stratified and leave-book-out, achieving F1 scores of 0.86 and 0.83 with AraBERTv2, respectively. We also evaluated five large language models through few-shot in-context learning, including open-source models and the closed-source Gemini 3 Pro, with Gemini 3 Pro achieving the highest LLM F1 score of 0.59. Supervised results degraded under out-of-domain evaluation; however, joint multi-domain training reduced this gap to less than a 1% F1 loss, demonstrating that domain-diverse training data is key to robust Arabic NER, though broader validation beyond the experiments reported is needed. AdabNER and its annotation guidelines are publicly available at https://doi.org/10.5281/zenodo.19468385.
Mitigating Lost in Multi-turn Conversation via Curriculum RL with Verifiable Accuracy and Abstention Rewards
PDF ↗Large Language Models demonstrate strong capabilities in single-turn instruction following but suffer from Lost-in-Conversation (LiC), a degradation in performance as information is revealed progressively in multi-turn settings. Motivated by the current progress on Reinforcement Learning with Verifiable Rewards (RLVR), we propose Curriculum Reinforcement Learning with Verifiable Accuracy and Abstention Rewards (RLAAR), a framework that encourages models not only to generate correct answers, but also to judge the solvability of questions in the multi-turn conversation setting. Our approach employs a competence-gated curriculum that incrementally increases dialogue difficulty (in terms of instruction shards), stabilizing training while promoting reliability. Using multi-turn, on-policy rollouts and a mixed-reward system, RLAAR teaches models to balance problem-solving with informed abstention, reducing premature answering behaviors that cause LiC. Evaluated on LiC benchmarks, RLAAR significantly mitigates LiC performance decay (62.6% to 75.1%) and improves calibrated abstention rates (33.5% to 73.4%). Together, these results provide a practical recipe for building multi-turn reliable and trustworthy LLMs.
Hallucination remains a major challenge for the safe and trustworthy deployment of large language models (LLMs) in factual content generation. Prior work has explored confidence estimation as an effective approach to hallucination detection, but often relies on post-hoc self-consistency methods that require computationally expensive sampling. Verbalized confidence offers a more efficient alternative, but existing approaches are largely limited to short-form question answering (QA) tasks and do not generalize well to open-ended generation. In this paper, we propose LoVeC (Long-form Verbalized Confidence), a novel reinforcement learning (RL)–based method that trains LLMs to append an on-the-fly numerical confidence score to each generated statement during long-form generation. The confidence score serves as a direct and interpretable signal of the factuality of generation. We introduce two evaluation settings, free-form tagging and iterative tagging, to assess different verbalized confidence estimation methods. Experiments on three long-form QA datasets show that our RL-trained models achieve better calibration and generalize robustly across domains. Also, our method is highly efficient, being 20 \times faster than traditional self-consistency methods while achieving better calibration.
Reinforcement learning (RL) paradigms have demonstrated strong performance on reasoning-intensive tasks such as code generation. However, limited trajectory diversity often leads to diminishing returns, which constrains the achievable performance ceiling. Search-enhanced RL alleviates this issue by introducing structured exploration, which remains constrained by the single-agent policy priors. Meanwhile, leveraging multiple interacting policies can acquire more diverse exploratory signals, but existing approaches are typically decoupled from structured search. We propose MARS^2 (Multi-Agent Reinforced Tree-Search Scaling), a unified RL framework in which multiple independently-optimized agents collaborate within a shared tree-structured search environment. MARS^2 models the search tree as a learnable multi-agent interaction environment, enabling heterogeneous agents to collaboratively generate and refine candidate solutions within a shared search topology. To support effective learning, we introduce a path-level group advantage formulation based on tree-consistent reward shaping, which facilitates effective credit assignment across complex search trajectories. Experiments on code generation benchmarks show that MARS^2 consistently improves performance across diverse model combinations and training settings, demonstrating the effectiveness of coupling multi-agent collaboration with tree search for enhancing reinforcement learning. Our code is publicly available at https://github.com/TsinghuaC3I/MARTI.
Reinforcement learning (RL) has emerged as a powerful paradigm for improving the reasoning capabilities of large language models (LLMs). Despite its success, RL faces fundamental challenges, including low sample efficiency and a strong dependence on the quality of the base model: while some models improve rapidly with limited RL updates, others require substantial training data to achieve meaningful gains. Recent studies suggest that the patterns of thinking tokens play a critical role in RL performance, and that supervised fine-tuning (SFT) on datasets exhibiting desirable reasoning patterns can reduce reliance on base models and better prepare LLMs for RL. However, how to automatically discover such patterns across tasks remains unclear. In this work, we describe thinking token patterns with reasoning primitives and argue that initializing LLMs with diverse, high-quality primitives is crucial for stable and efficient RL training. We propose Tailor, a pipeline that automatically discovers such reasoning primitives and curates SFT datasets to prepare LLMs for RL. Extensive experiments on mathematical and logical reasoning benchmarks demonstrate that Tailor consistently improves downstream RL performance, outperforming strong baselines, including methods with expert domain knowledge.
How Instruction and Reasoning Data shape Post-Training: Data Quality through the Lens of Layer-wise Gradients
PDF ↗As the post-training of large language models (LLMs) advances from instruction-following to complex reasoning tasks, understanding how different data affect finetuning dynamics remains largely unexplored. In this paper, we present a spectral analysis of layer-wise gradients induced by low/high-quality instruction and reasoning data for LLM post-training. Our analysis reveals that widely-studied metrics for data evaluation, e.g., IFD, InsTag, Difficulty, and Reward, can be explained and unified by spectral properties computed from gradients’ singular value decomposition (SVD). Specifically, higher-quality data are usually associated with lower nuclear norms and higher effective ranks. Notably, effective rank exhibits better robustness and resolution than nuclear norm in capturing subtle quality differences. For example, reasoning data achieves substantially higher effective ranks than instruction data, implying richer gradient structures on more complex tasks. Our experiments also highlight that models within the same family share similar gradient patterns regardless of their sizes, whereas different model families diverge significantly. Providing a unified view on the effects of data quality across instruction and reasoning data, this work illuminates the interplay between data quality and training stability, shedding novel insights into developing better data exploration strategies for post-training.
ChemReason-Bench: Benchmarking Large Language Models for Procedural Reasoning in Experimental Chemistry
PDF ↗Experimental protocols in organic synthesis specify not only the intended transformation but also an executable sequence of operations and conditions. While recent language models show strong chemistry knowledge, widely used evaluations remain less diagnostic of procedure-level decision making. In this setting, correctness requires consistent step ordering, feasibility under stated conditions, faithful entity-role grounding, and schema-parseable outputs that can be automatically validated against operational constraints. We present ChemReason-Bench, a human-validated benchmark for verifiable experimental procedure reasoning built on a structured representation with explicit placeholders and a unified schema, enabling automatic checks of many operational constraints. From 500 reactions, we instantiate 7306 benchmark tasks across six complementary formats: ordering, step validation, condition validation, schema-constrained completion, contrastive choice, and evidence-grounded rationalization. We further release a large-scale instantiation of the same templates for downstream adaptation studies, kept disjoint from the evaluation set. Using a unified evaluation protocol, we benchmark diverse open-source, proprietary, and domain-specific models and observe clear variation across the capability surface. We also report controlled adaptation experiments in the appendix, where supervised fine-tuning improves small models, preference optimization adds limited gains in our setting, and a gap remains to the strongest evaluated systems.
How predictable a word is can be generally quantified in two ways: using human responses to the cloze task or using probabilities from language models (LMs). When used as predictors of processing effort, LM probabilities outperform probabilities derived from cloze data. However, it is important to establish that LM probabilities do so for the right reasons, since different predictors can lead to different scientific conclusions about the role of prediction in language comprehension. We present evidence for three hypotheses about the apparent advantage of LM probabilities: not suffering from low resolution, distinguishing semantically similar words, and accurately assigning probabilities to low-frequency words. These results call for efforts to improve the resolution of cloze studies, coupled with experiments on whether human-like prediction is also as sensitive to the fine-grained distinctions made by LM probabilities.