Large language models (LLMs) for code rely on subword tokenizers, such as byte-pair encoding (BPE), learned from mixed natural language text and programming language code but driven by statistics rather than grammar. As a result, semantically identical code snippets can be tokenized differently depending on superficial factors such as whitespace or identifier naming. To measure the impact of this misalignment, we introduce TokDrift, a framework that applies semantic-preserving rewrite rules to create code variants differing only in tokenization. Across nine code LLMs, including large ones with over 30B parameters, even minor formatting changes can cause substantial shifts in model behavior. Layer-wise analysis shows that the issue originates in early embeddings, where subword segmentation fails to capture grammar token boundaries. Our findings identify misaligned tokenization as a hidden obstacle to reliable code understanding and generation, highlighting the need for grammar-aware tokenization for future code LLMs.
论文检索
输入标题、作者或关键词，从 9,460 篇学术成果中精准定位
MathSight: A Benchmark Exploring Have Vision-Language Models Really Seen in University-Level Mathematical Reasoning?
PDF ↗Recent advances in Vision-Language Models (VLMs) have achieved impressive progress in multimodal mathematical reasoning. Yet, howmuch visual information truly contributes to reasoning remains unclear. Existing benchmarks report strong overall performance but seldom isolate the role of the image modality, leaving open whether VLMs genuinely leverage visual understanding or merely depend on linguistic priors. To address this, we present MathSight, a university-level multimodal mathematical reasoning benchmark designed to disentangle and quantify the effect of visual input. Each problem includes multiple visual variants—original, hand-drawn, photocaptured—and a text-only condition for controlled comparison. Experiments on state-of-the-art VLMs reveal a consistent trend: the contribution of visual information diminishes with increasing problem difficulty. Remarkably, Qwen3-VL without any image input surpasses both its multimodal variants and GPT-5, underscoring the need for benchmarks like MathSight to advance genuine vision-grounded reasoning in future models. The project page is available at https://cnu-bot-group.github.io/MathSight/.
Retrieving music using natural language descriptions has improved with contrastive audio–text models such as CLAP, but current systems remain limited to coarse semantic queries. When descriptions specify fine-grained musical attributes such as tempo, key, chord progression, or rhythmic structure, existing models often fail to retrieve the correct audio. We show that this limitation stems from the contrastive learning objective itself: despite being trained on long captions, CLAP-based models effectively utilize only the first few tokens, discarding much of the information encoded in detailed prompts. Then, we propose FIGMA (Fine-Grained Music Retrieval), a multi-view contrastive architecture that addresses this limitation by jointly optimizing global audio–text alignment and frame-level, token-wise alignment. This design enables FIGMA to capture both high-level semantic context and fine-grained musical attributes within a unified representation space. Moreover, we formalize the task of Fine-Grained Music Retrieval and construct Fine-Grained Music Caption dataset (FGMCaps), a large-scale dataset of 380K music–caption pairs for training along with a 10K test set, both annotated with tempo, key, chord progression, beat count, as well as genre and mood. Extensive experiments demonstrate that FIGMA consistently outperforms existing CLAP-based music retrieval models across multiple music retrieval benchmarks, including out-of-domain evaluations, with relative improvements of up to 73.3%.
Large Language Models have shown remarkable potential as autonomous agents, but their effectiveness in knowledge-intensive tasks remains limited by passive knowledge utilization. We introduce KARL (Knowledge-Augmented Reinforcement Learning), a framework that enables LLM agents to dynamically explore structured knowledge sources through multi-turn interactions. Unlike existing retrieval-augmented approaches, KARL empowers agents to proactively decide when and what knowledge to acquire during task execution. Our framework incorporates online reinforcement learning with curiosity-driven reward shaping, explicitly incentivizing knowledge exploration while optimizing tool-use behaviors end-to-end. Extensive evaluation across six structured knowledge benchmarks demonstrates that KARL achieves state-of-the-art performance, with our Qwen2.5-14B-based agent significantly outperforming GPT-4o, Claude-4, and o4-mini on both knowledge graph and database tasks.Source code is available at https://github.com/THUDM/KARL.
Multimodal Large Language Models (MLLMs) face critical privacy challenges due to the indiscriminate memorization of sensitive data. Existing unlearning methods, largely adapted from Euclidean paradigms, suffer from a geometric mismatch: they fail to disentangle specific instances from general concepts, causing catastrophic forgetting or unsafe substitution. We introduce LOTUS (Lorentz Transport for Unlearning Strategies), a framework for surgical semantic pruning within the Lorentz manifold. Leveraging hyperbolic geometry’s hierarchical nature, LOTUS employs an Inverted Entailment Cone Loss to sever the inheritance of sensitive concepts and a Lorentz Transport mechanism to align pruned features within the tangent space, ensuring compatibility with Euclidean backbones via a safety refusal prior. Experiments on MLLMU-Bench with LLaVA and Qwen show that LOTUS significantly outperforms baselines, effectively erasing targeted visual data while preserving general utility.
Understanding a discourse requires tracking entities and the relations that hold between them. While Large Language Models (LLMs) perform well on relational reasoning, the mechanism by which they bind entities, relations, and attributes remains unclear. We study discourse-level relational binding and show that LLMs encode it via a Cell-based Binding Representation (CBR): a low-dimensional linear subspace in which each “cell” corresponds to an entity–relation index pair, and bound attributes are retrieved from the corresponding cell during inference. Using controlled multi-sentence data annotated with entity and relation indices, we identify the CBR subspace by decoding these indices from attribute-token activations with Partial Least Squares regression. Across domains and two model families, the indices are linearly decodable and form a grid-like geometry in the projected space. We further find that context-specific CBR representations are related by translation vectors in activation space, enabling cross-context transfer. Finally, activation patching shows that manipulating this subspace systematically changes relational predictions and that perturbing it disrupts performance, providing causal evidence that LLMs rely on CBR for relational binding.
Toward Robust Evaluation for Multilingual Grammatical Error Correction: Can Large Language Models Replace Human References?
PDF ↗A standard method for evaluating grammatical error correction systems severely underestimates performance, as it compares outputsagainst a small, fixed set of human references, despite the large space of possible valid corrections. Prior research has shown that using aclosest-gold reference – i.e., a human reference generated with respect to the system output rather than the original text – yields more accurate performance estimates. Yet, producing such references for each system individually is costly. We introduce an automated method for generating closest-gold references by prompting a large language model (LLM) with system outputs. We find that performance scores computed using automatic closest-gold references correlate well with human closest-golds, whereas standard reference-based evaluations show weak or no correlation.Building on this insight, we use both fixed human references and closest-gold references generated by Claude and Llama to compare theperformance of supervised models and GPT-4 across 14 benchmarks spanning 12 languages. Consequently, while prior work has shown that GPT-4 appears to lag behind traditional models, we demonstrate that this is due to the failures of the standard evaluation method that systematically underestimates GPT-4 performance more severely than that of supervised models. We show that a more appropriate evaluation approach, based on the closest gold method, reveals that GPT-4 outperforms traditional state-of- the-art models on almost all languages.
Lost in Simulation: LLM-Simulated Users are Unreliable Proxies for Human Users in Agentic Evaluations
PDF ↗Agentic benchmarks increasingly rely on LLM-simulated users to scalably evaluate agent performance, yet the robustness, validity, and fairness of this approach remain unexamined. Through a user study with participants across the United States, India, Kenya, and Nigeria, we investigate whether LLM-simulated users serve as reliable proxies for real human users in evaluating agents on τ-Bench retail tasks. We find that user simulators lack robustness, with agent success rates varying up to 9 percentage points across different user LLMs. Furthermore, simulated users systematically miscalibrate performance, underestimating success on challenging tasks while overestimating moderately difficult ones. African American Vernacular English (AAVE) speakers experience consistently worse success rates and calibration errors than Standard American English (SAE) speakers, with disparities compounding significantly with age. We also find simulated users to be a differentially effective proxy for different populations, performing worst for AAVE and Indian English. Additionally, simulated users introduce conversational artifacts and surface different failure patterns than human users. These findings demonstrate that current evaluation practices risk misrepresenting agent capabilities across diverse user populations and may obscure real-world deployment challenges.
SPARD: Self-Paced Curriculum for RL Alignment via Integrating Reward Dynamics and Data Utility
PDF ↗The evolution of Large Language Models (LLMs) is shifting the focus from single, verifiable tasks toward complex, open-ended real-world scenarios, imposing significant challenges on the post-training phase. In these settings, the scale and complexity of reward systems have grown significantly, transitioning toward multi-objective formulations that encompass a comprehensive spectrum of model capabilities and application contexts. However, traditional methods typically rely on fixed reward weights, ignoring non-stationary learning dynamics and struggling with data heterogeneity across dimensions. To address these issues, we propose SPARD, a framework that establishes an automated, self-paced curriculum by perceiving learning progress to dynamically adjust multi-objective reward weights and data importance, thereby synchronizing learning intent with data utility for optimal performance. Extensive experiments across multiple benchmarks demonstrate that SPARD significantly enhances model capabilities across all domains. Our code is publicly available at https://github.com/USTC-StarTeam/SPARD.
Complex reasoning with Large Language Models (LLMs) demands a careful balance between accuracy and computational cost. Verification is crucial for reliability but faces trade-off: robust process-based verifiers are computationally prohibitive, while fast verifiers lack precision. We introduce flexive, a unified generative verifier designed to navigate this trade-off by dynamically allocating compute between rapid fast thinking and deliberative slow thinking. A key innovation is our training strategy: we use Group Relative Policy Optimization (GRPO) to specifically enhance the reliability of the fast mode. This targeted training generalizes effectively, elevating the slow mode to state-of-the-art open-source performance. To deploy flexive, we propose the solve-detect-verify (SDV) pipeline. Moving beyond static Best-of-N ranking, SDV employs an iterative refinement process that utilizes likelihood-based probing to detect solution completion, curtailing overthinking, and leverages flexive’s feedback for targeted correction. Solve-detect-verify establishes a new open-source state-of-the-art on ProcessBench, outperforming GenPRM-32B while requiring ~2.3x fewer TFLOPS and 15x less training data. On AIME 2024, the full SDV pipeline achieves 83.3% accuracy, surpassing strong baselines while using significantly fewer tokens.
HalluCitation Matters: Revealing the Impact of Hallucinated References with 300 Hallucinated Papers in ACL Conferences
PDF ↗Recently, we have often observed hallucinated citations or references that do not correspond to any existing work in papers under review, preprints, or published papers. Such hallucinated citations pose a serious concern to scientific reliability. When they appear in accepted papers, they may also negatively affect the credibility of conferences. In this study, we refer to hallucinated citations as “HalluCitation” and systematically investigate their prevalence and impact. We analyze all papers published at ACL, NAACL, and EMNLP in 2024 and 2025, including main conference, Findings, and workshop papers. Our analysis reveals that nearly 300 papers contain at least one HalluCitation, most of which were published in 2025. Notably, half of these papers were identified at EMNLP 2025, the most recent conference, indicating that this issue is rapidly increasing. Moreover, more than 100 such papers were accepted as main conference and Findings papers at EMNLP 2025, affecting the credibility.
Characterizing and Evaluating Working Emotion Vocabularies in Multilingual Large Language Models
PDF ↗Prior work evaluating emotion and affective understanding in large language models (LLMs) typically rely on predetermined label sets or focus on a singular evaluation task (e.g., emotion detection). We consider affective states, referring to the much broader variety of terms people use to label their emotional experiences. We evaluate multilingual language models’ understanding of affective states in English and Spanish through three different tasks: 1) _identification_, where models predict an affective state given text, 2) _expression_, where models generate text expressing a given affective state, and 3) _verification_, where models report whether a given term refers to an affective state. We show that performance on one task is not necessarily predictive of performance on another. Using these three tasks, we then begin to explore when and why models struggle to understand particular affective states compared to others. We examine systematic patterns in the affective state terms that are well and poorly understood by models, characterizing the working emotion vocabulary of LLMs.
Sparse Mixture-of-Experts (MoE) models have achieved remarkable scalability, yet they remain vulnerable to hallucinations, particularly when processing long-tail knowledge. We identify that this fragility stems from static Top-k routing: routers tend to favor high-frequency patterns over rare factual associations. Consequently, "specialist experts" possessing critical long-tail knowledge are often assigned low gating scores and remain "dormant"—under-prioritized for specific tokens despite their proven causal importance on other inputs. To address this, we propose Counterfactual Routing (CoR), a training-free inference framework designed to awaken these dormant experts. CoR integrates layer-wise perturbation analysis with the Counterfactual Expert Impact (CEI) metric to dynamically shift computational resources from syntax-dominant to knowledge-intensive layers while maintaining a constant total activation count, effectively retrieving causally decisive experts via virtual ablation. Extensive experiments on TruthfulQA, FACTOR, and TriviaQA demonstrate that CoR improves factual accuracy by 3.1% on average without increasing the inference budget, establishing a superior Pareto frontier compared to static scaling strategies.
One Single Hub Text Breaks CLIP: Identifying Vulnerabilities in Cross-Modal Encoders via Hubness
PDF ↗The hubness problem, in which hub embeddings are close to many unrelated examples, occurs often in high-dimensional embedding spaces and may pose a practical threat for purposes such as information retrieval and automatic evaluation metrics.In particular, since cross-modal similarity between text and images cannot be calculated by direct comparisons, such as string matching, cross-modal encoders that project different modalities into a shared space are helpful for various cross-modal applications, and thus, the existence of hubs may pose practical threats.To reveal the vulnerabilities of cross-modal encoders, we propose a method for identifying the hub embedding and its corresponding hub text.Experiments on image captioning evaluation in MSCOCO and nocaps along with image-to-text retrieval tasks in MSCOCO and Flickr30k showed that our method can identify a single hub text that unreasonably achieves comparable or higher similarity scores than human-written reference captions in many images, thereby revealing the vulnerabilities in cross-modal encoders.
ReasMark: A Robust Watermark for Attributing LLM Reasoning Under Knowledge Distillation Attacks
PDF ↗Reasoning-enhanced large language models rely on intermediate reasoning signals to solve complex, multi-step tasks, making reasoning behavior a valuable form of intellectual property. Meanwhile, knowledge distillation enables an adversary to replicate this behavior in a realistic black-box setting by repeatedly querying a deployed model on a target domain and training a local student to imitate its outputs, including reasoning traces. Existing LLM watermarks primarily operate on surface text and decoding-time token biases, and thus fail to provide reliable attribution of reasoning behavior once it is transferred through knowledge distillation. ReasMark entangles the watermark with the target-domain input distribution by selecting watermark tokens from high-frequency prompts, so distillation queries naturally activate it. It then embeds the watermark by score-conditioned losses that create a detectable reasoning-length gap for black-box verification. Comprehensive experiments across multiple LLMs, datasets, and distillation settings demonstrate that ReasMark consistently outperforms existing baselines while preserving task utility.
Aligning Large Language Models (LLMs) with diverse and potentially conflicting human values necessitates navigating complex multi-objective landscapes. However, existing prompt-conditioned approaches face a critical training-inference discrepancy: they rely on ground-truth scores during training while requiring manual user-specification at inference. We introduce prediction of implicit preferences to bridge this gap while reducing user burden. To this end, we propose Self-Guided Alignment (SGA), a framework that transforms passive reward dependency into an intrinsic adaptive sensing capability. It employs a dual-head architecture to unify preference internalization with conditional generation, enabling the model to learn a latent mapping between raw prompts and preference profiles. Through adaptive preference sensing, the model autonomously predicts the latent preference score to self-guide the generation, thereby eliminating the need for manual specification at inference. Extensive experiments across diverse model scales demonstrate that SGA often outperforms state-of-the-art baselines, achieving superior multi-objective trade-offs and improved preference alignment. Code is available at https://github.com/python-yyds/SGA.
For constructing text embeddings, mean pooling, which averages token embeddings, is the standard approach. This paper examines whether mean pooling actually works well in real models. First, we note that mean pooling can collapse information beyond the first-order statistics of the token embeddings, such as second-order statistics that capture their spatial structure, potentially mapping distinct token embedding distributions to similar text embeddings. Motivated by this concern, we propose a simple metric to quantify such a collapse induced by mean pooling. Then, using this metric, we empirically measure how often this collapse occurs in actual models and texts, and find that modern text encoders are robust to this collapse. In particular, contrastive fine-tuned text encoders tend to be less prone to the collapse than their pretrained backbone models. We also find that the robustness of these text encoders lies in the concentration of token embeddings within each text. In addition, we find that robustness to the collapse, as quantified by our proposed metric, correlates with downstream task performance. Overall, our findings offer a new perspective on why modern text encoders remain effective despite relying on seemingly coarse mean pooling.
From Word Sequences to Behavioral Sequences: Adapting Modeling and Evaluation Paradigms for Longitudinal NLP
PDF ↗While NLP typically treats documents as independent and unordered samples, in longitudinal studies, this assumption rarely holds: documents are nested within authors and ordered in time, forming person-indexed, time-ordered \textit{behavioral sequences}.Here, we demonstrate the need for and propose a longitudinal modeling and evaluation paradigm that consequently updates four parts of the NLP pipeline: (1) evaluation splits aligned to generalization over people (\textit{cross-sectional}) and/or time (\textit{prospective}); (2) accuracy metrics separating between-person differences from within-person dynamics; (3) sequence inputs to incorporate history by default; and (4) model internals that support different \textit{coarseness} of latent state over histories (pooled summaries, explicit dynamics, or interaction-based models).We demonstrate the issues ensued by traditional pipeline and our proposed improvements on a dataset of 17k daily diary transcripts paired with PTSD symptom severity from 238 participants, finding that traditional document-level evaluation can yield substantially different and sometimes reversed conclusions compared to our ecologically valid modeling and evaluation. We tie our results to a broader discussion motivating a shift from word-sequence evaluation toward \textit{behavior-sequence} paradigms for NLP.
High-quality scientific data is critical for advancing LLMs, yet academic literature remains largely underutilized. This work addresses the fundamental question: How can we systematically unlock scientific data’s value for pre-training? First, we construct a large-scale raw scientific corpus but identify a critical Learnability Gap, revealing that direct pre-training yields negligible gains. To bridge this, we develop a multi-stage pipeline featuring content cleaning and pedagogical augmentation, resulting in SciPedia, a 900B-token corpus. Finally, we establish a controlled verification framework: we develop SciPedia-Eval benchmark and conduct 600B tokens of continued pre-training (CPT) starting from transparent base models (3B/7B) trained from scratch. Compared to a CPT baseline trained with general-purpose data, our approach with SciPedia data boosts average performance by +2.12 (3B) and +2.95 (7B), reaching +5.60 and +8.40 on in-domain tasks. This setup further allows us to derive empirical guidelines for data composition and model configurations.
DyBBT: Dynamic Balance via Bandit-inspired Targeting for Dialog Policy with Cognitive Dual Systems
PDF ↗Task oriented dialog systems often rely on static exploration strategies that do not adapt to dynamic dialog contexts, leading to inefficient exploration and suboptimal performance. We propose DyBBT, a novel dialog policy learning framework that formalizes the exploration challenge through a structured cognitive state space \mathcal{C} that captures dialog progression, user uncertainty, and slot dependency. DyBBT proposes a bandit-inspired meta-controller that dynamically switches between a fast intuitive inference (System 1) and a slow deliberative reasoner (System 2) based on real-time cognitive states and visitation counts. Extensive experiments on single- and multi-domain benchmarks show that DyBBT achieves SOTA performance in success rate, efficiency, and generalization, with human evaluations confirming that its decisions are well-aligned with expert judgment.