论文检索

Linqing Chen, Hanmeng Zhong, Wentao Wu, Peng Zhou

Recent advancements in Multimodal Large Language Models (MLLMs) have largely been driven by aligning visual encoders with pre-trained Large Language Models (LLMs). While effective, the geometric nature of this alignment remains under-explored. Existing methods often assume a symmetric interaction between visual and textual modalities, implying that both spaces adapt to each other. In this paper, we challenge this assumption and propose the "Text Space as Anchor" hypothesis. We argue that the semantic space of LLMs is rigid, anisotropic, and dominant; thus, effective cross-modal alignment may be an asymmetric projection of visual features onto this pre-existing text manifold without distorting it. We identify a potential issue in current parameter-efficient tuning paradigms where task-specific visual adjustments inadvertently disrupt the projector’s geometry, leading to "catastrophic forgetting" of the alignment mechanism itself. To address this, we introduce Anchor-Preserving Projection (APP), a novel method that regularizes the projector to maintain the geometric structure of the text embedding space via spectral filtering. Extensive experiments on 8 diverse cross-modal tasks and 3 pure language benchmarks demonstrate that APP preserves the LLM’s inherent linguistic capabilities (e.g., MMLU, GSM8K) and reduces object hallucination significantly better than standard fine-tuning methods. We release our code.

Yuansen Liu, Yixuan Tang, Anthony Kum Hoe Tung

Current LLM safety research predominantly focuses on mitigating **Goal Hijacking**, preventing attackers from redirecting a model’s high-level objective (e.g., from "summarizing emails" to "phishing users"). In this paper, we argue that this perspective is incomplete and highlight a critical vulnerability in **Reasoning Alignment**. We expose the inherent fragility of current alignment techniques by proposing a new adversarial prompt attack paradigm: **Reasoning Hijacking**. To demonstrate this vulnerability, we instantiate it via the **Criteria Attack**, which subverts model judgments by injecting spurious decision criteria without altering the high-level task goal. Unlike Goal Hijacking, which attempts to override the system prompt, Reasoning Hijacking keeps the task goal intact but manipulates the model’s decision-making logic by injecting spurious reasoning shortcuts. Through extensive experiments on three different tasks (toxic comment, negative review, and spam detection), we demonstrate that even state-of-the-art models are highly fragile, consistently prioritizing injected heuristic shortcuts over rigorous semantic analysis. Crucially, because the model’s explicit intent remains aligned with the user’s instructions, these attacks can bypass defenses designed to detect goal deviation (e.g., SecAlign, StruQ), revealing a fundamental blind spot in the current safety landscape. Data and code are available at [https://github.com/Yuan-Hou/criteria_attack](https://github.com/Yuan-Hou/criteria_attack).

Kaixiong Gong, Kaituo Feng, Bohao Li, Yibing Wang, Mofan Cheng, Shijia Yang, Jiaming Han, Benyou Wang, Yutong Bai, Zhuoran Yang 等

Recent multimodal large language models (MLLMs), such as GPT-4o, Gemini 1.5/2.5 Pro, and Reka Core, have advanced audio-visual reasoning capabilities, achieving strong performance in tasks like cross-modal understanding and generation. However, our DeafTest uncovers unanticipated failures: most of the state-of-the-art MLLMs struggle with very simple audio tasks, such as distinguishing louder sounds or sound counting. This raises a fundamental question—does a deficiency in low-level audio perception constrain higher-level audio-visual reasoning? To address this, we introduce AV-Odyssey Bench—a comprehensive benchmark of 4,555 meticulously designed problems that integrate text, audio, and visual modalities. Each task requires models to unify cross-modal reasoning, leveraging synchronized audio-visual cues to infer solutions. By structuring questions as multiple-choice, we ensure objective, reproducible evaluations without reliance on subjective human or LLM-based judgments. Through comprehensive benchmarking of closed-source and open-source models, we showcase: (i) current MLLMs lack robust audio-visual integration ability and (ii) performance on DeafTest (Pearson’s r = 0.945) strongly correlates with AV-Odyssey accuracy. These findings challenge assumptions about models’ multimodal proficiency and highlight fundamental audio perception as a reasoning bottleneck. We believe that our results provide concrete guidance for future dataset design, alignment strategies, and architectures.

Khotso Selialia, Antoine Nzeyimana, Fatima M. Anwar

Multilingual neural machine translation (MNMT) models degrade in performance as input context length increases, causing positional encoding schemes to misinterpret token distances. Existing absolute and relative positional encodings rely on fixed token indices and implicitly assume uniform semantic density, which breaks down for long-context inputs. We introduce DCARPE, a tokenization-aware adaptive positional encoding that conditions relative positional bias on input-level sequence length and fragmentation statistics, allowing the model to reinterpret positional distance when tokenization-induced inflation arises rather than semantic factors. Evaluations on JW300 and out-of-distribution FLORES-200 demonstrate consistent improvements in long-context robustness, achieving gains of up to +10.81 ChrF++ and +8.00 BLEU over baselines.

Shivam Adarsh, Maria Maistro, Christina Lioma

Large Language Models (LLMs) often encode whether a statement is true as a vector in their residual stream activations. These vectors, also known as truth vectors, have been studied in prior work, however how they change when context is introduced remains unexplored. We study this question by measuring (1) the directional change (\theta) between the truth vectors with and without context and (2) the relative magnitude of the truth vectors upon adding context. Across four LLMs and four datasets, we find that (1) truth vectors are roughly orthogonal in early layers, converge in middle layers, and may stabilize or continue increasing in later layers; (2) adding context generally increases the truth vector magnitude, i.e., the separation between true and false representations in the activation space is amplified; (3) larger models distinguish relevant from irrelevant context mainly through directional change (\theta), while smaller models show this distinction through magnitude differences. We also find that context conflicting with parametric knowledge produces larger geometric changes than parametrically aligned context. Collectively, these findings provide a geometric characterization of how context transforms the truth vector in the activation space of LLMs.

Ryo Yoshida, Shinnosuke Isono, Taiga Someya, Yohei Oseki, Tatsuki Kuribayashi

Surprisal theory hypothesizes that the difficulty of human sentence processing increases linearly with surprisal, the negative log-probability of a word given its context. Computational psycholinguistics has tested this hypothesis using language models (LMs) as proxies for human prediction. While surprisal derived from recent neural LMs generally captures human processing difficulty on naturalistic corpora that predominantly consist of simple sentences, it severely underestimates processing difficulty on sentences that require syntactic disambiguation (garden-path effects). This leads to the claim that the processing difficulty of such sentences cannot be reduced to surprisal, although it remains possible that neural LMs simply differ from humans in next-word prediction. In this paper, we investigate whether it is truly impossible to construct a neural LM that can explain garden-path effects via surprisal. Specifically, instead of evaluating off-the-shelf neural LMs, we fine-tune these LMs on garden-path sentences so as to better align surprisal-based reading-time estimates with actual human reading times. Our results show that fine-tuned LMs do not overfit and successfully capture human reading slowdowns on held-out garden-path items; they even improve predictive power for human reading times on naturalistic corpora and preserve their general LM capabilities. These results provide an existence proof for a neural LM that can explain both garden-path effects and naturalistic reading times via surprisal, but also raise a theoretical question: what kind of evidence can truly falsify surprisal theory?

Terry Jingchen Zhang, Gopal Dev, Ning Wang, Max Obreiter, Punya Syon Pandey, Keenan Samway, Wenyuan Jiang, Yinya Huang, Bernhard Schölkopf, Mrinmaya Sachan 等

Post-cutoff performance decay of LLMs has been widely interpreted as a temporal signal for benchmark contamination, where public information released before the training cutoff may have been included into training corpora and inflated model performance by memorization. We critically examine this view and demonstrate that this temporal signal is highly sensitive to how benchmark questions are constructed, even if the underlying source material remains invariant. Specifically, we show that LLM-transformed questions can produce remarkably different temporal patterns compared to fill-in-the-blank (cloze) questions directly retrieved from the very same documents. We validate this effect on prior benchmarks that report clear post-cutoff decay (LiveCodeBench), and show that a simple LLM-driven transformation of the same problems can effectively remove the temporal pattern. We further provide a mechanistic understanding of this phenomenon using influence function analysis. Overall, our results suggest that post-cutoff performance decay is a sensitive contamination signal, motivating more robust contamination probes for reliable LLM evaluation.

Haotian Lu, Yuchen Mou, Bingzhe Wu

Warning: This paper may contain content that could be disturbing or offensive. Content moderation in online platforms faces persistent challenges due to the evolving complexity of user-generated content and the limitations of traditional rule-based and machine learning approaches. While recent advances in large language models (LLMs) have enabled more sophisticated moderation via direct prompting or fine-tuning, these approaches often exhibit limited generalization, interpretability, and adaptability to unseen or ambiguous cases.In this work, we propose a novel moderation framework that leverages analogical examples to enhance rule induction and decision reliability. Our approach integrates end-to-end optimization of analogical retrieval, rule generation, and moderation classification, enabling the dynamic adaptation of moderation rules to diverse content scenarios. Through comprehensive experiments, we demonstrate that our method significantly outperforms both rule-injected fine-tuning baselines and multi-stage static RAG pipelines in terms of moderation accuracy and rule quality. Further evaluations—including human assessments and external model generalization tests confirm the superiority of rules generated by our framework in terms of clarity, interpretability, and applicability. These findings highlight the potential of analogical example-driven methods for advancing robust, explainable, and generalizable content moderation in real-world applications.

Jun Seo Kim, Hyemi Kim, Woo Joo OH, Hongjin Cho, Hochul Lee, Hye Hyeon Kim

Cognitive distortions have been closely linked to mental health disorders, yet their automatic detection remains challenging due to contextual ambiguity, co-occurrence, and semantic overlap. We propose a novel framework that combines Large Language Models (LLMs) with a Multiple-Instance Learning (MIL) architecture to enhance interpretability and expression-level reasoning. Each utterance is decomposed into Emotion, Logic, and Behavior (ELB) components, which are processed by LLMs to infer multiple distortion instances, each with a predicted type, expression, and model-assigned salience score. These instances are integrated via a Multi-View Gated Attention mechanism for final classification. Experiments on Korean (KoACD) and English (Therapist QA) datasets demonstrate that incorporating ELB and LLM-inferred salience scores improves classification performance, especially for distortions with high interpretive ambiguity. Our results suggest a psychologically grounded and generalizable approach for fine-grained reasoning in mental health NLP. The dataset and implementation details are publicly accessible.

Kazuki Tsutsukawa

Keypoint-based action recognition offers robustness to appearance variations and provides privacy-preserving representation. However, existing zero-shot (ZS) approaches largely emphasize human motion while underutilizing contextual information, particularly human–object interactions. Moreover, extending keypoint-based ZS models to few-shot scenarios remains insufficiently explored. We propose Instance-aware Semantic Alignment and Transfer (InsAT), a unified framework for ZS recognition and zero-to-few-shot (Z2F) adaptation that leverages instance-level language descriptions. InsAT aligns textual descriptions of humans, objects, and their interactions with visual representations derived from human and object keypoints, enabling effective transfer of interaction knowledge from seen to unseen action classes. To support Z2F adaptation, we introduce Instance-level Visual Adaptation, a parameter-free mechanism that improves recognition by incorporating instance-level contextual cues without updating model weights. Extensive experiments demonstrate that InsAT substantially outperforms prior keypoint-based ZS methods and achieves competitive performance relative to large vision–language models, while remaining data-efficient and robust.

Afrozah Nadeem, Mark Dras, Usman Naseem

Large Language Models (LLMs) increasingly shape public discourse, yet most evaluations of economic and political bias have focused on high-resource Western languages and contexts. This leaves a blind spots in low-resource, multilingual regions such as Pakistan, where linguistic identity is closely tied to regional, religious, and political ideologies. We present a systematic evaluation of political bias in 13 state-of-the-art LLMs across five Pakistani languages: Urdu, Punjabi, Sindhi, Pashto, and Balochi. Our framework integrates a culturally adapted Political Compass Test (PCT) with multi-level framing analysis, capturing both ideological stance (economic/social axes) and stylistic framing (content, tone, emphasis). The prompts are aligned with 11 socio-political themes specific to the Pakistani context. The results show that while LLMs significantly reflect liberal-left orientations consistent with Western training data, they exhibit more authoritarian framing in regional languages, highlighting language-conditioned ideological modulation. We also identify model-specific bias patterns in all languages. These findings show the need for culturally grounded multilingual bias examining frameworks in NLP. Code and dataset are available.

Yihua Zhu, Qianying Liu, Jiaxin Wang, Fei Cheng, Chaoran Liu, Akiko Aizawa, Sadao Kurohashi, Hidetoshi Shimodaira

Autoregressive LLMs perform well on relational tasks that require linking entities via relational words (e.g., father/son, friend), but it is unclear whether they learn the logical semantics of such relations (e.g., symmetry and inversion logic) and, if so, whether reversal-type failures arise from missing relational semantics or left-to-right order bias. We propose a controlled Knowledge Graph-based synthetic framework that generates text from symmetric/inverse triples, train GPT-style autoregressive models from scratch, and evaluate memorization, logical inference, and in-context generalization to unseen entities to address these questions. We find a sharp phase transition in which relational semantics emerge with sufficient logic-bearing supervision, even in shallow (2–3 layer) models, and that successful generalization aligns with stable intermediate-layer signals. Finally, order-matched forward/reverse tests and a diffusion baseline indicate that reversal failures are primarily driven by autoregressive order bias rather than deficient inversion semantics.

Cuong Chi Le, Minh V.T. Pham, Tung Duy Vu, Cuong Duc Van, Huy N. Phan, Hoang N. Phan, Tien N. Nguyen

Specifications are vital for ensuring program correctness, yet writing them manually remains challenging and time-intensive. Recent large language model (LLM)-based methods have shown successes in generating specifications such as postconditions, but existing single-pass prompting often yields inaccurate results. In this paper, we present SpecMind, a novel framework for postcondition generation that treats LLMs as interactive and exploratory reasoners rather than one-shot generators. SpecMind employs feedback-driven multi-turn prompting approaches, enabling the model to iteratively refine candidate postconditions by incorporating implicit and explicit correctness feedback, while autonomously deciding when to stop. This process fosters deeper code comprehension and improves alignment with true program behavior via exploratory attempts. Our empirical evaluation shows that SpecMind significantly outperforms state-of-the-art approaches in both accuracy and completeness of generated postconditions.

Siddhant Bhambri, Upasana Biswas, Subbarao Kambhampati

Recent advances in reasoning-oriented Large Language Models (LLMs) have been driven by the introduction of Chain-of-Thought (CoT) traces, where models generate intermediate reasoning traces before producing an answer. These traces, as in DeepSeek R1, are not only used to guide model inference but also serve as supervision signals for Knowledge Distillation (KD) to improve smaller models. A prevailing but under-examined implicit assumption is that these CoT traces when emitted at inference time are both semantically correct and interpretable for the end-users. While there are reasons to believe that these intermediate tokens help improve solution accuracy, in this work, we question their validity (semantic correctness) and interpretability to the end user. To isolate the effect of trace semantics, we design experiments in the Question Answering (QA) domain using a rule-based problem decomposition method. This enables us to create Supervised Fine-Tuning (SFT) datasets for LLMs where - each QA problem is paired with either verifiably correct or incorrect CoT traces, while always providing the correct final solution. Trace correctness at inference time is then evaluated by checking the accuracy of every sub-step in decomposed reasoning chains. To assess end-user interpretability, we finetune LLMs with three additional types of CoT traces: R1 traces, R1 trace summaries, and post-hoc explanations of R1 traces. We further conduct a human-subject study with 100 participants asking them to rate the interpretability of each trace type on a standardized Likert scale. Our experiments reveal two key findings - (1) CoT trace correctness is not reliably correlated with the model’s generation of correct final answers: correct traces led to correct solutions only for 28% test-set problems while incorrect traces don’t necessarily degrade solution accuracy. (2) In end-user interpretability studies, fine-tuning on verbose R1 traces produced the best model performance but these traces were rated as least interpretable by users, scoring on average 3.39 for interpretability and 4.59 for cognitive load metrics on a 5-point Likert scale. In contrast, the decomposed traces that are judged significantly more interpretable don’t lead to comparable solution accuracy. Together, these findings challenge the assumption in question suggesting that researchers and practitioners should decouple model supervision objectives from end-user-facing trace design.

Dongjun Kim, Jeongho Yoon, Chanjun Park, Heuiseok Lim

Dense retrieval in multilingual settings often searches over mixed-language collections, yet multilingual embeddings encode language identity alongside semantics. This language signal can inflate similarity for same-language pairs and crowd out relevant evidence written in other languages. We propose LANGSAE EDITING, a post-hoc sparse autoencoder trained on pooled embeddings that enables controllable removal of language-identity signal directly in vector space. The method identifies language-associated latent units using cross-language activation statistics, suppresses these units at inference time, and reconstructs embeddings in the original dimensionality, making it compatible with existing vector databases without retraining the base encoder or re-encoding raw text. Experiments across multiple languages show consistent improvements in ranking quality and cross-language coverage, with especially strong gains for script-distinct languages.

Yijie Li, Xi Cao, Yuan Sun, Quulgan Minggad, Abdulla Ablikim, Jia Qing Cai Wang

Despite the rapid advancement of LLMs, their performance on linguistically and culturally diverse minority languages within a unified national context remains underexplored. We present CMiLBench, a collection of hierarchical multitask benchmarks designed to translate theoretical notions of “diversity in unity” into practical evaluation for three representative Chinese minority languages: Tibetan, Mongolian, and Uyghur. CMiLBench comprises 24,663 instances across 5 difficulty levels and 17 tasks spanning foundational ability, cultural specificity, and safety alignment. We adopt existing dataset adaptation, minority knowledge construction, and high-resource benchmark translation to construct CMiLBench. We assess 14 state-of-the-art commercial and open-source LLMs with a hybrid framework that integrates automatic metrics and LLM-as-a-Judge scoring. The comparative experimental results reveal the gap between theoretical capability and practical utility. CMiLBench serves as a foundational and scalable evaluation resource to bridge the digital language divide and promote the informatization and intelligentization of low-resource Chinese minority languages.

Haoyue Zhang, Hualei Zhang, Xiaosong Ma, Jie Zhang, Song Guo

Large Language Models (LLMs) exhibit enhanced capabilities by Chain-of-Thought reasoning. However, the extended reasoning sequences introduce significant GPU memory overhead due to increased key-value (KV) cache. Existing KV cache compression methods mitigate memory bottlenecks but struggle in long reasoning tasks. In this paper, we analyze attention patterns in reasoning tasks and reveal a **Token Importance Recurrence** phenomenon: a large proportion of tokens regain high attention after multiple decoding steps, which is failed to capture by existing works and may lead to unpredictable eviction on such periodically critical tokens. To address this, we propose **LazyEviction**, an observation window-based lagged eviction framework retaining latent recurring tokens by prioritized eviction based on tokens’ recurrence patterns. Extensive experiments demonstrate that LazyEviction reduces KV cache by 50% 70% while maintaining comparable accuracy, outperforming existing KV cache baselines. Our implementation code can be found at https://github.com/Halo-949/LazyEviction.

Jialiang Guo, Fucheng Xiong, Xu He, Haodong Zhao, Xingyang li, Ke Zeng, Xunliang Cai

Reinforcement Learning with Verifiable Rewards (RLVR) has become the standard paradigm for enhancing reasoning capabilities in Large Language Models, yet on-policy algorithms like GRPO suffer from sample inefficiency. Current experience replay methods for RLVR typically replay correct trajectories to consolidate learned reasoning patterns and accelerate convergence, but overlook the vast failure space. This work investigates how to effectively replay failure trajectories. We find that the high heterogeneity of failures renders random replay ineffective, and that high-value negatives should be both gradient-efficient and structurally proximal to correct solutions. To this end, we propose NexGRPO, which employs mid-confidence gating to filter invalid noise and saturated errors, and utilizes boundary failure sampling to retrieve boundary errors semantically similar to correct solutions for targeted refinement. Extensive experiments on mathematical and general reasoning benchmarks demonstrate that NexGRPO outperforms strong baaselines and achieves improved out-of-distribution generalization.

Ziping Ye, Gourab Dey, Christos Christodoulopoulos, Charith Peris, Anil Ramakrishna, Weitong Ruan, Aram Galstyan, Kai-Wei Chang, Rahul Gupta, Ninareh Mehrabi

We introduce SWAN (Semantic Watermarking with Abstract Meaning Representation), a novel framework that embeds watermark signatures into the semantic structure of a sentence using Abstract Meaning Representation (AMR). In contrast to existing watermarking methods, which typically encode signatures by adjusting token selection preferences during text generation, SWAN embeds the signature directly in the sentence’s semantic representation. As the signature is encoded at the semantic structure level, any paraphrase that preserves meaning, automatically preserves the signature. SWAN is training-free: watermark injection is achieved by prompting an LLM to generate sentences guided by a selected AMR template while maintaining contextual coherence, and detection uses an off-the-shelf AMR parser followed by a simple one-proportion z-test. Empirical evaluation on the RealNews benchmark shows SWAN matches state-of-the-art detection performance on unaltered watermarked text, while significantly improving robustness against paraphrasing, increasing detection AUC by up to 13.9 percentage points compared to prior methods. These results demonstrate that SWAN’s approach of anchoring watermarks in AMR semantic structures provides a simple, effective, and prompt-based method for robust text provenance verification under paraphrasing, opening new avenues for semantic-level watermarking research.

Qiaoyu Zheng, Zehan Ma, Yijing Zhang, Qiqi Wang, Huijia Li, Qian Liu

Large language models (LLMs) are playing an increasingly pivotal role in LegalAI. However, existing benchmarks are primarily tailored for legal professionals, emphasizing deep reasoning and explainability. While public-facing legal applications demand outputs that are direct, actionable, and accessible, a need largely overlooked by current evaluation frameworks. To bridge this gap, we propose a public-oriented LegalAI benchmark grounded in legal functionalism and genre analysis. Specifically, we categorize public legal demands into two core tasks: Instant Question Answering and Legal Text Generation. We further introduce three public-oriented evaluation dimensions: legal normativity, content relevance, and format usability, which collectively assess the practical validity and user readiness of model outputs. To reflect real-world lay user usage, we evaluate 17 LLMs on Pub-LawBench using only simple prompts and Chain-of-Thought under a vanilla inference setting, excluding complex techniques like RAG or agent-based methods inaccessible to non-experts. Experiments reveal limitations of current LLMs in delivering effective public-oriented legal assistance, highlighting the need for more user-centric model development and benchmarking. Our code and datasets are available for review at https://anonymous.4open.science/r/P-LawBench-E565/.