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

Thursday 8 January

Computation and Language

ASKD-Whisper: Adaptive Self-knowledge Distillation for Efficient and Low-Latency Automatic Speech Recognition

Junseok Lee, Nahun Kim, Sangyong Lee et al. (+1 more)

Knowledge distillation (KD) is one of the most effective paradigms for compressing large-scale foundation models into deployable architectures. In the context of Automatic Speech Recognition (ASR), previous studies have predominantly focused on forcing the student model to strictly mimic the predictive distribution of a massive teacher model. However, this static dependency often presents an inherent trade-off: while the student rapidly acquires basic linguistic representations, it simultaneously inherits the teacher's domain-specific blind spots and over-confident hallucinations, leading to a severe decline in out-of-distribution generalization capacity. To effectively mitigate this issue, we propose Adaptive Self-Knowledge Distillation (ASKD), a dynamic curriculum framework. ASKD systematically decays the dependency on the teacher's distribution as training progresses-thereby unlocking the student's independent reasoning capacity-and subsequently employs a self-knowledge distillation phase to act as a structural regularizer. By applying ASKD, we distill the massive Whisper architecture into a compact variant, ASKD-Whisper. In our comprehensive evaluations across diverse acoustic domains, ASKD-Whisper not only achieves a 5x speedup in inference latency but also outperforms its teacher model by yielding a 1.07% lower word error rate (WER). These results demonstrate that ASKD effectively prevents teacher-induced overfitting and establishes a new state-of-the-art for generalizable model compression.
Read the full abstract

All authors Junseok Lee, Nahun Kim, Sangyong Lee, Chang-Jae Chun

Computation and Language · Cross-listed from cs.LG

SPINAL – Scaling-law and Preference Integration in Neural Alignment Layers

Arion Das, Partha Pratim Saha, Amit Dhanda et al. (+3 more)

Direct Preference Optimization (DPO) is a principled, scalable alternative to RLHF for aligning large language models from pairwise preferences, but its internal geometric footprint remains undercharacterized, limiting audits, checkpoint comparisons, and failure prediction. We introduce SPINAL (Scaling-law and Preference Integration in Neural Alignment Layers), a diagnostic that measures how alignment reshapes representations across depth by tracing localized structural change layer by layer. Across model families, DPO produces a layerwise calibration effect concentrated in the final decoder blocks (often layers 21-30), where preference gradients most directly affect the next-token distribution. SPINAL encodes each checkpoint as a depth trace over (layer index, contraction score, transport score). The contraction score summarizes how quickly the tail of a layer's spectrum decays (how fast small modes vanish); higher values indicate stronger contraction into fewer effective directions. The transport score summarizes how much the token distribution shifts between adjacent layers using a bounded overlap measure; lower values indicate shorter, smoother steps through representation space. Aligned checkpoints show a late-layer ramp-up in contraction and a smooth reduction in transport, consistent with tightened and stabilized policy mass, while unaligned models trace higher-curvature, more entropic, and geometrically incoherent depth paths. Overall, alignment is geometrically localized: the final layers encode the dominant preference-induced corrections. SPINAL turns this localization into a practical audit signal, quantifying where alignment concentrates, how strongly it manifests, and when it begins to destabilize during training.
Read the full abstract

All authors Arion Das, Partha Pratim Saha, Amit Dhanda, Vinija Jain, Aman Chadha, Amitava Das

Computation and Language · Cross-listed from cs.LG

Geometry-Aware Hallucination Detection in Large Language Models

Bodla Krishna Vamshi, Rohan Bhatnagar and Haizhao Yang

Large language models (LLMs) frequently generate factually incorrect or unsupported content, commonly referred to as hallucinations. Prior work has explored decoding strategies, retrieval augmentation, and supervised fine-tuning for hallucination detection, while recent studies show that in-context learning (ICL) can substantially influence factual reliability. However, existing ICL demonstration selection methods often rely on surface-level similarity heuristics and exhibit limited robustness across tasks and models. We propose GA-ICL, a geometry-aware demonstration sampling framework for selecting in-context demonstrations that leverages latent representations extracted from frozen LLMs. By jointly modeling local manifold structure and class-aware prototype geometry, GA-ICL selects demonstrations based on their proximity to learned prototypes rather than lexical or embedding similarity alone. Across factual verification (FEVER) and hallucination detection (HaluEval) benchmarks, GA-ICL outperforms standard ICL selection baselines in the majority of evaluated settings, with particularly strong gains on dialogue and summarization tasks. The method remains robust under temperature perturbations and model variation, indicating improved stability compared to heuristic retrieval strategies. While lexical retrieval can remain competitive in certain question-answering regimes at smaller model scales, our results demonstrate that geometry-aware prototype selection provides a reliable and training-light approach for hallucination detection without modifying LLM parameters. Extended evaluations on Phi-14B and Qwen3-32B confirm that GA-ICL scales effectively to larger models, outperforming all compared baselines including on QA tasks where smaller models show boundary-condition limitations, offering a principled direction for improved ICL demonstration selection.
Read the full abstract

Computation and Language · Cross-listed from cs.CY

Political Alignment in Large Language Models: A Multidimensional Audit of Psychometric Identity and Behavioral Bias

Adib Sakhawat, Tahsin Islam, Takia Farhin et al. (+3 more)

As large language models (LLMs) are increasingly deployed, understanding how they express political positioning is important for evaluating alignment and downstream effects. We audit 26 contemporary LLMs using three political psychometric inventories (Political Compass, SapplyValues, 8Values) and a news bias labeling task. To test robustness, inventories are administered across multiple semantic prompt variants and analyzed with a two-way ANOVA separating model and prompt effects. Most models cluster in a similar ideological region, with 96.3% located in the Libertarian-Left quadrant of the Political Compass, and model identity explaining most variance across prompt variants ($\eta^2 > 0.90$). Cross-instrument comparisons suggest that the Political Compass social axis aligns more strongly with cultural progressivism than authority-related measures ($r=-0.64$). We observe differences between open-weight and closed-source models and asymmetric performance in detecting extreme political bias in downstream classification. Regression analysis finds that psychometric ideological positioning does not significantly predict classification errors, providing no evidence of a statistically significant relationship between conversational ideological identity and task-level behavior. These findings suggest that single-axis evaluations are insufficient and that multidimensional auditing frameworks are important to characterize alignment behavior in deployed LLMs. Our code and data are publicly available at https://github.com/sakhadib/PolAlignLLM.
Read the full abstract

All authors Adib Sakhawat, Tahsin Islam, Takia Farhin, Syed Rifat Raiyan, Hasan Mahmud, Md Kamrul Hasan

Computation and Language · Cross-listed from cs.LG

MLB: A Scenario-Driven Benchmark for Evaluating Large Language Models in Clinical Applications

Qing He, Dongsheng Bi, Jianrong Lu et al. (+20 more)

The proliferation of Large Language Models (LLMs) presents transformative potential for healthcare, yet practical deployment is hindered by the absence of frameworks that assess real-world clinical utility. Existing benchmarks test static knowledge, failing to capture the dynamic, application-oriented capabilities required in clinical practice. To bridge this gap, we introduce a Medical LLM Benchmark MLB, a comprehensive benchmark evaluating LLMs on both foundational knowledge and scenario-based reasoning. MLB is structured around five core dimensions: Medical Knowledge (MedKQA), Safety and Ethics (MedSE), Medical Record Understanding (MedRU), Smart Services (SmartServ), and Smart Healthcare (SmartCare). The benchmark integrates 22 datasets (17 newly curated) from diverse Chinese clinical sources, covering 64 clinical specialties. Its design features a rigorous curation pipeline involving 300 licensed physicians. Besides, we provide a scalable evaluation methodology, centered on a specialized judge model trained via Supervised Fine-Tuning (SFT) on expert annotations. Our comprehensive evaluation of 10 leading models reveals a critical translational gap: while the top-ranked model, Kimi-K2-Instruct (77.3% accuracy overall), excels in structured tasks like information extraction (87.8% accuracy in MedRU), performance plummets in patient-facing scenarios (61.3% in SmartServ). Moreover, the exceptional safety score (90.6% in MedSE) of the much smaller Baichuan-M2-32B highlights that targeted training is equally critical. Our specialized judge model, trained via SFT on a 19k expert-annotated medical dataset, achieves 92.1% accuracy, an F1-score of 94.37%, and a Cohen's Kappa of 81.3% for human-AI consistency, validating a reproducible and expert-aligned evaluation protocol. MLB thus provides a rigorous framework to guide the development of clinically viable LLMs.
Read the full abstract

All authors Qing He, Dongsheng Bi, Jianrong Lu, Minghui Yang, Zixiao Chen, Jiacheng Lu, Jing Chen, Nannan Du, Xiao Cu, Sijing Wu, Peng Xiang, Yinyin Hu, Yi Guo, Chunpu Li, Shaoyang Li, Zhuo Dong, Ming Jiang, Shuai Guo, Liyun Feng, Jin Peng, Jian Wang, Jinjie Gu, Junwei Liu

Computation and Language · Cross-listed from cs.CV

Thinking with Map: Reinforced Parallel Map-Augmented Agent for Geolocalization

Yuxiang Ji, Yong Wang, Ziyu Ma et al. (+6 more)

The image geolocalization task aims to predict the location where an image was taken anywhere on Earth using visual clues. Existing large vision-language model (LVLM) approaches leverage world knowledge, chain-of-thought reasoning, and agentic capabilities, but overlook a common strategy used by humans – using maps. In this work, we first equip the model Thinking with Map ability and formulate it as an agent-in-the-map loop. We develop a two-stage optimization scheme for it, including agentic reinforcement learning (RL) followed by parallel test-time scaling (TTS). The RL strengthens the agentic capability of model to improve sampling efficiency, and the parallel TTS enables the model to explore multiple candidate paths before making the final prediction, which is crucial for geolocalization. To evaluate our method on up-to-date and in-the-wild images, we further present MAPBench, a comprehensive geolocalization training and evaluation benchmark composed entirely of real-world images. Experimental results show that our method outperforms existing open- and closed-source models on most metrics, specifically improving Acc@500m from 8.0% to 22.1% compared to Gemini-3-Pro with Google Search/Map grounded mode.
Read the full abstract

All authors Yuxiang Ji, Yong Wang, Ziyu Ma, Yiming Hu, Hailang Huang, Xuecai Hu, Guanhua Chen, Liaoni Wu, Xiangxiang Chu

Computation and Language

Large Language Models Are Bad Dice Players: LLMs Struggle to Generate Random Numbers from Statistical Distributions

Minda Zhao, Yilun Du and Mengyu Wang

As large language models (LLMs) transition from chat interfaces to integral components of stochastic pipelines and systems approaching general intelligence, the ability to faithfully sample from specified probability distributions has become a functional requirement rather than a theoretical curiosity. We present the first large-scale, statistically powered audit of native probabilistic sampling in frontier LLMs, benchmarking 11 models across 15 distributions. To disentangle failure modes, we employ a dual-protocol design: Batch Generation, where a model produces $N{=}1000$ samples within one response, and Independent Requests, comprising $N{=}1000$ stateless calls. We observe a sharp protocol asymmetry: batch generation achieves only modest statistical validity, with a 7% median pass rate, while independent requests collapse almost entirely, with 10 of 11 models passing none of the distributions. Beyond this asymmetry, we reveal that sampling fidelity degrades monotonically with distributional complexity and aggravates as the sampling horizon $N$ increases. Finally, we demonstrate how the propagation of these failures into downstream real-world application tasks introduces systematic biases: models fail to enforce uniform answer-position constraints in Multiple Choice Question generation and systematically violate demographic targets in attribute-constrained text-to-image prompt synthesis. These findings indicate that current LLMs lack a functional internal sampler, necessitating external tools for applications requiring statistical guarantees.
Read the full abstract

Computation and Language

Glitter: Visualizing Lexical Surprisal for Readability in Administrative Texts

Jan Černý, Ivana Kvapilíková and Silvie Cinková

This work investigates how measuring information entropy of text can be used to estimate its readability. We propose a visualization framework that can be used to approximate information entropy of text using multiple language models and visualize the result. The end goal is to use this method to estimate and improve readability and clarity of administrative or bureaucratic texts. Our toolset is available as a libre software on https://github.com/ufal/Glitter.
Read the full abstract

Computation and Language

Same Claim, Different Judgment: Benchmarking Scenario-Induced Bias in Multilingual Financial Misinformation Detection

Zhiwei Liu, Yupen Cao, Yuechen Jiang et al. (+22 more)

Large language models (LLMs) have been widely applied across various domains of finance. Since their training data are largely derived from human-authored corpora, LLMs may inherit a range of human biases. Behavioral biases can lead to instability and uncertainty in decision-making, particularly when processing financial information. However, existing research on LLM bias has mainly focused on direct questioning or simplified, general-purpose settings, with limited consideration of the complex real-world financial environments and high-risk, context-sensitive, multilingual financial misinformation detection tasks MFMD. In this work, we propose MFMDScen, a comprehensive benchmark for evaluating behavioral biases of LLMs in MFMD across diverse economic scenarios. In collaboration with financial experts, we construct three types of complex financial scenarios: (i) role- and personality-based, (ii) role- and region-based, and (iii) role-based scenarios incorporating ethnicity and religious beliefs. We further develop a multilingual financial misinformation dataset covering English, Chinese, Greek, and Bengali. By integrating these scenarios with misinformation claims, MFMDScen enables a systematic evaluation of 22 mainstream LLMs. Our findings reveal that pronounced behavioral biases persist across both commercial and open-source models. This project is available at https://github.com/lzw108/FMD.
Read the full abstract

All authors Zhiwei Liu, Yupen Cao, Yuechen Jiang, Mohsinul Kabir, Polydoros Giannouris, Chen Xu, Ziyang Xu, Tianlei Zhu, Md. Tariquzzaman, Triantafillos Papadopoulos, Yan Wang, Lingfei Qian, Xueqing Peng, Zhuohan Xie, Ye Yuan, Saeed Almheiri, Abdulrazzaq Alnajjar, Mingbin Chen, Harry Stuart, Paul Thompson, Prayag Tiwari, Alejandro Lopez-Lira, Xue Liu, Jimin Huang, Sophia Ananiadou

Computation and Language · Cross-listed from cs.AI

Conformity and Social Impact on AI Agents

Alessandro Bellina, Giordano De Marzo and David Garcia

As AI agents increasingly operate in multi-agent environments, understanding their collective behavior becomes critical for predicting the dynamics of artificial societies. This study examines conformity, the tendency to align with group opinions under social pressure, in large multimodal language models functioning as AI agents. By adapting classic visual experiments from social psychology, we investigate how AI agents respond to group influence as social actors. Our experiments reveal that AI agents exhibit a systematic conformity bias, aligned with Social Impact Theory, showing sensitivity to group size, unanimity, task difficulty, and source characteristics. Critically, AI agents achieving near-perfect performance in isolation become highly susceptible to manipulation through social influence. This vulnerability persists across model scales: while larger models show reduced conformity on simple tasks due to improved capabilities, they remain vulnerable when operating at their competence boundary. These findings reveal fundamental security vulnerabilities in AI agent decision-making that could enable malicious manipulation, misinformation campaigns, and bias propagation in multi-agent systems, highlighting the urgent need for safeguards in collective AI deployments.
Read the full abstract

Computation and Language · Cross-listed from cs.AI

The Persona Paradox: Medical Personas as Behavioral Priors in Clinical Language Models

Tassallah Abdullahi, Shrestha Ghosh, Hamish S Fraser et al. (+5 more)

Persona conditioning can be viewed as a behavioral prior for large language models (LLMs) and is often assumed to confer expertise and improve safety in a monotonic manner. However, its effects on high-stakes clinical decision-making remain poorly characterized. We systematically evaluate persona-based control in clinical LLMs, examining how professional roles (e.g., Emergency Department physician, nurse) and interaction styles (bold vs.\ cautious) influence behavior across models and medical tasks. We assess performance on clinical triage and patient-safety tasks using multidimensional evaluations that capture task accuracy, calibration, and safety-relevant risk behavior. We find systematic, context-dependent, and non-monotonic effects: Medical personas improve performance in critical care tasks, yielding gains of up to $\sim+20\%$ in accuracy and calibration, but degrade performance in primary-care settings by comparable margins. Interaction style modulates risk propensity and sensitivity, but it's highly model-dependent. While aggregated LLM-judge rankings favor medical over non-medical personas in safety-critical cases, we found that human clinicians show moderate agreement on safety compliance (average Cohen's $\kappa = 0.43$) but indicate a low confidence in 95.9% of their responses on reasoning quality. Our work shows that personas function as behavioral priors that introduce context-dependent trade-offs rather than guarantees of safety or expertise. The code is available at https://github.com/rsinghlab/Persona\_Paradox.
Read the full abstract

All authors Tassallah Abdullahi, Shrestha Ghosh, Hamish S Fraser, Daniel León Tramontini, Adeel Abbasi, Ghada Bourjeily, Carsten Eickhoff, Ritambhara Singh

Computation and Language

Lost in Execution: On the Multilingual Robustness of Tool Calling in Large Language Models

Zheng Luo, T Pranav Kutralingam, Ogochukwu N Okoani et al. (+3 more)

Large Language Models (LLMs) are increasingly deployed as agents that invoke external tools through structured function calls. While recent work reports strong tool-calling performance under standard English-centric evaluations, the robustness of tool calling under multilingual user interactions remains underexplored. In this work, we introduce MLCL, a diagnostic benchmark, and conduct a systematic evaluation of multilingual tool calling across Chinese, Hindi, and the low-resource language Igbo. Through fine-grained error analysis, we show that many failures occur despite correct intent understanding and tool selection. We identify parameter value language mismatch as a dominant failure mode, where models generate semantically appropriate parameter values in the user's language, violating language-invariant execution conventions. We further evaluate several inference-time system strategies and find that while these strategies substantially reduce language-induced execution errors, none of them can fully recover English-level performance.
Read the full abstract

All authors Zheng Luo, T Pranav Kutralingam, Ogochukwu N Okoani, Wanpeng Xu, Hua Wei, Xiyang Hu

Computation and Language

The Table of Media Bias Elements: A sentence-level taxonomy of media bias types and propaganda techniques

Tim Menzner and Jochen L. Leidner

Public debates about "left-" or "right-wing" news overlook the fact that bias is usually conveyed by concrete linguistic manoeuvres that transcend any single political spectrum. We therefore shift the focus from where an outlet allegedly stands to how partiality is expressed in individual sentences. Drawing on 26,464 sentences collected from newsroom corpora, user submissions and our own browsing, we iteratively combine close-reading, interdisciplinary theory and pilot annotation to derive a fine-grained, sentence-level taxonomy of media bias and propaganda. The result is a two-tier schema comprising 38 elementary bias types, arranged in six functional families and visualised as a "table of media-bias elements". For each type we supply a definition, real-world examples, cognitive and societal drivers, and guidance for recognition. A quantitative survey of a random 155-sentence sample illustrates prevalence differences, while a cross-walk to the best-known NLP and communication-science taxonomies reveals substantial coverage gains and reduced ambiguity.
Read the full abstract

Computation and Language · Cross-listed from cs.LG

TIME: Temporally Intelligent Meta-reasoning Engine for Context-Triggered Explicit Reasoning

Susmit Das

Reasoning-oriented language models typically expose explicit reasoning as a long, front-loaded chain of "thinking" tokens before the main output, either always enabled or externally toggled at inference time. Although this can help on arithmetic, coding, and other multi-step tasks, it is costly, weakens claim-level auditability, and does not allow the model to re-trigger explicit reasoning once presentation has begun. In dialogue, these limitations are compounded by weak sensitivity to temporal structure: unless time is explicitly stated in text, standard models treat replies separated by seconds and replies separated by weeks as equivalent. We introduce TIME (Temporally Intelligent Meta-reasoning Engine), a behavioral alignment framework that learns explicit reasoning as a context-triggered control policy rather than a fixed response mode. TIME augments dialogue with optional ISO 8601 <time> tags, tick events that represent silent time passage, and short <think> blocks that may appear anywhere in a response. Using a four-phase curriculum, including a small maximally diverse full-batch alignment stage, we train Qwen3 dense models to invoke brief, in-place reasoning bursts only when contextual cues warrant them, while keeping user-facing output compact. We also introduce TIMEBench, a diagnostic benchmark for evaluating reasoning from temporal cues in dialogue. Across 4B-32B scales, TIME improves TIMEBench scores over the corresponding base Qwen3 models in both thinking and no-thinking modes while reducing explicit reasoning tokens by roughly an order of magnitude. Beyond score improvements, TIME induces a distinct behavioral shift: explicit reasoning becomes more compact and more responsive to contextual cues. Code, training data, and benchmark artifacts are publicly available.
Read the full abstract

Computation and Language

GDPO: Group reward-Decoupled Normalization Policy Optimization for Multi-reward RL Optimization

Shih-Yang Liu, Xin Dong, Ximing Lu et al. (+10 more)

As language models become increasingly capable, users expect them to provide not only accurate responses but also behaviors aligned with diverse human preferences across a variety of scenarios. To achieve this, Reinforcement learning (RL) pipelines have begun incorporating multiple rewards, each capturing a distinct preference, to guide models toward these desired behaviors. However, recent work has defaulted to apply Group Relative Policy Optimization (GRPO) under multi-reward setting without examining its suitability. In this paper, we demonstrate that directly applying GRPO to normalize distinct rollout reward combinations causes them to collapse into identical advantage values, reducing the resolution of the training signal and resulting in suboptimal convergence and, in some cases, early training failure. We then introduce Group reward-Decoupled Normalization Policy Optimization (GDPO), a new policy optimization method to resolve these issues by decoupling the normalization of individual rewards, more faithfully preserving their relative differences and enabling more accurate multi-reward optimization, along with substantially improved training stability. We compare GDPO with GRPO across three tasks: tool calling, math reasoning, and coding reasoning, evaluating both correctness metrics (accuracy, bug ratio) and constraint adherence metrics (format, length). Across all settings, GDPO consistently outperforms GRPO, demonstrating its effectiveness and generalizability for multi-reward reinforcement learning optimization.
Read the full abstract

All authors Shih-Yang Liu, Xin Dong, Ximing Lu, Shizhe Diao, Peter Belcak, Mingjie Liu, Min-Hung Chen, Hongxu Yin, Yu-Chiang Frank Wang, Kwang-Ting Cheng, Yejin Choi, Jan Kautz, Pavlo Molchanov

Computation and Language

AI Application Gives Users Real-Time Feedback on the Level of Peace in the Social Media Videos They Watch

P. Gilda, P. Dungarwal, A. Thongkham et al. (+11 more)

Most people now get their news from videos on social media, such as YouTube and Facebook, rather than through curated journalism. "We become what we behold." The content and tone of language plays an essential role in starting or ending conflicts. "Hate Speech" can enhance conflict, "Peace Speech" can enhance peace. We developed an application that measures, in real time, these aspects of speech from YouTube videos, which can give users helpful feedback on their own media diet. We used two approaches: 1) supervised machine learning. Language in the text of online news media text was tagged by surveys that measure the level of peace in those countries. One fully connected feedforward and 2 convolutional neural networks trained on that data were $\sim 97\%$ accurate in predicting levels of peace in the test set and $\sim 70\%$ accurate in another distinct news text data set, but did not generalize to YouTube videos, suggesting that written text is different than transcribed spoken language. 2) social science dimensions. There is no similar external data to tag the text in the YouTube video transcripts. We therefore used 2 word-level sentiment analysis (SA) and 6 context-level large language models (LLMs) to measure 5 social dimensions in peace identified by 59 social science studies: compassion-contempt, news-opinion, promotion-prevention, creativity-order, nuance-simplification. LLMs more closely matched the values by 3 human coders on 52 videos, $r^2\sim0.60$ than SA, at $r^2\sim0.03$. Results: LLMs successfully measured social dimensions important in peace in YouTube videos, compared to human coders. These results serve as the basis of an analysis engine that can give users and content creators feedback on their own media diet and creations.
Read the full abstract

All authors P. Gilda, P. Dungarwal, A. Thongkham, E. T. Ajayi, S. Choudhary, T. M. Terol, C. Lam, J. P. Araujo, M. McFadyen-Mungalln, L. S. Liebovitch, P. T. Coleman, H. West, K. Sieck, S. Carter

Computation and Language · Cross-listed from cs.CV

Mechanisms of Prompt-Induced Hallucination in Vision-Language Models

William Rudman, Michal Golovanevsky, Dana Arad et al. (+4 more)

Large vision-language models (VLMs) are highly capable, yet often hallucinate by favoring textual prompts over visual evidence. We study this failure mode in a controlled object-counting setting, where the prompt overstates the number of objects in the image (e.g., asking a model to describe four waterlilies when only three are present). At low object counts, models often correct the overestimation, but as the number of objects increases, they increasingly conform to the prompt regardless of the discrepancy. Through mechanistic analysis of three VLMs, we identify a small set of attention heads whose ablation substantially reduces prompt-induced hallucinations (PIH) by at least 40% without additional training. Across models, PIH-heads mediate prompt copying in model-specific ways. We characterize these differences and show that PIH ablation increases correction toward visual evidence. Our findings offer insights into the internal mechanisms driving prompt-induced hallucinations, revealing model-specific differences in how these behaviors are implemented.
Read the full abstract

All authors William Rudman, Michal Golovanevsky, Dana Arad, Yonatan Belinkov, Ritambhara Singh, Carsten Eickhoff, Kyle Mahowald

Computation and Language

LELA: an LLM-based Entity Linking Approach with Zero-Shot Domain Adaptation

Samy Haffoudhi, Fabian M. Suchanek and Nils Holzenberger

Entity linking (mapping ambiguous mentions in text to entities in a knowledge base) is a foundational step in tasks such as knowledge graph construction, question-answering, and information extraction. Our method, LELA, is a modular coarse-to-fine approach that leverages the capabilities of large language models (LLMs), and works with different target domains, knowledge bases and LLMs, without any fine-tuning phase. Our experiments across various entity linking settings show that LELA is highly competitive with fine-tuned approaches, and substantially outperforms the non-fine-tuned ones.
Read the full abstract

Computation and Language · Cross-listed from cs.AI

Observations and Remedies for Large Language Model Bias in Self-Consuming Performative Loop

Yaxuan Wang, Zhongteng Cai, Yujia Bao et al. (+2 more)

The rapid advancement of large language models (LLMs) has led to growing interest in using synthetic data to train future models. However, this creates a self-consuming retraining loop, where models are trained on their own outputs and may cause performance drops and induce emerging biases. In real-world applications, previously deployed LLMs may influence the data they generate, leading to a dynamic system driven by user feedback. For example, if a model continues to underserve users from a group, less query data will be collected from this particular demographic of users. In this study, we introduce the concept of Self-Consuming Performative Loop (SCPL) and investigate the role of synthetic data in shaping bias during these dynamic iterative training processes under controlled performative feedback. This controlled setting is motivated by the inaccessibility of real-world user preference data from dynamic production systems, and enables us to isolate and analyze feedback-driven bias evolution in a principled manner. We focus on two types of loops, including the typical retraining setting and the incremental fine-tuning setting, which is largely underexplored. Through experiments on three real-world tasks, we find that the performative loop increases preference bias and decreases disparate bias. We design a reward-based rejection sampling strategy to mitigate the bias, moving towards more trustworthy self-improving systems.
Read the full abstract

All authors Yaxuan Wang, Zhongteng Cai, Yujia Bao, Xueru Zhang, Yang Liu

Computation and Language

Inside Out: Evolving User-Centric Core Memory Trees for Long-Term Personalized Dialogue Systems

Jihao Zhao, Ding Chen, Zhaoxin Fan et al. (+5 more)

Existing long-term personalized dialogue systems struggle to reconcile unbounded interaction streams with finite context constraints, often succumbing to memory noise accumulation, reasoning degradation, and persona inconsistency. To address these challenges, this paper proposes Inside Out, a framework that utilizes a globally maintained PersonaTree as the carrier of long-term user profiling. By constraining the trunk with an initial schema and updating the branches and leaves, PersonaTree enables controllable growth, achieving memory compression while preserving consistency. Moreover, we train a lightweight MemListener via reinforcement learning with process-based rewards to produce structured, executable, and interpretable {ADD, UPDATE, DELETE, NO_OP} operations, thereby supporting the dynamic evolution of the personalized tree. During response generation, PersonaTree is directly leveraged to enhance outputs in latency-sensitive scenarios; when users require more details, the agentic mode is triggered to introduce details on-demand under the constraints of the PersonaTree. Experiments show that PersonaTree outperforms full-text concatenation and various personalized memory systems in suppressing contextual noise and maintaining persona consistency. Notably, the small MemListener model achieves memory-operation decision performance comparable to, or even surpassing, powerful reasoning models such as DeepSeek-R1-0528 and Gemini-3-Pro.
Read the full abstract

All authors Jihao Zhao, Ding Chen, Zhaoxin Fan, Kerun Xu, Mengting Hu, Bo Tang, Feiyu Xiong, Zhiyu Li

Computation and Language

Reverse-engineering NLI: A study of the meta-inferential properties of Natural Language Inference

Rasmus Blanck, Bill Noble and Stergios Chatzikyriakidis

Natural Language Inference (NLI) has been an important task for evaluating language models for Natural Language Understanding, but the logical properties of the task are poorly understood and often mischaracterized. Understanding the notion of inference captured by NLI is key to interpreting model performance on the task. In this paper we formulate three possible readings of the NLI label set and perform a comprehensive analysis of the meta-inferential properties they entail. Focusing on the SNLI dataset, we exploit (1) NLI items with shared premises and (2) items generated by LLMs to evaluate models trained on SNLI for meta-inferential consistency and derive insights into which reading of the logical relations is encoded by the dataset.
Read the full abstract

Computation and Language

RelayLLM: Efficient Reasoning via Collaborative Decoding

Chengsong Huang, Tong Zheng, Langlin Huang et al. (+3 more)

Large Language Models (LLMs) for complex reasoning is often hindered by high computational costs and latency, while resource-efficient Small Language Models (SLMs) typically lack the necessary reasoning capacity. Existing collaborative approaches, such as cascading or routing, operate at a coarse granularity by offloading entire queries to LLMs, resulting in significant computational waste when the SLM is capable of handling the majority of reasoning steps. To address this, we propose RelayLLM, a novel framework for efficient reasoning via token-level collaborative decoding. Unlike routers, RelayLLM empowers the SLM to act as an active controller that dynamically invokes the LLM only for critical tokens via a special command, effectively "relaying" the generation process. We introduce a two-stage training framework, including warm-up and Group Relative Policy Optimization (GRPO) to teach the model to balance independence with strategic help-seeking. Empirical results across six benchmarks demonstrate that RelayLLM achieves an average accuracy of 49.52%, effectively bridging the performance gap between the two models. Notably, this is achieved by invoking the LLM for only 1.07% of the total generated tokens, offering a 98.2% cost reduction compared to performance-matched random routers.
Read the full abstract

All authors Chengsong Huang, Tong Zheng, Langlin Huang, Jinyuan Li, Haolin Liu, Jiaxin Huang

Computation and Language

DocDancer: Towards Agentic Document-Grounded Information Seeking

Qintong Zhang, Xinjie Lv, Jialong Wu et al. (+8 more)

Document Question Answering (DocQA) focuses on answering questions grounded in given documents, yet existing DocQA agents lack effective tool utilization and largely rely on closed-source models. In this work, we introduce DocDancer, an end-to-end trained open-source Doc agent. We formulate DocQA as an information-seeking problem and propose a tool-driven agent framework that explicitly models document exploration and comprehension. To enable end-to-end training of such agents, we introduce an Exploration-then-Synthesis data synthesis pipeline that addresses the scarcity of high-quality training data for DocQA. Training on the synthesized data, the trained models on two long-context document understanding benchmarks, MMLongBench-Doc and DocBench, show their effectiveness. Further analysis provides valuable insights for the agentic tool design and synthetic data.
Read the full abstract

All authors Qintong Zhang, Xinjie Lv, Jialong Wu, Baixuan Li, Zhengwei Tao, Guochen Yan, Huanyao Zhang, Bin Wang, Jiahao Xu, Haitao Mi, Wentao Zhang

Computation and Language · Cross-listed from cs.CV

A Two-Stage Multitask Vision-Language Framework for Explainable Crop Disease Visual Question Answering

Md. Zahid Hossain, Most. Sharmin Sultana Samu, Md. Rakibul Islam et al. (+1 more)

Visual question answering (VQA) for crop disease analysis requires accurate visual understanding and reliable language generation. In this work, we present a lightweight and explainable vision-language framework for crop and disease identification from leaf images. The proposed approach integrates a Swin Transformer vision encoder with sequence-to-sequence language decoders. The vision encoder is first trained in a multitask setup for both plant and disease classification, and then frozen while the text decoders are trained, forming a two-stage training strategy that enhances visual representation learning and cross-modal alignment. We evaluate the model on the large-scale Crop Disease Domain Multimodal (CDDM) dataset using both classification and natural language generation metrics. Experimental results demonstrate near-perfect recognition performance, achieving 99.94% plant classification accuracy and 99.06% disease classification accuracy, along with strong BLEU, ROUGE and BERTScore results. Without fine-tuning, the model further generalizes well to the external PlantVillageVQA benchmark, achieving 83.18% micro accuracy in the VQA task. Our lightweight design outperforms larger vision-language baselines while using significantly fewer parameters. Explainability is assessed through Grad-CAM and token-level attribution, providing interpretable visual and textual evidence for predictions. Qualitative results demonstrate robust performance under diverse user-driven queries, highlighting the effectiveness of task-specific visual pretraining and the two-stage training methodology for crop disease visual question answering. An interactive demo of the proposed Swin-T5 model is publicly available as a Gradio-based application at https://huggingface.co/spaces/Zahid16/PlantDiseaseVQAwithSwinT5 for community use.
Read the full abstract

All authors Md. Zahid Hossain, Most. Sharmin Sultana Samu, Md. Rakibul Islam, Md. Siam Ansary

Computation and Language

Agent-as-a-Judge

Runyang You, Hongru Cai, Caiqi Zhang et al. (+5 more)

LLM-as-a-Judge has revolutionized AI evaluation by leveraging large language models for scalable assessments. However, as evaluands become increasingly complex, specialized, and multi-step, the reliability of LLM-as-a-Judge has become constrained by inherent biases, shallow single-pass reasoning, and the inability to verify assessments against real-world observations. This has catalyzed the transition to Agent-as-a-Judge, where agentic judges employ planning, tool-augmented verification, multi-agent collaboration, and persistent memory to enable more robust, verifiable, and nuanced evaluations. Despite the rapid proliferation of agentic evaluation systems, the field lacks a unified framework to navigate this shifting landscape. To bridge this gap, we present the first comprehensive survey tracing this evolution. Specifically, we identify key dimensions that characterize this paradigm shift and establish a developmental taxonomy. We organize core methodologies and survey applications across general and professional domains. Furthermore, we analyze frontier challenges and identify promising research directions, ultimately providing a clear roadmap for the next generation of agentic evaluation.
Read the full abstract

All authors Runyang You, Hongru Cai, Caiqi Zhang, Qiancheng Xu, Meng Liu, Tiezheng Yu, Yongqi Li, Wenjie Li

Computation and Language · Cross-listed from cs.AI

Token-Level LLM Collaboration via FusionRoute

Nuoya Xiong, Yuhang Zhou, Hanqing Zeng et al. (+5 more)

Large language models (LLMs) exhibit strengths across diverse domains. However, achieving strong performance across these domains with a single general-purpose model typically requires scaling to sizes that are prohibitively expensive to train and deploy. On the other hand, while smaller domain-specialized models are much more efficient, they struggle to generalize beyond their training distributions. To address this dilemma, we propose FusionRoute, a robust and effective token-level multi-LLM collaboration framework in which a lightweight router simultaneously (i) selects the most suitable expert at each decoding step and (ii) contributes a complementary logit that refines or corrects the selected expert's next-token distribution via logit addition. Unlike existing token-level collaboration methods that rely solely on fixed expert outputs, we provide a theoretical analysis showing that pure expert-only routing is fundamentally limited: unless strong global coverage assumptions hold, it cannot in general realize the optimal decoding policy. By augmenting expert selection with a trainable complementary generator, FusionRoute expands the effective policy class and enables recovery of optimal value functions under mild conditions. Empirically, across both Llama-3 and Gemma-2 families and diverse benchmarks spanning mathematical reasoning, code generation, and instruction following, FusionRoute outperforms both sequence- and token-level collaboration, model merging, and direct fine-tuning, while remaining competitive with domain experts on their respective tasks.
Read the full abstract

All authors Nuoya Xiong, Yuhang Zhou, Hanqing Zeng, Zhaorun Chen, Furong Huang, Shuchao Bi, Lizhu Zhang, Zhuokai Zhao

Computation and Language

How Human is AI? Examining the Impact of Emotional Prompts on Artificial and Human and Responsiveness

Florence Bernays, Marco Henriques Pereira and Jochen Menges

This research examines how the emotional tone of human-AI interactions shapes ChatGPT and human behavior. In a between-subject experiment, we asked participants to express a specific emotion while working with ChatGPT (GPT-4.0) on two tasks, including writing a public response and addressing an ethical dilemma. We found that compared to interactions where participants maintained a neutral tone, ChatGPT showed greater improvement in its answers when participants praised ChatGPT for its responses. Expressing anger towards ChatGPT also led to a higher albeit smaller improvement relative to the neutral condition, whereas blaming ChatGPT did not improve its answers. When addressing an ethical dilemma, ChatGPT prioritized corporate interests less when participants expressed anger towards it, while blaming increases its emphasis on protecting the public interest. Additionally, we found that people used more negative, hostile, and disappointing expressions in human-human communication after interactions during which participants blamed rather than praised for their responses. Together, our findings demonstrate that the emotional tone people apply in human-AI interactions not only shape ChatGPT's outputs but also carry over into subsequent human-human communication.
Read the full abstract

Computation and Language · Cross-listed from cs.DL

Semantically Orthogonal Framework for Citation Classification: Disentangling Intent and Content

Changxu Duan and Zhiyin Tan

Understanding the role of citations is essential for research assessment and citation-aware digital libraries. However, existing citation classification frameworks often conflate citation intent (why a work is cited) with cited content type (what part is cited), limiting their effectiveness in auto classification due to a dilemma between fine-grained type distinctions and practical classification reliability. We introduce SOFT, a Semantically Orthogonal Framework with Two dimensions that explicitly separates citation intent from cited content type, drawing inspiration from semantic role theory. We systematically re-annotate the ACL-ARC dataset using SOFT and release a cross-disciplinary test set sampled from ACT2. Evaluation with both zero-shot and fine-tuned Large Language Models demonstrates that SOFT enables higher agreement between human annotators and LLMs, and supports stronger classification performance and robust cross-domain generalization compared to ACL-ARC and SciCite annotation frameworks. These results confirm SOFT's value as a clear, reusable annotation standard, improving clarity, consistency, and generalizability for digital libraries and scholarly communication infrastructures. All code and data are publicly available on GitHub https://github.com/zhiyintan/SOFT.
Read the full abstract

Computation and Language · Cross-listed from cs.DL

Multi-Disciplinary Dataset Discovery from Citation-Verified Literature Contexts

Zhiyin Tan and Changxu Duan

Identifying suitable datasets for a research question remains challenging because existing dataset search engines rely heavily on metadata quality and keyword overlap, which often fail to capture the semantic intent of scientific investigation. We introduce a literature-driven framework that discovers datasets from citation contexts in scientific papers, enabling retrieval grounded in actual research use rather than metadata availability. Our approach combines large-scale citation-context extraction, schema-guided dataset recognition with Large Language Models, and provenance-preserving entity resolution. We evaluate the system on eight survey-derived computer science queries and find that it achieves substantially higher recall than Google Dataset Search and DataCite Commons, with normalized recall ranging from an average of 47.47% to a highest value of 81.82%. Beyond recovering gold-standard datasets, the method also surfaces additional datasets not documented in the surveys. Expert assessments across five top-level Fields of Science indicate that a substantial portion of the additional datasets are considered high utility, and some are regarded as novel for the specific topics chosen by the experts. These findings establish citation-context mining as an effective and generalizable paradigm for dataset discovery, particularly in settings where datasets lack sufficient or reliable metadata. To support reproducibility and future extensions, we release our code, evaluation datasets, and results on GitHub (https://github.com/Fireblossom/citation-context-dataset-discovery).
Read the full abstract

Computation and Language

Code-Mix Sentiment Analysis on Hinglish Tweets

Aashi Garg, Aneshya Das, Arshi Arya et al. (+2 more)

The effectiveness of brand monitoring in India is increasingly challenged by the rise of Hinglish–a hybrid of Hindi and English–used widely in user-generated content on platforms like Twitter. Traditional Natural Language Processing (NLP) models, built for monolingual data, often fail to interpret the syntactic and semantic complexity of this code-mixed language, resulting in inaccurate sentiment analysis and misleading market insights. To address this gap, we propose a high-performance sentiment classification framework specifically designed for Hinglish tweets. Our approach fine-tunes mBERT (Multilingual BERT), leveraging its multilingual capabilities to better understand the linguistic diversity of Indian social media. A key component of our methodology is the use of subword tokenization, which enables the model to effectively manage spelling variations, slang, and out-of-vocabulary terms common in Romanized Hinglish. This research delivers a production-ready AI solution for brand sentiment tracking and establishes a strong benchmark for multilingual NLP in low-resource, code-mixed environments.
Read the full abstract

All authors Aashi Garg, Aneshya Das, Arshi Arya, Anushka Goyal, Aditi