Peer Revenue Play credits. Real papers.

Papers

Fresh from arXiv, typeset properly. Highlighted words trade on the market.

Computer Science Computation and Language

cs.CL Watch Computation and Language

Friday 16 January

Computation and Language · Cross-listed from cs.LG

Knowledge is Not Enough: Injecting RL Skills for Continual Adaptation

Pingzhi Tang, Yiding Wang and Muhan Zhang

Large Language Models (LLMs) face the "knowledge cutoff" challenge, where their frozen parametric memory prevents direct internalization of new information. While Supervised Fine-Tuning (SFT) is commonly used to update model knowledge, it often updates factual content without reliably improving the model's ability to use the newly incorporated information for question answering or decision-making. Reinforcement Learning (RL) is essential for acquiring reasoning skills; however, its high computational cost makes it impractical for efficient online adaptation. We empirically observe that the parameter updates induced by SFT and RL are nearly orthogonal. Based on this observation, we propose Parametric Skill Transfer (PaST), a framework that supports modular skill transfer for efficient and effective knowledge adaptation. By extracting a domain-agnostic Skill Vector from a source domain, we can linearly inject knowledge manipulation skills into a target model after it has undergone lightweight SFT on new data. Experiments on knowledge-incorporation QA (SQuAD, LooGLE) and agentic tool-use benchmarks (ToolBench) demonstrate the effectiveness of our method. On SQuAD, PaST outperforms the state-of-the-art self-editing SFT baseline by up to 9.9 points. PaST further scales to long-context QA on LooGLE with an 8.0-point absolute accuracy gain, and improves zero-shot ToolBench success rates by +10.3 points on average with consistent gains across tool categories, indicating strong scalability and cross-domain transferability of the Skill Vector.
Read the full abstract

Computation and Language

Reasoning in Trees: Improving Retrieval-Augmented Generation for Multi-Hop Question Answering

Yuling Shi, Maolin Sun, Zijun Liu et al. (+4 more)

Retrieval-Augmented Generation (RAG) has demonstrated significant effectiveness in enhancing large language models (LLMs) for complex multi-hop question answering (QA). For multi-hop QA tasks, current iterative approaches predominantly rely on LLMs to self-guide and plan multi-step exploration paths during retrieval, leading to substantial challenges in maintaining reasoning coherence across steps from inaccurate query decomposition and error propagation. To address these issues, we introduce Reasoning Tree Guided RAG (RT-RAG), a novel hierarchical framework for complex multi-hop QA. RT-RAG systematically decomposes multi-hop questions into explicit reasoning trees, minimizing inaccurate decomposition through structured entity analysis and consensus-based tree selection that clearly separates core queries, known entities, and unknown entities. Subsequently, a bottom-up traversal strategy employs iterative query rewriting and refinement to collect high-quality evidence, thereby mitigating error propagation. Comprehensive experiments show that RT-RAG substantially outperforms state-of-the-art methods by 7.0% F1 and 6.0% EM, demonstrating the effectiveness of RT-RAG in complex multi-hop QA.
Read the full abstract

All authors Yuling Shi, Maolin Sun, Zijun Liu, Mo Yang, Yixiong Fang, Tianran Sun, Xiaodong Gu

Computation and Language

How DDAIR you? Disambiguated Data Augmentation for Intent Recognition

Galo Castillo-López, Alexis Lombard, Nasredine Semmar et al. (+1 more)

Large Language Models (LLMs) are effective for data augmentation in classification tasks like intent detection. In some cases, they inadvertently produce examples that are ambiguous with regard to untargeted classes. We present DDAIR (Disambiguated Data Augmentation for Intent Recognition) to mitigate this problem. We use Sentence Transformers to detect ambiguous class-guided augmented examples generated by LLMs for intent recognition in low-resource scenarios. We identify synthetic examples that are semantically more similar to another intent than to their target one. We also provide an iterative re-generation method to mitigate such ambiguities. Our findings show that sentence embeddings effectively help to (re)generate less ambiguous examples, and suggest promising potential to improve classification performance in scenarios where intents are loosely or broadly defined.
Read the full abstract

All authors Galo Castillo-López, Alexis Lombard, Nasredine Semmar, Gaël de Chalendar

Computation and Language

FactCorrector: A Graph-Inspired Approach to Long-Form Factuality Correction of Large Language Models

Javier Carnerero-Cano, Massimiliano Pronesti, Radu Marinescu et al. (+6 more)

Large language models (LLMs) are widely used in knowledge-intensive applications but often generate factually incorrect responses. A promising approach to rectify these flaws is correcting LLMs using feedback. Therefore, in this paper, we introduce FactCorrector, a new post-hoc correction method that adapts across domains without retraining and leverages structured feedback about the factuality of the original response to generate a correction. To support rigorous evaluations of factuality correction methods, we also develop the VELI5 benchmark, a novel dataset containing systematically injected factual errors and ground-truth corrections. Experiments on VELI5 and several popular long-form factuality datasets show that the FactCorrector approach significantly improves factual precision while preserving relevance, outperforming strong baselines. We release our code at https://ibm.biz/factcorrector.
Read the full abstract

All authors Javier Carnerero-Cano, Massimiliano Pronesti, Radu Marinescu, Tigran Tchrakian, James Barry, Jasmina Gajcin, Yufang Hou, Alessandra Pascale, Elizabeth Daly

Computation and Language

Language of Thought Shapes Output Diversity in Large Language Models

Shaoyang Xu and Wenxuan Zhang

Output diversity is crucial for Large Language Models as it underpins pluralism and creativity. In this work, we reveal that controlling the language used during model thinking-the language of thought-provides a novel and structural source of output diversity. Our preliminary study shows that different thinking languages occupy distinct regions in a model's thinking space. Based on this observation, we study two repeated sampling strategies under multilingual thinking-Single-Language Sampling and Mixed-Language Sampling-and conduct diversity evaluation on outputs that are controlled to be in English, regardless of the thinking language used. Across extensive experiments, we demonstrate that switching the thinking language from English to non-English languages consistently increases output diversity, with a clear and consistent positive correlation such that languages farther from English in the thinking space yield larger gains. We further show that aggregating samples across multiple thinking languages yields additional improvements through compositional effects, and that scaling sampling with linguistic heterogeneity expands the model's diversity ceiling. Finally, we show that these findings translate into practical benefits in pluralistic alignment scenarios, leading to broader coverage of cultural knowledge and value orientations in LLM outputs. Our code is publicly available at https://github.com/iNLP-Lab/Multilingual-LoT-Diversity.
Read the full abstract

Computation and Language

MultiCaption: Detecting disinformation using multilingual visual claims

Rafael Martins Frade, Rrubaa Panchendrarajan and Arkaitz Zubiaga

Online disinformation poses an escalating threat to society, driven increasingly by the rapid spread of misleading content across both multimedia and multilingual platforms. While automated fact-checking methods have advanced in recent years, their effectiveness remains constrained by the scarcity of datasets that reflect these real-world complexities. To address this gap, we first present MultiCaption, a new dataset specifically designed for detecting contradictions in visual claims. Pairs of claims referring to the same image or video were labeled through multiple strategies to determine whether they contradict each other. The resulting dataset comprises 11,088 visual claims in 64 languages, offering a unique resource for building and evaluating misinformation-detection systems in truly multimodal and multilingual environments. We then provide comprehensive experiments using transformer-based architectures, natural language inference models, and large language models, establishing strong baselines for future research. The results show that MultiCaption is more challenging than standard NLI tasks, requiring task-specific finetuning for strong performance. Moreover, the gains from multilingual training and testing highlight the dataset's potential for building effective multilingual fact-checking pipelines without relying on machine translation.
Read the full abstract

Computation and Language

T$^\star$: Progressive Block Scaling for Masked Diffusion Language Models Through Trajectory Aware Reinforcement Learning

Hanchen Xia, Baoyou Chen, Yutang Ge et al. (+2 more)

We present T$^\star$, a simple TraceRL-based training curriculum for progressive block-size scaling in masked diffusion language models (MDMs). Starting from an AR-initialized small-block MDM, T$^\star$ transitions smoothly to larger blocks, enabling higher-parallelism decoding with minimal performance degradation on math reasoning benchmarks. Moreover, further analysis suggests that T$^\star$ may actually converge to an alternative decoding schedule that achieves comparable performance.
Read the full abstract

All authors Hanchen Xia, Baoyou Chen, Yutang Ge, Guojiang Zhao, Siyu Zhu

Computation and Language

DOREMI: Optimizing Long Tail Predictions in Document-Level Relation Extraction

Laura Menotti, Stefano Marchesin and Gianmaria Silvello

Document-Level Relation Extraction (DocRE) presents significant challenges due to its reliance on cross-sentence context and the long-tail distribution of relation types, where many relations have scarce training examples. In this work, we introduce DOcument-level Relation Extraction optiMizing the long taIl (DOREMI), an iterative framework that enhances underrepresented relations through minimal yet targeted manual annotations. Unlike previous approaches that rely on large-scale noisy data or heuristic denoising, DOREMI actively selects the most informative examples to improve training efficiency and robustness. DOREMI can be applied to any existing DocRE model and is effective at mitigating long-tail biases, offering a scalable solution to improve generalization on rare relations.
Read the full abstract

Computation and Language · Cross-listed from cs.AI

TANDEM: Temporal-Aware Neural Detection for Multimodal Hate Speech

Girish A. Koushik, Helen Treharne and Diptesh Kanojia

Social media platforms are increasingly dominated by long-form multimodal content, where harmful narratives are constructed through a complex interplay of audio, visual, and textual cues. While automated systems can flag hate speech with high accuracy, they often function as "black boxes" that fail to provide the granular, interpretable evidence, such as precise timestamps and target identities, required for effective human-in-the-loop moderation. In this work, we introduce TANDEM, a unified framework that transforms audio-visual hate detection from a binary classification task into a structured reasoning problem. Our approach employs a novel tandem reinforcement learning strategy where vision-language and audio-language models optimize each other through self-constrained cross-modal context, stabilizing reasoning over extended temporal sequences without requiring dense frame-level supervision. Experiments across three benchmark datasets demonstrate that TANDEM significantly outperforms zero-shot and context-augmented baselines, achieving 0.73 F1 in target identification on HateMM (a 30% improvement over state-of-the-art) while maintaining precise temporal grounding. We further observe that while binary detection is robust, differentiating between offensive and hateful content remains challenging in multi-class settings due to inherent label ambiguity and dataset imbalance. More broadly, our findings suggest that structured, interpretable alignment is achievable even in complex multimodal settings, offering a blueprint for the next generation of transparent and actionable online safety moderation tools.
Read the full abstract

Computation and Language

The Growing Gains and Pains of Iterative Web Corpora Crawling: Insights from South Slavic CLASSLA-web 2.0 Corpora

Taja Kuzman Pungeršek, Peter Rupnik, Vít Suchomel et al. (+1 more)

Crawling national top-level domains has proven to be highly effective for collecting texts in less-resourced languages. This approach has been recently used for South Slavic languages and resulted in the largest general corpora for this language group: the CLASSLA-web 1.0 corpora. Building on this success, we established a continuous crawling infrastructure for iterative national top-level domain crawling across South Slavic and related webs. We present the first outcome of this crawling infrastructure - the CLASSLA-web 2.0 corpus collection, with substantially larger web corpora containing 17.0 billion words in 38.1 million texts in seven languages: Bosnian, Bulgarian, Croatian, Macedonian, Montenegrin, Serbian, and Slovenian. In addition to genre categories, the new version is also automatically annotated with topic labels. Comparing CLASSLA-web 2.0 with its predecessor reveals that only one-fifth of the texts overlap, showing that re-crawling after just two years yields largely new content. However, while the new web crawls bring growing gains, we also notice growing pains - a manual inspection of top domains reveals a visible degradation of web content, as machine-generated sites now contribute a significant portion of texts.
Read the full abstract

All authors Taja Kuzman Pungeršek, Peter Rupnik, Vít Suchomel, Nikola Ljubešić

Computation and Language · Cross-listed from cs.SD

FlashLabs Chroma 1.0: A Real-Time End-to-End Spoken Dialogue Model with Personalized Voice Cloning

Tanyu Chen, Tairan Chen, Kai Shen et al. (+4 more)

Recent end-to-end spoken dialogue systems leverage speech tokenizers and neural audio codecs to enable LLMs to operate directly on discrete speech representations. However, these models often exhibit limited speaker identity preservation, hindering personalized voice interaction. In this work, we present Chroma 1.0, the first open-source, real-time, end-to-end spoken dialogue model that achieves both low-latency interaction and high-fidelity personalized voice cloning. Chroma achieves sub-second end-to-end latency through an interleaved text-audio token schedule (1:2) that supports streaming generation, while maintaining high-quality personalized voice synthesis across multi-turn conversations. Our experimental results demonstrate that Chroma achieves a 10.96% relative improvement in speaker similarity over the human baseline, with a Real-Time Factor (RTF) of 0.43, while maintaining strong reasoning and dialogue capabilities. Our code and models are publicly available at https://github.com/FlashLabs-AI-Corp/FlashLabs-Chroma and https://huggingface.co/FlashLabs/Chroma-4B .
Read the full abstract

All authors Tanyu Chen, Tairan Chen, Kai Shen, Zhenghua Bao, Zhihui Zhang, Man Yuan, Yi Shi

Computation and Language

Integrity Shield A System for Ethical AI Use & Authorship Transparency in Assessments

Ashish Raj Shekhar, Shiven Agarwal, Priyanuj Bordoloi et al. (+3 more)

Large Language Models (LLMs) can now solve entire exams directly from uploaded PDF assessments, raising urgent concerns about academic integrity and the reliability of grades and credentials. Existing watermarking techniques either operate at the token level or assume control over the model's decoding process, making them ineffective when students query proprietary black-box systems with instructor-provided documents. We present Integrity Shield, a document-layer watermarking system that embeds schema-aware, item-level watermarks into assessment PDFs while keeping their human-visible appearance unchanged. These watermarks consistently prevent MLLMs from answering shielded exam PDFs and encode stable, item-level signatures that can be reliably recovered from model or student responses. Across 30 exams spanning STEM, humanities, and medical reasoning, Integrity Shield achieves exceptionally high prevention (91-94% exam-level blocking) and strong detection reliability (89-93% signature retrieval) across four commercial MLLMs. Our demo showcases an interactive interface where instructors upload an exam, preview watermark behavior, and inspect pre/post AI performance & authorship evidence.
Read the full abstract

All authors Ashish Raj Shekhar, Shiven Agarwal, Priyanuj Bordoloi, Yash Shah, Tejas Anvekar, Vivek Gupta

Computation and Language

Efficient Multilingual Name Type Classification Using Convolutional Networks

Davor Lauc

We present a convolutional neural network approach for classifying proper names by language and entity type. Our model, Onomas-CNN X, combines parallel convolution branches with depthwise-separable operations and hierarchical classification to process names efficiently on CPU hardware. We evaluate the architecture on a large multilingual dataset covering 104 languages and four entity types (person, organization, location, other). Onomas-CNN X achieves 92.1% accuracy while processing 2,813 names per second on a single CPU core - 46 times faster than fine-tuned XLM-RoBERTa with comparable accuracy. The model reduces energy consumption by a factor of 46 compared to transformer baselines. Our experiments demonstrate that specialized CNN architectures remain competitive with large pre-trained models for focused NLP tasks when sufficient training data exists.
Read the full abstract

Computation and Language · Cross-listed from cs.SE

ABC-Bench: Benchmarking Agentic Backend Coding in Real-World Development

Jie Yang, Honglin Guo, Li Ji et al. (+11 more)

The evolution of Large Language Models (LLMs) into autonomous agents has expanded the scope of AI coding from localized code generation to complex, repository-level, and execution-driven problem solving. However, current benchmarks predominantly evaluate code logic in static contexts, neglecting the dynamic, full-process requirements of real-world engineering, particularly in backend development which demands rigorous environment configuration and service deployment. To address this gap, we introduce ABC-Bench, a benchmark explicitly designed to evaluate agentic backend coding within a realistic, executable workflow. Using a scalable automated pipeline, we curated 224 practical tasks spanning 8 languages and 19 frameworks from open-source repositories. Distinct from previous evaluations, ABC-Bench require the agents to manage the entire development lifecycle from repository exploration to instantiating containerized services and pass the external end-to-end API tests. Our extensive evaluation reveals that even state-of-the-art models struggle to deliver reliable performance on these holistic tasks, highlighting a substantial disparity between current model capabilities and the demands of practical backend engineering. Our code is available at https://github.com/OpenMOSS/ABC-Bench.
Read the full abstract

All authors Jie Yang, Honglin Guo, Li Ji, Jiazheng Zhou, Rui Zheng, Zhikai Lei, Shuo Zhang, Zhiheng Xi, Shichun Liu, Yuxin Wang, Bo Wang, Yining Zheng, Tao Gui, Xipeng Qiu

Computation and Language · Cross-listed from cs.LG

Spurious Rewards Paradox: Mechanistically Understanding How RLVR Activates Memorization Shortcuts in LLMs

Lecheng Yan, Ruizhe Li, Guanhua Chen et al. (+5 more)

Reinforcement Learning with Verifiable Rewards (RLVR) is highly effective for enhancing LLM reasoning, yet recent evidence shows models like Qwen 2.5 achieve significant gains even with spurious or incorrect rewards. We investigate this phenomenon and identify a "Perplexity Paradox": spurious RLVR triggers a divergence where answer-token perplexity drops while prompt-side coherence degrades, suggesting the model is bypassing reasoning in favor of memorization. Using Path Patching, Logit Lens, JSD analysis, and Neural Differential Equations, we uncover a hidden Anchor-Adapter circuit that facilitates this shortcut. We localize a Functional Anchor in the middle layers (L18-20) that triggers the retrieval of memorized solutions, followed by Structural Adapters in later layers (L21+) that transform representations to accommodate the shortcut signal. Finally, we demonstrate that scaling specific MLP keys within this circuit allows for bidirectional causal steering-artificially amplifying or suppressing contamination-driven performance. Our results provide a mechanistic roadmap for identifying and mitigating data contamination in RLVR-tuned models. Code is available at https://github.com/idwts/How-RLVR-Activates-Memorization-Shortcuts.
Read the full abstract

All authors Lecheng Yan, Ruizhe Li, Guanhua Chen, Qing Li, Jiahui Geng, Wenxi Li, Longyue Wang, Chenyang Lyu

Computation and Language

CoG: Controllable Graph Reasoning via Relational Blueprints and Failure-Aware Refinement over Knowledge Graphs

Yuanxiang Liu, Songze Li, Xiaoke Guo et al. (+4 more)

Large Language Models (LLMs) have demonstrated remarkable reasoning capabilities but often grapple with reliability challenges like hallucinations. While Knowledge Graphs (KGs) offer explicit grounding, existing paradigms of KG-augmented LLMs typically exhibit cognitive rigidity–applying homogeneous search strategies that render them vulnerable to instability under neighborhood noise and structural misalignment leading to reasoning stagnation. To address these challenges, we propose CoG, a training-free framework inspired by Dual-Process Theory that mimics the interplay between intuition and deliberation. First, functioning as the fast, intuitive process, the Relational Blueprint Guidance module leverages relational blueprints as interpretable soft structural constraints to rapidly stabilize the search direction against noise. Second, functioning as the prudent, analytical process, the Failure-Aware Refinement module intervenes upon encountering reasoning impasses. It triggers evidence-conditioned reflection and executes controlled backtracking to overcome reasoning stagnation. Experimental results on three benchmarks demonstrate that CoG significantly outperforms state-of-the-art approaches in both accuracy and efficiency.
Read the full abstract

All authors Yuanxiang Liu, Songze Li, Xiaoke Guo, Zhaoyan Gong, Qifei Zhang, Huajun Chen, Wen Zhang

Computation and Language

Spectral Characterization and Mitigation of Sequential Knowledge Editing Collapse

Chi Zhang, Mengqi Zhang, Xiaotian Ye et al. (+5 more)

Sequential knowledge editing in large language models often causes catastrophic collapse of the model's general abilities, especially for parameter-modifying methods. Existing approaches mitigate this issue through heuristic constraints on parameter updates, yet the mechanisms underlying such degradation remain insufficiently understood. In this work, we present a spectral analysis of sequential knowledge editing and show that a model's general abilities are closely associated with dominant singular directions of pretrained weight matrices. These directions are highly sensitive to perturbations and are progressively disrupted by repeated edits, closely tracking the collapse in both editing efficacy and general performance. Building on this insight, we propose REVIVE, a plug-and-play framework that stabilizes sequential editing by explicitly preserving the dominant singular subspace. REVIVE represents parameter updates in the spectral basis of the original weights and filters components that would interfere with the protected region. Extensive experiments across multiple models and benchmarks show that REVIVE consistently improves editing efficacy while substantially preserving general abilities under long-horizon sequential editing, including extreme settings with up to 20,000 edits.
Read the full abstract

All authors Chi Zhang, Mengqi Zhang, Xiaotian Ye, Runxi Cheng, Zisheng Zhou, Ying Zhou, Pengjie Ren, Zhumin Chen

Computation and Language · Cross-listed from cs.SD

SonicBench: Dissecting the Physical Perception Bottleneck in Large Audio Language Models

Yirong Sun, Yanjun Chen, Xin Qiu et al. (+8 more)

Large Audio Language Models (LALMs) excel at semantic and paralinguistic tasks, yet their ability to perceive the fundamental physical attributes of audio such as pitch, loudness, and spatial location remains under-explored. To bridge this gap, we introduce SonicBench, a psychophysically grounded benchmark that systematically evaluates 12 core physical attributes across five perceptual dimensions. Unlike previous datasets, SonicBench uses a controllable generation toolbox to construct stimuli for two complementary paradigms: recognition (absolute judgment) and comparison (relative judgment). This design allows us to probe not only sensory precision but also relational reasoning capabilities, a domain where humans typically exhibit greater proficiency. Our evaluation reveals a substantial deficiency in LALMs' foundational auditory understanding; most models perform near random guessing and, contrary to human patterns, fail to show the expected advantage on comparison tasks. Furthermore, explicit reasoning yields minimal gains. However, our linear probing analysis demonstrates crucially that frozen audio encoders do successfully capture these physical cues (accuracy at least 60%), suggesting that the primary bottleneck lies in the alignment and decoding stages, where models fail to leverage the sensory signals they have already captured.
Read the full abstract

All authors Yirong Sun, Yanjun Chen, Xin Qiu, Gang Zhang, Hongyu Chen, Daokuan Wu, Chengming Li, Min Yang, Dawei Zhu, Wei Zhang, Xiaoyu Shen

Computation and Language

Budget-Aware Anytime Reasoning with LLM-Synthesized Preference Data

Xuanming Zhang, Shwan Ashrafi, Aziza Mirsaidova et al. (+5 more)

We study the reasoning behavior of large language models (LLMs) under limited computation budgets. In such settings, producing useful partial solutions quickly is often more practical than exhaustive reasoning, which incurs high inference costs. Many real-world tasks, such as trip planning, require models to deliver the best possible output within a fixed reasoning budget. We introduce an anytime reasoning framework and the Anytime Index, a metric that quantifies how effectively solution quality improves as reasoning tokens increase. To further enhance efficiency, we propose an inference-time self-improvement method using LLM-synthesized preference data, where models learn from their own reasoning comparisons to produce better intermediate solutions. Experiments on NaturalPlan (Trip), AIME, and GPQA datasets show consistent gains across Grok-3, GPT-oss, GPT-4.1/4o, and LLaMA models, improving both reasoning quality and efficiency under budget constraints.
Read the full abstract

All authors Xuanming Zhang, Shwan Ashrafi, Aziza Mirsaidova, Amir H. Rezaeian, Miguel Ballesteros, Lydia B. Chilton, Zhou Yu, Dan Roth

Computation and Language

From Interpretability to Performance: Optimizing Retrieval Heads for Long-Context Language Models

Youmi Ma and Naoaki Okazaki

Advances in mechanistic interpretability have identified special attention heads, known as retrieval heads, that are responsible for retrieving information from the context. However, the role of these retrieval heads in improving model performance remains unexplored. This work investigates whether retrieval heads can be leveraged to enhance the long-context capabilities of LLMs. Specifically, we propose RetMask, a method that generates training signals by contrasting normal model outputs with those from an ablated variant in which the retrieval heads are masked. This mechanism-based approach achieves substantial improvements: +2.28 points on HELMET at 128K for Llama-3.1, with +70% gains on generation with citation and +32% on passage re-ranking, while preserving performance on general tasks. Experiments across four models in three families demonstrate that RetMask consistently improves long-context performance, where gains correlate with the sparsity of the retrieval score distribution: models with sparser distributions, where retrieval capabilities are concentrated in a small set of heads, respond more strongly, while those with less sparse distributions show more modest gains. These results validate the functional role of retrieval heads and show that mechanistic insights can be transformed into performance enhancements.
Read the full abstract

Computation and Language

Finding the Translation Switch: Discovering and Exploiting the Task-Initiation Features in LLMs

Xinwei Wu, Heng Liu, Xiaohu Zhao et al. (+6 more)

Large Language Models (LLMs) frequently exhibit strong translation abilities, even without task-specific fine-tuning. However, the internal mechanisms governing this innate capability remain largely opaque. To demystify this process, we leverage Sparse Autoencoders (SAEs) and introduce a novel framework for identifying task-specific features. Our method first recalls features that are frequently co-activated on translation inputs and then filters them for functional coherence using a PCA-based consistency metric. This framework successfully isolates a small set of **translation initiation** features. Causal interventions demonstrate that amplifying these features steers the model towards correct translation, while ablating them induces hallucinations and off-task outputs, confirming they represent a core component of the model's innate translation competency. Moving from analysis to application, we leverage this mechanistic insight to propose a new data selection strategy for efficient fine-tuning. Specifically, we prioritize training on **mechanistically hard** samples-those that fail to naturally activate the translation initiation features. Experiments show this approach significantly improves data efficiency and suppresses hallucinations. Furthermore, we find these mechanisms are transferable to larger models of the same family. Our work not only decodes a core component of the translation mechanism in LLMs but also provides a blueprint for using internal model mechanism to create more robust and efficient models. The codes are available at https://github.com/flamewei123/AAAI26-translation-Initiation-Features.
Read the full abstract

All authors Xinwei Wu, Heng Liu, Xiaohu Zhao, Yuqi Ren, Linlong Xu, Longyue Wang, Deyi Xiong, Weihua Luo, Kaifu Zhang

Computation and Language · Cross-listed from cs.AI

AdaMARP: An Adaptive Multi-Agent Interaction Framework for General Immersive Role-Playing

Zhenhua Xu, Dongsheng Chen, Shuo Wang et al. (+4 more)

LLM role-playing aims to portray arbitrary characters in interactive narratives, yet existing systems often suffer from limited immersion and adaptability. They typically under-model dynamic environmental information and assume largely static scenes and casts, offering insufficient support for multi-character orchestration, scene transitions, and on-the-fly character introduction. We propose an adaptive multi-agent role-playing framework, AdaMARP, featuring an immersive message format that interleaves [Thought], (Action), <Environment>, and Speech, together with an explicit Scene Manager that governs role-playing through discrete actions (init_scene, pick_speaker, switch_scene, add_role, end) accompanied by rationales. To train these capabilities, we construct AdaRPSet for the Actor Model and AdaSMSet for supervising orchestration decisions, and introduce AdaptiveBench for trajectory-level evaluation. Experiments across multiple backbones and model scales demonstrate consistent improvements: AdaRPSet enhances character consistency, environment grounding, and narrative coherence, with an 8B actor outperforming several commercial LLMs, while AdaSMSet enables smoother scene transitions and more natural role introductions, surpassing Claude Sonnet 4.5 using only a 14B LLM.
Read the full abstract

All authors Zhenhua Xu, Dongsheng Chen, Shuo Wang, Jian Li, Chengjie Wang, Meng Han, Yabiao Wang

Computation and Language

NOVA: NOise-aware Verbal Confidence CAlibration for Robust Large Language Models in RAG Systems

Jiayu Liu, Rui Wang, Qing Zong et al. (+9 more)

Accurately assessing model confidence is essential for deploying large language models (LLMs) in mission-critical factual domains. While retrieval-augmented generation (RAG) is widely adopted to improve grounding, confidence calibration in RAG settings remains poorly understood. We conduct a systematic study across four benchmarks, revealing that LLMs exhibit poor calibration performance especially when noisy contexts are retrieved. Specifically, contradictory or irrelevant evidence tends to exacerbate the model's overconfidence issue. To address this, we propose NOVA Rules (NOise-Aware Verbal Confidence CAlibration Rules) to provide a principled foundation for resolving overconfidence under noise. We further design NOVA, a noise-aware calibration framework that synthesizes supervision from ~2K HotpotQA examples guided by these rules. By performing supervised fine-tuning (SFT) with this data, NOVA equips models with intrinsic noise awareness without relying on stronger teacher models. Empirical results show that NOVA yields substantial gains, improving ECE scores by 10.9% in-domain and 8.0% out-of-domain. By bridging the gap between retrieval noise and verbal calibration, NOVA paves the way for both accurate and epistemically reliable LLMs.
Read the full abstract

All authors Jiayu Liu, Rui Wang, Qing Zong, Yumeng Wang, Cheng Qian, Qingcheng Zeng, Tianshi Zheng, Haochen Shi, Dadi Guo, Baixuan Xu, Chunyang Li, Yangqiu Song

Computation and Language

Redefining Machine Simultaneous Interpretation: From Incremental Translation to Human-Like Strategies

Qianen Zhang, Zeyu Yang and Satoshi Nakamura

Simultaneous Machine Translation (SiMT) requires high-quality translations under strict real-time constraints, which traditional policies with only READ/WRITE actions cannot fully address. We extend the action space of SiMT with four adaptive actions: Sentence_Cut, Drop, Partial_Summarization and Pronominalization, which enable real-time restructuring, omission, and simplification while preserving semantic fidelity. We adapt these actions in a large language model (LLM) framework and construct training references through action-aware prompting. To evaluate both quality and word-level monotonicity, we further develop a latency-aware TTS pipeline that maps textual outputs to speech with realistic timing. Experiments on the ACL60/60 English-Chinese, English-German and English-Japanese benchmarks show that our framework consistently improves semantic metrics and achieves lower delay compared to reference translations and salami-based baselines. Notably, combining Drop and Sentence_Cut leads to consistent improvements in the balance between fluency and latency. These results demonstrate that enriching the action space of LLM-based SiMT provides a promising direction for bridging the gap between human and machine interpretation.
Read the full abstract

Computation and Language

When Personalization Misleads: Understanding and Mitigating Hallucinations in Personalized LLMs

Zhongxiang Sun, Yi Zhan, Chenglei Shen et al. (+4 more)

Personalized large language models (LLMs) adapt model behavior to individual users to enhance user satisfaction, yet personalization can inadvertently distort factual reasoning. We show that when personalized LLMs face factual queries, there exists a phenomenon where the model generates answers aligned with a user's prior history rather than the objective truth, resulting in personalization-induced hallucinations that degrade factual reliability and may propagate incorrect beliefs, due to representational entanglement between personalization and factual representations. To address this issue, we propose Factuality-Preserving Personalized Steering (FPPS), a lightweight inference-time approach that mitigates personalization-induced factual distortions while preserving personalized behavior. We further introduce PFQABench, the first benchmark designed to jointly evaluate factual and personalized question answering under personalization. Experiments across multiple LLM backbones and personalization methods show that FPPS substantially improves factual accuracy while maintaining personalized performance.
Read the full abstract

All authors Zhongxiang Sun, Yi Zhan, Chenglei Shen, Weijie Yu, Xiao Zhang, Ming He, Jun Xu

Computation and Language

ZPD Detector: Data Selection via Capability-Difficulty Alignment for Large Language Models

Bo Yang, Yunkui Chen, Lanfei Feng et al. (+2 more)

As the cost of training large language models continues to increase and high-quality training data become increasingly scarce, selecting high-value samples or synthesizing effective training data under limited data budgets has emerged as a critical research problem. Most existing data selection methods rely on static criteria, such as difficulty, uncertainty, or heuristics, and fail to model the evolving relationship between the model and the data. Inspired by the educational theory of the Zone of Proximal Development (ZPD), we propose ZPD Detector, a data selection framework that adopts a bidirectional perspective between models and data by explicitly modeling the alignment between sample difficulty and the model's current capability. ZPD Detector integrates difficulty calibration, model capability estimation based on Item Response Theory (IRT), and a capability-difficulty matching score to dynamically identify the most informative samples at each learning stage, improving data utilization efficiency; moreover, this dynamic matching strategy provides new insights into training strategy design. All code and data will be released after our work be accepted to support reproducible researc
Read the full abstract

All authors Bo Yang, Yunkui Chen, Lanfei Feng, Yu Zhang, Shijian Li

Computation and Language · Cross-listed from cs.CR

AJAR: Adaptive Jailbreak Architecture for Red-teaming

Yipu Dou and Wang Yang

Large language model (LLM) safety evaluation is moving from content moderation to action security as modern systems gain persistent state, tool access, and autonomous control loops. Existing jailbreak frameworks still leave a gap between adaptive multi-turn attacks and agentic runtimes: attack algorithms are usually packaged as monolithic scripts, while agent harnesses rarely expose explicit abstractions for rollback, tool simulation, or strategy switching. We present AJAR, a red-teaming framework that exposes multi-turn jailbreak algorithms as callable MCP services and lets an Auditor Agent orchestrate them inside a tool-aware runtime built on Petri. AJAR integrates three representative attacks, namely Crescendo, ActorAttack, and X-Teaming, under a shared service interface for planning, prompt generation, optimization, evaluation, and context control. On 200 HarmBench validation behaviors, AJAR improves X-Teaming from 65.0% to 76.0% attack success rate (ASR), reaches 80% cumulative success one turn earlier than the native implementation, and reproduces Crescendo more effectively than PyRIT (91.0% vs. 87.5% ASR). Behavior-level analysis shows that these gains are concentrated in hard categories and frequently depend on rollback-enabled transcript repair. We further show that tool access reshapes rather than uniformly enlarges the attack surface: ActorAttack rises from 51.0% to 56.0% ASR with tools, whereas Crescendo drops from 91.0% to 78.0% and X-Teaming from 76.0% to 55.5%, with the sharpest declines appearing in categories that rely on long semantic buildup. These results position AJAR as a practical foundation for evaluating multi-turn jailbreaks under realistic agent constraints. Code and data are available at https://github.com/douyipu/ajar.
Read the full abstract

Computation and Language

Steering Language Models Before They Speak: Logit-Level Interventions

Hyeseon An, Shinwoo Park, Hyundong Jin et al. (+1 more)

Controllable generation requires language models to realize output characteristics such as reading level, politeness, and toxicity. Existing steering methods are often indirect, require access to internal activations, or depend on auxiliary trained models. We propose SWAI, a training-free inference-time method that addresses these limitations by steering directly in logit space using corpus-derived token statistics. SWAI computes z-normalized one-vs-rest log-odds scores from labeled corpora and biases high-scoring tokens only within the model's top-K candidate set, allowing control to favor target-characteristic tokens while preserving contextually plausible choices. Across readability, politeness, and toxicity control, SWAI consistently improves over prompt-based and prior logit-level baselines without modifying model parameters, accessing internal layers, or training an auxiliary model. Selectivity and lookup-table ablations show that the gains come from target-specific statistical scores rather than generic logit perturbation. These results indicate that effective steering does not require learned controllers when the logit intervention is guided by target-specific statistics under high-probability candidates.
Read the full abstract

All authors Hyeseon An, Shinwoo Park, Hyundong Jin, Yo-Sub Han

Computation and Language

Multi-Stage Patient Role-Playing Framework for Realistic Clinical Interactions

Shijie Jiang, Zefan Zhang, Kehua Zhu et al. (+2 more)

The simulation of realistic clinical interactions plays a pivotal role in advancing clinical Large Language Models (LLMs) and supporting medical diagnostic education. Existing approaches and benchmarks rely on generic or LLM-generated dialogue data, which limits the authenticity and diversity of doctor-patient interactions. In this work, we propose the first Chinese patient simulation dataset (Ch-PatientSim), constructed from realistic clinical interaction scenarios to comprehensively evaluate the performance of models in emulating patient behavior. Patients are simulated based on a five-dimensional persona structure. To address issues of the persona class imbalance, a portion of the dataset is augmented using few-shot generation, followed by manual verification. We evaluate various state-of-the-art LLMs and find that most produce overly formal responses that lack individual personality. To address this limitation, we propose a training-free Multi-Stage Patient Role-Playing (MSPRP) framework, which decomposes interactions into three stages to ensure both personalization and realism in model responses. Experimental results demonstrate that our approach significantly improves model performance across multiple dimensions of patient simulation.
Read the full abstract

All authors Shijie Jiang, Zefan Zhang, Kehua Zhu, Tian Bai, Ruihong Zhao

Computation and Language · Cross-listed from cs.CV

PatientVLM Meets DocVLM: Pre-Consultation Dialogue Between Vision-Language Models for Efficient Diagnosis

K Lokesh, Abhirama Subramanyam Penamakuri, Uday Agarwal et al. (+4 more)

Traditionally, AI research in medical diagnosis has largely centered on image analysis. While this has led to notable advancements, the absence of patient-reported symptoms continues to hinder diagnostic accuracy. To address this, we propose a Pre-Consultation Dialogue Framework (PCDF) that mimics real-world diagnostic procedures, where doctors iteratively query patients before reaching a conclusion. Specifically, we simulate diagnostic dialogues between two vision-language models (VLMs): a DocVLM, which generates follow-up questions based on the image and dialogue history, and a PatientVLM, which responds using a symptom profile derived from the ground-truth diagnosis. We additionally conducted a small-scale clinical validation of the synthetic symptoms generated by our framework, with licensed clinicians confirming their clinical relevance, symptom coverage, and overall realism. These findings indicate that the resulting DocVLM-PatientVLM interactions form coherent, multi-turn consultations paired with images and diagnoses, which we then use to fine-tune the DocVLM. This dialogue-based supervision leads to substantial gains over image-only training, highlighting the value of realistic symptom elicitation for diagnosis.
Read the full abstract

All authors K Lokesh, Abhirama Subramanyam Penamakuri, Uday Agarwal, Apoorva Challa, Shreya K Gowda, Somesh Gupta, Anand Mishra