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

Tuesday 27 January

Computation and Language

When Benchmarks Leak: Inference-Time Decontamination for LLMs

Jianzhe Chai, Yu Zhe and Jun Sakuma

Benchmark-based evaluation is the de facto standard for comparing large language models (LLMs). However, its reliability is increasingly threatened by test set contamination, where test samples or their close variants leak into training data and artificially inflate reported performance. To address this issue, prior work has explored two main lines of mitigation. One line attempts to identify and remove contaminated benchmark items before evaluation, but this inevitably alters the evaluation set itself and becomes unreliable when contamination is moderate or severe. The other line preserves the benchmark and instead suppresses contaminated behavior at evaluation time; however, such interventions often interfere with normal inference and lead to noticeable performance degradation on clean inputs. We propose DeconIEP, a decontamination framework that operates entirely during evaluation by applying small, bounded perturbations in the input embedding space. Guided by a relatively less-contaminated reference model, DeconIEP learns an instance-adaptive perturbation generator that steers the evaluated model away from memorization-driven shortcut pathways. Across multiple open-weight LLMs and benchmarks, extensive empirical results show that DeconIEP achieves strong decontamination effectiveness while incurring only minimal degradation in benign utility.
Read the full abstract

Computation and Language

Formula-One Prompting: A Composable Equation-First Prefix for Applied Mathematics

Natapong Nitarach, Pittawat Taveekitworachai and Kunat Pipatanakul

This paper introduces Formula Prompting (FP) and Formula-One Prompting (F-1), two single-call methods that elicit governing equations before solving applied-math problems. Chain-of-Thought (CoT) and Program-of-Thought (PoT) prompting improve mathematical reasoning by eliciting reasoning traces or code-like structures learned during pretraining. This suggests a diagnostic question: which useful pretraining patterns remain under-elicited? Using infini-gram-mini, we scan 81.7 trillion pretraining tokens and find that, in curated corpora such as DataComp-LM, equation-centered language appears 121x more often than code and 3.79x more often than step-by-step narration, yet standard prompting methods do not explicitly elicit equation formulation. FP asks the model to formalize a problem's governing equations before solving; F-1 extends FP with a composable Phase 2 that selects Direct, CoT, or PoT-style solving in the same call. Across five reasoning models and four applied-math benchmarks (finance, physics, cryptography, competition math), F-1 outperforms CoT by 5.76 pp and PoT by 8.42 pp on average, with the largest gain of 13.30 pp on FinanceMath, while topping the accuracy-token efficiency frontier at only 68 prompt tokens of overhead. Variant ablations identify the equation-formalization prefix, not the strategy menu, as the primary driver: adding CoT or PoT on top of the prefix yields no further gain, and 73.3% of remaining failures occur downstream of a correct Phase-1 equation.
Read the full abstract

Computation and Language

MetaGen: Self-Evolving Roles and Topologies for Multi-Agent LLM Reasoning

Yimeng Wang, Jiaxing Zhao, Hongbin Xie et al. (+6 more)

Large language models are increasingly deployed as multi-agent systems, where specialized roles communicate and collaborate through structured interactions to solve complex tasks that often exceed the capacity of a single agent. However, most existing systems still rely on a fixed role library and an execution-frozen interaction topology, a rigid design choice that frequently leads to task mismatch, prevents timely adaptation when new evidence emerges during reasoning, and further inflates inference cost. We introduce MetaGen, a training-free framework that adapts both the role space and the collaboration topology at inference time, without updating base model weights. MetaGen generates and rewrites query-conditioned role specifications to maintain a controllable dynamic role pool, then instantiates a constrained execution graph around a minimal backbone. During execution, it iteratively updates role prompts and adjusts structural decisions using lightweight feedback signals. Experiments on code generation and multi-step reasoning benchmarks show that MetaGen improves the accuracy and cost tradeoff over strong multi-agent baselines.
Read the full abstract

All authors Yimeng Wang, Jiaxing Zhao, Hongbin Xie, Hexing Ma, Yuzhen Lei, Shuangxue Liu, Xuan Song, Zichen Zhang, Haoran Zhang

Computation and Language

ReToP: Learning to Rewrite Electronic Health Records for Clinical Prediction

Jesus Lovon-Melgarejo, Jose G. Moreno, Christine Damase-Michel et al. (+1 more)

Electronic Health Records (EHRs) provide crucial information for clinical decision-making. However, their high-dimensionality, heterogeneity, and sparsity make clinical prediction challenging. Large Language Models (LLMs) allowed progress towards addressing this challenge by leveraging parametric medical knowledge to enhance EHR data for clinical prediction tasks. Despite the significant achievements made so far, most of the existing approaches are fundamentally task-agnostic in the sense that they deploy LLMs as EHR encoders or EHR completion modules without fully integrating signals from the prediction tasks. This naturally hinders task performance accuracy. In this work, we propose Rewrite-To-Predict (ReToP), an LLM-based framework that addresses this limitation through an end-to-end training of an EHR rewriter and a clinical predictor. To cope with the lack of EHR rewrite training data, we generate synthetic pseudo-labels using clinical-driven feature selection strategies to create diverse patient rewrites for fine-tuning the EHR rewriter. ReToP aligns the rewriter with prediction objectives using a novel Classifier Supervised Contribution (CSC) score that enables the EHR rewriter to generate clinically relevant rewrites that directly enhance prediction. Our ReToP framework surpasses strong baseline models across three clinical tasks on MIMIC-IV. Moreover, the analysis of ReToP shows its generalizability to unseen datasets and tasks with minimal fine-tuning while preserving faithful rewrites and emphasizing task-relevant predictive features.
Read the full abstract

All authors Jesus Lovon-Melgarejo, Jose G. Moreno, Christine Damase-Michel, Lynda Tamine

Computation and Language · Cross-listed from cs.LG

Group Distributionally Robust Optimization-Driven Reinforcement Learning for LLM Reasoning

Kishan Panaganti, Zhenwen Liang, Wenhao Yu et al. (+2 more)

Recent progress in Large Language Model (LLM) reasoning is increasingly driven by the refinement of post-training loss functions and alignment strategies. However, standard Reinforcement Learning (RL) paradigms like Group Relative Policy Optimization (GRPO) remain constrained by static uniformity: uniform prompt sampling and a fixed number of rollouts per prompt. For heterogeneous, heavy-tailed reasoning data, this creates structural inefficiencies that waste compute on already-solved patterns while under-training the long tail of hard problems. To address this, we propose Multi-Adversary Group Distributionally Robust Optimization (GDRO), an optimization-first framework that moves beyond uniform reasoning models by dynamically adapting the training distribution. We introduce an Online Difficulty Classifier that partitions prompts into dynamic pass@k difficulty groups. We then propose two independent GDRO games for post-training: (1) Prompt-GDRO, which employs an EMA-debiased multiplicative-weights bandit sampler to target the intensive difficulty margin and upweight persistently hard groups without frequency bias; and (2) Rollout-GDRO, which uses a shadow-price controller to reallocate rollouts across groups, maximizing gradient variance reduction on hard tasks under a fixed mean budget (compute-neutral). We provide no-regret guarantees for both controllers and additionally a variance-proxy analysis motivating a square-root optimal rollout allocation for Rollout-GDRO. We validate our framework on the DAPO 14.1k dataset using Qwen3-Base models. Prompt-GDRO and Rollout-GDRO achieve average relative gains of +10.6% and +10.1%, respectively, in pass@8 accuracy across 1.7B, 4B, and 8B scales compared to the GRPO baseline. Qualitative analysis shows an emergent curriculum: the adversaries shift resources to the evolving reasoning frontier, enhancing the reasoning model's performance.
Read the full abstract

All authors Kishan Panaganti, Zhenwen Liang, Wenhao Yu, Haitao Mi, Dong Yu

Computation and Language

DART: Diffusion-Inspired Speculative Decoding for Fast LLM Inference

Fuliang Liu, Xue Li, Ketai Zhao et al. (+7 more)

Speculative decoding is an effective and lossless approach for accelerating LLM inference. However, existing widely adopted model-based draft designs, such as EAGLE3, improve accuracy at the cost of multi-step autoregressive inference, resulting in high drafting latency and ultimately rendering the drafting stage itself a performance bottleneck. Inspired by diffusion-based large language models (dLLMs), we propose DART, which leverages parallel generation to reduce drafting latency. DART predicts logits for multiple future masked positions in parallel within a single forward pass based on hidden states of the target model, thereby eliminating autoregressive rollouts in the draft model while preserving a lightweight design. Based on these parallel logit predictions, we further introduce an efficient tree pruning algorithm that constructs high-quality draft token trees with N-gram-enforced semantic continuity. DART substantially reduces draft-stage overhead while preserving high draft accuracy, leading to significantly improved end-to-end decoding speed. Experimental results demonstrate that DART achieves a 2.03x–3.44x wall-clock time speedup across multiple datasets, surpassing EAGLE3 by 30% on average and offering a practical speculative decoding framework. Code is released at https://github.com/fvliang/DART.
Read the full abstract

All authors Fuliang Liu, Xue Li, Ketai Zhao, Yinxi Gao, Ziyan Zhou, Zhonghui Zhang, Zhibin Wang, Wanchun Dou, Sheng Zhong, Chen Tian

Computation and Language

Riddle Quest : The Enigma of Words

Niharika Sri Parasa, Chaitali Diwan and Srinath Srinivasa

Riddles are concise linguistic puzzles that describe an object or idea through indirect, figurative, or playful clues. They are a longstanding form of creative expression, requiring the solver to interpret hints, recognize patterns, and draw inferences to identify the answers. In this work, we introduce a simple pipeline for creating and evaluating analogy-based riddles. The system includes a triples creator that builds structured facts about a concept, a semantic mapper that selects attributes useful for analogy, a stylized generator that turns them into riddle clues, and a validator that collects all possible answers the riddle could point to. We use this validator to study whether large language models can recover the full answer set for different riddle types. Our case study shows that while models often guess the main intended answer, they frequently miss other valid interpretations. This highlights the value of riddles as a lightweight tool for examining reasoning coverage and ambiguity handling in language models.
Read the full abstract

Computation and Language

DiaDem: Advancing Dialogue Descriptions in Audiovisual Video Captioning for Multimodal Large Language Models

Xinlong Chen, Weihong Lin, Jingyun Hua et al. (+10 more)

Accurate dialogue description in audiovisual video captioning is crucial for downstream understanding and generation tasks. However, existing models generally struggle to produce faithful dialogue descriptions within audiovisual captions. To mitigate this limitation, we propose DiaDem, a powerful audiovisual video captioning model capable of generating captions with more precise dialogue descriptions while maintaining strong overall performance. We first synthesize a high-quality dataset for SFT, then employ a difficulty-partitioned two-stage GRPO strategy to further enhance dialogue descriptions. To enable systematic evaluation of dialogue description capabilities, we introduce DiaDemBench, a comprehensive benchmark designed to evaluate models across diverse dialogue scenarios, emphasizing both speaker attribution accuracy and utterance transcription fidelity in audiovisual captions. Extensive experiments on DiaDemBench reveal even commercial models still exhibit substantial room for improvement in dialogue-aware captioning. Notably, DiaDem not only outperforms the Gemini series in dialogue description accuracy but also achieves competitive performance on general audiovisual captioning benchmarks, demonstrating its overall effectiveness.
Read the full abstract

All authors Xinlong Chen, Weihong Lin, Jingyun Hua, Linli Yao, Yue Ding, Bozhou Li, Bohan Zeng, Yang Shi, Qiang Liu, Yuanxing Zhang, Pengfei Wan, Liang Wang, Tieniu Tan

Computation and Language

RPO-RAG: Aligning Small LLMs with Relation-aware Preference Optimization for Knowledge Graph Question Answering

Kaehyun Um, KyuHwan Yeom, Haerim Yang et al. (+3 more)

Large Language Models (LLMs) have recently demonstrated remarkable reasoning abilities, yet hallucinate on knowledge-intensive tasks. Retrieval-augmented generation (RAG) mitigates this issue by grounding answers in external sources, e.g., knowledge graphs (KGs). However, existing KG-based RAG approaches rely on semantics-unaware path sampling and are weakly aligned with KG reasoning objectives, which limits further accuracy gains. They also feed retrieved paths directly into the reasoner without organizing them into answer-centered reasoning paths, hindering small LLMs' ability to leverage the retrieved knowledge. Furthermore, prior works predominantly rely on large LLMs (e.g., ChatGPT/GPT-4) or assume backbones above 7B parameters, leaving sub-7B models underexplored. We address this gap with RPO-RAG, the first KG-based RAG framework specifically designed for small LLMs, to the best of our knowledge. RPO-RAG introduces three key innovations: (1) a query-path semantic sampling strategy that provides informative supervisory signals; (2) a relation-aware preference optimization that aligns training with intermediate KG reasoning signals (e.g., relation); and (3) an answer-centered prompt design that organizes entities and reasoning paths in an interpretable format. Extensive experiments on two benchmark Knowledge Graph Question Answering (KGQA) datasets, WebQSP and CWQ, demonstrate that RPO-RAG effectively bridges the performance gap between small and large language models. On WebQSP, it improves F1 by up to 8.8%, reflecting enhanced answer precision, while on CWQ it achieves new state-of-the-art results among models under 8B parameters in both Hit and F1. Overall, RPO-RAG substantially improves the reasoning capability of small LLMs, even under 3B parameters-highlighting their potential for resource-efficient and practical on-device KGQA applications.
Read the full abstract

All authors Kaehyun Um, KyuHwan Yeom, Haerim Yang, Minyoung Choi, Hyeongjun Yang, Kyong-Ho Lee

Computation and Language

DREAMSTATE: Diffusing States and Parameters for Recurrent Large Language Models

Liu Xiao

Modern Recurrent Neural Networks (RNNs), such as RWKV, are distinguished by their powerful short-range modeling capabilities and efficient fixed-size states, which constitute a core advantage over standard Transformers. However, there is a significant lack of research into their internal state as an editable knowledge representation. To fill this gap, we first explore the representational properties of the RWKV state by proposing the DREAMSTATE framework. This framework utilizes a conditional Diffusion Transformer (DiT) to directly model the probability manifold of the state, enabling its generation and editing. The structural nature of this representation is validated through t-SNE visualizations and controlled generation experiments. After successfully uncovering and modeling the state's representational potential, we further propose a novel hybrid architecture that combines the local advantages of RNNs with global context adaptability. This architecture features a parallel DiT that processes a variable-length global context to dynamically generate and adjust the core recurrent module's WKV parameters, transforming the fixed recurrence mechanism into a context-aware dynamic function. Experiments demonstrate that this hybrid model can be trained stably via a multi-objective loss, validating its design feasibility. Our work not only opens a new research direction for RNN state representation but also provides a concrete architectural reference for future model design. The code is publicly available at: https://huggingface.co/2dgx41s/DreamState.
Read the full abstract

Computation and Language

A Hybrid Supervised-LLM Pipeline for Actionable Suggestion Mining in Unstructured Customer Reviews

Aakash Trivedi, Aniket Upadhyay, Pratik Narang et al. (+2 more)

Extracting actionable suggestions from customer reviews is essential for operational decision-making, yet these directives are often embedded within mixed-intent, unstructured text. Existing approaches either classify suggestion-bearing sentences or generate high-level summaries, but rarely isolate the precise improvement instructions businesses need. We evaluate a hybrid pipeline combining a high-recall RoBERTa classifier trained with a precision-recall surrogate to reduce unrecoverable false negatives with a controlled, instruction-tuned LLM for suggestion extraction, categorization, clustering, and summarization. Across real-world hospitality and food datasets, the hybrid system outperforms prompt-only, rule-based, and classifier-only baselines in extraction accuracy and cluster coherence. Human evaluations further confirm that the resulting suggestions and summaries are clear, faithful, and interpretable. Overall, our results show that hybrid reasoning architectures achieve meaningful improvements fine-grained actionable suggestion mining while highlighting challenges in domain adaptation and efficient local deployment.
Read the full abstract

All authors Aakash Trivedi, Aniket Upadhyay, Pratik Narang, Dhruv Kumar, Praveen Kumar

Computation and Language

How Do Transformers Learn to Associate Tokens: Gradient Leading Terms Bring Mechanistic Interpretability

Shawn Im, Changdae Oh, Zhen Fang et al. (+1 more)

Semantic associations such as the link between "bird" and "flew" are foundational for language modeling as they enable models to go beyond memorization and instead generalize and generate coherent text. Understanding how these associations are learned and represented in language models is essential for connecting deep learning with linguistic theory and developing a mechanistic foundation for large language models. In this work, we analyze how these associations emerge from natural language data in attention-based language models through the lens of training dynamics. By leveraging a leading-term approximation of the gradients, we develop closed-form expressions for the weights at early stages of training that explain how semantic associations first take shape. Through our analysis, we reveal that each set of weights of the transformer has closed-form expressions as simple compositions of three basis functions (bigram, token-interchangeability, and context mappings), reflecting the statistics of the text corpus and uncovering how each component of the transformer captures semantic associations based on these compositions. Experiments on real-world LLMs demonstrate that our theoretical weight characterizations closely match the learned weights, and qualitative analyses further show how our theorem shines light on interpreting the learned associations in transformers.
Read the full abstract

All authors Shawn Im, Changdae Oh, Zhen Fang, Sharon Li

Computation and Language

Do Images Speak Louder than Words? Investigating the Effect of Textual Misinformation in VLMs

Chi Zhang, Wenxuan Ding, Jiale Liu et al. (+3 more)

Vision-Language Models (VLMs) have shown strong multimodal reasoning capabilities on Visual-Question-Answering (VQA) benchmarks. However, their robustness against textual misinformation remains under-explored. While existing research has studied the effect of misinformation in text-only domains, it is not clear how VLMs arbitrate between contradictory information from different modalities. To bridge the gap, we first propose the CONTEXT-VQA (i.e., Conflicting Text) dataset, consisting of image-question pairs together with systematically generated persuasive prompts that deliberately conflict with visual evidence. Then, a thorough evaluation framework is designed and executed to benchmark the susceptibility of various models to these conflicting multimodal inputs. Comprehensive experiments over 11 state-of-the-art VLMs reveal that these models are indeed vulnerable to misleading textual prompts, often overriding clear visual evidence in favor of the conflicting text, and show an average performance drop of over 48.2% after only one round of persuasive conversation. Our findings highlight a critical limitation in current VLMs and underscore the need for improved robustness against textual manipulation.
Read the full abstract

All authors Chi Zhang, Wenxuan Ding, Jiale Liu, Mingrui Wu, Qingyun Wu, Ray Mooney

Computation and Language

Transparency-First Medical Language Models: Datasheets, Model Cards, and End-to-End Data Provenance for Clinical NLP

Olaf Yunus Laitinen Imanov, Taner Yilmaz, Ayse Tuba Tugrul et al. (+11 more)

We introduce TeMLM, a set of transparency-first release artifacts for clinical language models. TeMLM unifies provenance, data transparency, modeling transparency, and governance into a single, machine-checkable release bundle. We define an artifact suite (TeMLM-Card, TeMLM-Datasheet, TeMLM-Provenance) and a lightweight conformance checklist for repeatable auditing. We instantiate the artifacts on Technetium-I, a large-scale synthetic clinical NLP dataset with 498,000 notes, 7.74M PHI entity annotations across 10 types, and ICD-9-CM diagnosis labels, and report reference results for ProtactiniumBERT (about 100 million parameters) on PHI de-identification (token classification) and top-50 ICD-9 code extraction (multi-label classification). We emphasize that synthetic benchmarks are valuable for tooling and process validation, but models should be validated on real clinical data prior to deployment.
Read the full abstract

All authors Olaf Yunus Laitinen Imanov, Taner Yilmaz, Ayse Tuba Tugrul, Melike Nesrin Zaman, Ozkan Gunalp, Duygu Erisken, Sila Burde Dulger, Rana Irem Turhan, Izzet Ozdemir, Derya Umut Kulali, Ozan Akbulut, Harun Demircioglu, Hasan Basri Kara, Berfin Tavan

Computation and Language

Leveraging Sentence-oriented Augmentation and Transformer-Based Architecture for Vietnamese-Bahnaric Translation

Tan Sang Nguyen, Quoc Nguyen Pham and Tho Quan

The Bahnar people, an ethnic minority in Vietnam with a rich ancestral heritage, possess a language of immense cultural and historical significance. The government places a strong emphasis on preserving and promoting the Bahnaric language by making it accessible online and encouraging communication across generations. Recent advancements in artificial intelligence, such as Neural Machine Translation (NMT), have brought about a transformation in translation by improving accuracy and fluency. This, in turn, contributes to the revival of the language through educational efforts, communication, and documentation. Specifically, NMT is pivotal in enhancing accessibility for Bahnaric speakers, making information and content more readily available. Nevertheless, the translation of Vietnamese into Bahnaric faces practical challenges due to resource constraints, especially given the limited resources available for the Bahnaric language. To address this, we employ state-of-the-art techniques in NMT along with two augmentation strategies for domain-specific Vietnamese-Bahnaric translation task. Importantly, both approaches are flexible and can be used with various neural machine translation models. Additionally, they do not require complex data preprocessing steps, the training of additional systems, or the acquisition of extra data beyond the existing training parallel corpora.
Read the full abstract

Computation and Language

PsyProbe: Proactive and Interpretable Dialogue through User State Modeling for Exploratory Counseling

Sohhyung Park, Hyunji Kang, Sungzoon Cho et al. (+1 more)

Recent advances in large language models have enabled mental health dialogue systems, yet existing approaches remain predominantly reactive, lacking systematic user state modeling for proactive therapeutic exploration. We introduce PsyProbe, a dialogue system designed for the exploration phase of counseling that systematically tracks user psychological states through the PPPPPI framework (Presenting, Predisposing, Precipitating, Perpetuating, Protective, Impact) augmented with cognitive error detection. PsyProbe combines State Builder for extracting structured psychological profiles, Memory Construction for tracking information gaps, Strategy Planner for Motivational Interviewing behavioral codes, and Response Generator with Question Ideation and Critic/Revision modules to generate contextually appropriate, proactive questions. We evaluate PsyProbe with 27 participants in real-world Korean counseling scenarios, including automatic evaluation across ablation modes, user evaluation, and expert evaluation by a certified counselor. The full PsyProbe model consistently outperforms baseline and ablation modes in automatic evaluation. User evaluation demonstrates significantly increased engagement intention and improved naturalness compared to baseline. Expert evaluation shows that PsyProbe substantially improves core issue understanding and achieves question rates comparable to professional counselors, validating the effectiveness of systematic state modeling and proactive questioning for therapeutic exploration.
Read the full abstract

All authors Sohhyung Park, Hyunji Kang, Sungzoon Cho, Dongil Kim

Computation and Language · Cross-listed from cs.AI

Payoff scaling shapes cooperation in LLM agents across languages

Trung-Kiet Huynh, Dao-Sy Duy-Minh, Thanh-Bang Cao et al. (+13 more)

Large language models (LLMs) are increasingly deployed as autonomous agents that negotiate, coordinate, and make decisions on behalf of human users. Understanding how they cooperate is therefore important for AI governance and machine behaviour research. We examine two factors: payoff magnitude and interaction language. Classical game theory predicts that multiplying all payoffs by a positive constant should not change behaviour because the underlying strategic incentives remain unchanged. We test this prediction using five widely adopted commercial LLMs playing an iterated prisoner's dilemma across ten payoff scales and five languages (English, French, Vietnamese, Chinese and Arabic). Contrary to theoretical expectations, changing only the numerical scale of payoffs substantially alters cooperation. Models respond differently, often in opposite directions, meaning that averages across models can conceal important behavioural heterogeneity. Sensitivity to payoff scale also depends strongly on prompt language. Moreover, payoff scale modifies the effects of persona instructions: at sub-unit stakes, where no payoff shown exceeds 1, all five models cooperate more under a selfish than a cooperative persona, whereas Claude and Gemini reverse this pattern at larger scales. These findings show that cooperation measures depend critically on payoff scale and language and should not be treated as fixed model traits.
Read the full abstract

All authors Trung-Kiet Huynh, Dao-Sy Duy-Minh, Thanh-Bang Cao, Phong-Hao Le, Hong-Dan Nguyen, Phu-Quy Nguyen-Lam, Minh-Luan Nguyen-Vo, Hong-Phat Pham, Phu-Hoa Pham, Thien-Kim Than, Chi-Nguyen Tran, Huy Tran, Gia-Thoai Tran-Le, Alessio Buscemi, Le Hong Trang, The Anh Han

Computation and Language

Optimizing Conversational Quality in Spoken Dialogue Systems with Reinforcement Learning from AI Feedback

Siddhant Arora, Jinchuan Tian, Jiatong Shi et al. (+4 more)

Reinforcement learning from human or AI feedback (RLHF/RLAIF) for speech-in/speech-out dialogue systems (SDS) remains underexplored, with prior work largely limited to single semantic rewards applied at the utterance level. Such setups overlook the multi-dimensional and multi-modal nature of conversational quality, which encompasses semantic coherence, audio naturalness, speaker consistency, emotion alignment, and turn-taking behavior. Moreover, they are fundamentally mismatched with duplex spoken dialogue systems that generate responses incrementally, where agents must make decisions based on partial utterances. We address these limitations with the first multi-reward RLAIF framework for SDS, combining semantic, audio-quality, and emotion-consistency rewards. To align utterance-level preferences with incremental, blockwise decoding in duplex models, we apply turn-level preference sampling and aggregate per-block log-probabilities within a single DPO objective. We present the first systematic study of preference learning for improving SDS quality in both multi-turn Chain-of-Thought and blockwise duplex models, and release a multi-reward DPO dataset to support reproducible research. Experiments show that single-reward RLAIF selectively improves its targeted metric, while joint multi-reward training yields consistent gains across semantic quality and audio naturalness. These results highlight the importance of holistic, multi-reward alignment for practical conversational SDS.
Read the full abstract

All authors Siddhant Arora, Jinchuan Tian, Jiatong Shi, Hayato Futami, Yosuke Kashiwagi, Emiru Tsunoo, Shinji Watanabe

Computation and Language · Cross-listed from cs.CY

Who's in Charge? Disempowerment Patterns in Real-World LLM Usage

Mrinank Sharma, Miles McCain, Raymond Douglas et al. (+1 more)

Although AI assistants are now deeply embedded in society, there has been limited empirical study of how their usage affects human empowerment. We present the first large-scale empirical analysis of disempowerment patterns in real-world AI assistant interactions, analyzing 1.5 million consumer Claude$.$ai conversations using a privacy-preserving approach. We focus on situational disempowerment potential, which occurs when AI assistant interactions risk leading users to form distorted perceptions of reality, make inauthentic value judgments, or act in ways misaligned with their values. Quantitatively, we find that severe forms of disempowerment potential occur in fewer than one in a thousand conversations, though rates are substantially higher in personal domains like relationships and lifestyle. Qualitatively, we uncover several concerning patterns, such as validation of persecution narratives and grandiose identities with emphatic sycophantic language, definitive moral judgments about third parties, and complete scripting of value-laden personal communications that users appear to implement verbatim. Analysis of historical trends reveals an increase in the prevalence of disempowerment potential over time. We also find that interactions with greater disempowerment potential receive higher user approval ratings, possibly suggesting a tension between short-term user preferences and long-term human empowerment. Our findings highlight the need for AI systems designed to robustly support human autonomy and flourishing.
Read the full abstract

All authors Mrinank Sharma, Miles McCain, Raymond Douglas, David Duvenaud

Computation and Language · Cross-listed from cs.LG

Principled Fine-tuning of LLMs from User-Edits: A Medley of Preference, Supervision, and Reward

Dipendra Misra, Aldo Pacchiano, Ta-Chung Chi et al. (+1 more)

We study how to fine-tune LLMs using user-edit deployment data consisting of a set of context, an agent's response, and user edits. This deployment data is naturally generated by users in applications such as LLMs-based writing assistants and coding agents. The _natural_ origin of user edits makes it a desired source for adapting and personalizing LLMs. In this setup, there emerges a unification of various feedback types namely preferences, supervised labels, and cost that are typically studied separately in the literature. In this paper, we initiate the theoretical investigation of learning from user edits. We first derive bounds for learning algorithms that learn from each of these feedback types. We prove that these algorithms have different trade-offs depending upon the user, data distribution, and model class. We then propose a simple ensembling procedure to jointly learn from these feedback types. On two domains adapted from Gao et al. 2024, we show our ensembling procedure outperforms these methods that learn from individual feedback. Further, we show that our proposed procedure can robustly adapt to different user-edit distributions at test time.
Read the full abstract

All authors Dipendra Misra, Aldo Pacchiano, Ta-Chung Chi, Ge Gao

Monday 26 January

Computation and Language

NLP Privacy Risk Identification in Social Media (NLP-PRISM): A Survey

Dhiman Goswami, Jai Kruthunz Naveen Kumar and Sanchari Das

Natural Language Processing (NLP) is integral to social media analytics but often processes content containing Personally Identifiable Information (PII), behavioral cues, and metadata raising privacy risks such as surveillance, profiling, and targeted advertising. To systematically assess these risks, we review 203 peer-reviewed papers and propose the NLP Privacy Risk Identification in Social Media (NLP-PRISM) framework, which evaluates vulnerabilities across six dimensions: data collection, preprocessing, visibility, fairness, computational risk, and regulatory compliance. Our analysis shows that transformer models achieve F1-scores ranging from 0.58-0.84, but incur a 1% - 23% drop under privacy-preserving fine-tuning. Using NLP-PRISM, we examine privacy coverage in six NLP tasks: sentiment analysis (16), emotion detection (14), offensive language identification (19), code-mixed processing (39), native language identification (29), and dialect detection (24) revealing substantial gaps in privacy research. We further found a (reduced by 2% - 9%) trade-off in model utility, MIA AUC (membership inference attacks) 0.81, AIA accuracy 0.75 (attribute inference attacks). Finally, we advocate for stronger anonymization, privacy-aware learning, and fairness-driven training to enable ethical NLP in social media contexts.
Read the full abstract

Computation and Language · Cross-listed from cs.HC

AI as Teammate or Tool? A Review of Human-AI Interaction in Decision Support

Most. Sharmin Sultana Samu, Nafisa Khan, Kazi Toufique Elahi et al. (+3 more)

The integration of Artificial Intelligence (AI) necessitates determining whether systems function as tools or collaborative teammates. In this study, by synthesizing Human-AI Interaction (HAI) literature, we analyze this distinction across four dimensions: interaction design, trust calibration, collaborative frameworks and healthcare applications. Our analysis reveals that static interfaces and miscalibrated trust limit AI efficacy. Performance hinges on aligning transparency with cognitive workflows, yet a fluency trap often inflates trust without improving decision-making. Consequently, an overemphasis on explainability leaves systems largely passive. Our findings show that current AI systems remain largely passive due to an overreliance on explainability-centric designs and that transitioning AI to an active teammate requires adaptive, context-aware interactions that support shared mental models and the dynamic negotiation of authority between humans and AI.
Read the full abstract

All authors Most. Sharmin Sultana Samu, Nafisa Khan, Kazi Toufique Elahi, Tasnuva Binte Rahman, Md. Rakibul Islam, Farig Sadeque

Computation and Language

Not the Example, but the Process: How Self-Generated Examples Enhance LLM Reasoning

Daehoon Gwak, Minseo Jung, Junwoo Park et al. (+4 more)

Recent studies have shown that Large Language Models (LLMs) can improve their reasoning performance through self-generated few-shot examples, achieving results comparable to manually curated in-context examples. However, the underlying mechanism behind these gains remains unclear, making it hard to decide when and how to apply the technique effectively. In this work, we argue that the key benefit arises not from the generated examples themselves but from the act of creating them. To validate this, on reasoning-intensive tasks across diverse LLM architectures, we systematically evaluate three prompting strategies for in-context learning: (1) Zero-shot prompting; (2) Integrated prompting, where LLMs create and solve problems within a single, unified prompt; and (3) Decoupled prompting, where self-generated examples are reused as in-context examples, but the context of their creation itself is excluded. We conduct experiments across five widely used model architectures, demonstrating that Integrated prompting consistently outperforms both Zero-shot and Decoupled prompting. In contrast, Decoupled prompting offers only marginal gains over Zero-shot. Further, for a more in-depth analysis, we conduct an attention analysis and observe significant differences in attention patterns between Integrated and Decoupled prompting. These findings suggest that the advantage of self-generation prompting comes from the process of problem creation, not the examples themselves, providing valuable insights for designing more effective prompting strategies.
Read the full abstract

All authors Daehoon Gwak, Minseo Jung, Junwoo Park, Minho Park, ChaeHun Park, Junha Hyung, Jaegul Choo

Computation and Language

Enhancing Action and Ingredient Modeling for Semantically Grounded Recipe Generation

Guoshan Liu, Bin Zhu, Yian Li et al. (+3 more)

Recent advances in Multimodal Large Language Models (MLMMs) have enabled recipe generation from food images, yet outputs often contain semantically incorrect actions or ingredients despite high lexical scores (e.g., BLEU, ROUGE). To address this gap, we propose a semantically grounded framework that predicts and validates actions and ingredients as internal context for instruction generation. Our two-stage pipeline combines supervised fine-tuning (SFT) with reinforcement fine-tuning (RFT): SFT builds foundational accuracy using an Action-Reasoning dataset and ingredient corpus, while RFT employs frequency-aware rewards to improve long-tail action prediction and ingredient generalization. A Semantic Confidence Scoring and Rectification (SCSR) module further filters and corrects predictions. Experiments on Recipe1M show state-of-the-art performance and markedly improved semantic fidelity.
Read the full abstract

All authors Guoshan Liu, Bin Zhu, Yian Li, Jingjing Chen, Chong-Wah Ngo, Yu-Gang Jiang

Computation and Language

CAST: Achieving Stable LLM-based Text Analysis for Data Analytics

Jinxiang Xie, Zihao Li, Wei He et al. (+3 more)

Text analysis of tabular data relies on two core operations: summarization for corpus-level theme extraction and tagging for row-level labeling. A critical limitation of employing large language models (LLMs) for these tasks is their inability to meet the high standards of output stability demanded by data analytics. To address this challenge, we introduce CAST (Consistency via Algorithmic Prompting and Stable Thinking), a framework that enhances output stability by constraining the model's latent reasoning path. CAST combines (i) Algorithmic Prompting to impose a procedural scaffold over valid reasoning transitions and (ii) Thinking-before-Speaking to enforce explicit intermediate commitments before final generation. To measure progress, we introduce CAST-S and CAST-T, stability metrics for bulleted summarization and tagging, and validate their alignment with human judgments. Experiments across publicly available benchmarks on multiple LLM backbones show that CAST consistently achieves the best stability among all baselines, improving Stability Score by up to 16.2%, while maintaining or improving output quality.
Read the full abstract

All authors Jinxiang Xie, Zihao Li, Wei He, Rui Ding, Shi Han, Dongmei Zhang

Computation and Language

Reranker Optimization via Geodesic Distances on k-NN Manifolds

Wen G. Gong

Current neural reranking approaches for retrieval-augmented generation (RAG) rely on cross-encoders or large language models (LLMs), requiring substantial computational resources and exhibiting latencies of 3-5 seconds per query. We propose Maniscope, a geometric reranking method that computes geodesic distances on k-nearest neighbor (k-NN) manifolds constructed over retrieved document candidates. This approach combines global cosine similarity with local manifold geometry to capture semantic structure that flat Euclidean metrics miss. Evaluating on eight BEIR benchmark datasets (1,233 queries), Maniscope outperforms HNSW graph-based baseline on the three hardest datasets (NFCorpus: +7.0%, TREC-COVID: +1.6%, AorB: +2.8% NDCG@3) while being 3.2x faster (4.7 ms vs 14.8 ms average). Compared to cross-encoder rerankers, Maniscope achieves within 2% accuracy at 10-45x lower latency. On TREC-COVID, LLM-Reranker provides only +0.5% NDCG@3 improvement over Maniscope at 840x higher latency, positioning Maniscope as a practical alternative for real-time RAG deployment. The method requires O(N D + M^2 D + M k log k) complexity where M << N , enabling sub-10 ms latency. We plan to release Maniscope as open-source software.
Read the full abstract

Computation and Language

From Transcripts to AI Agents: Knowledge Extraction, RAG Integration, and Robust Evaluation of Conversational AI Assistants

Krittin Pachtrachai, Petmongkon Pornpichitsuwan, Wachiravit Modecrua et al. (+1 more)

Building reliable conversational AI assistants for customer-facing industries remains challenging due to noisy conversational data, fragmented knowledge, and the requirement for accurate human hand-off - particularly in domains that depend heavily on real-time information. This paper presents an end-to-end framework for constructing and evaluating a conversational AI assistant directly from historical call transcripts. Incoming transcripts are first graded using a simplified adaptation of the PIPA framework, focusing on observation alignment and appropriate response behavior, and are filtered to retain only high-quality interactions exhibiting coherent flow and effective human agent responses. Structured knowledge is then extracted from curated transcripts using large language models (LLMs) and deployed as the sole grounding source in a Retrieval-Augmented Generation (RAG) pipeline. Assistant behavior is governed through systematic prompt tuning, progressing from monolithic prompts to lean, modular, and governed designs that ensure consistency, safety, and controllable execution. Evaluation is conducted using a transcript-grounded user simulator, enabling quantitative measurement of call coverage, factual accuracy, and human escalation behavior. Additional red teaming assesses robustness against prompt injection, out-of-scope, and out-of-context attacks. Experiments are conducted in the Real Estate and Specialist Recruitment domains, which are intentionally challenging and currently suboptimal for automation due to their reliance on real-time data. Despite these constraints, the assistant autonomously handles approximately 30 percents of calls, achieves near-perfect factual accuracy and rejection behavior, and demonstrates strong robustness under adversarial testing.
Read the full abstract

All authors Krittin Pachtrachai, Petmongkon Pornpichitsuwan, Wachiravit Modecrua, Touchapon Kraisingkorn

Computation and Language · Cross-listed from cs.AI

A Geometric Taxonomy of Hallucinations in LLMs

Javier Marín

Hallucinations in deployed language models can have real consequences for downstream decisions in domains such as healthcare, legal, and financial services. In production, detection has to run on what the deployed system can see: the query, the response, and often a source document. White-box access to model internals and multi-sample querying are not generally available behind a third-party API. Within this setting - black-box, single-pass, only question/answer available - the dominant baseline is NLI, which returns a value but no diagnosis when it fails. We argue that operating directly on the geometry of the embedding space provides detection methods whose successes and failures are interpretable as structural properties of contrastive sentence-encoder training \citep{wang2020understanding}. The contribution is: given an operationally-motivated taxonomy, geometry predicts which types of hallucination are detectable and which are not - and the predictions hold. We propose three operational types organized by the relation of the response embedding to the plausibility region of grounded responses on the unit hypersphere, and derive from the alignment objective a prediction for each: (1)query-proximate unfaithfulness is detectable by an angular ratio; (2)confabulation outside the plausibility region produces a directional signature that outperforms NLI on expert-annotated error; (3)factual errors sharing vocabulary and frame with correct answers are not separable by angular geometry. To validate on content resembling deployment, we built a 212-pair human-confabulated dataset across nine domains using provoked confabulation.
Read the full abstract

Computation and Language · Cross-listed from cs.LG

Is Finer Better? The Limits of Microscaling Formats in Large Language Models

Andrea Fasoli, Monodeep Kar, Chi-Chun Liu et al. (+4 more)

Microscaling data formats leverage per-block tensor quantization to enable aggressive model compression with limited loss in accuracy. Unlocking their potential for efficient training and inference necessitates hardware-friendly implementations that handle matrix multiplications in a native format and adopt efficient error-mitigation strategies. Herein, we report the emergence of a surprising behavior associated with microscaling quantization, whereas the output of a quantized model degrades as block size is decreased below a given threshold. This behavior clashes with the expectation that a smaller block size should allow for a better representation of the tensor elements. We investigate this phenomenon both experimentally and theoretically, decoupling the sources of quantization error behind it. Experimentally, we analyze the distributions of several Large Language Models and identify the conditions driving the anomalous behavior. Theoretically, we lay down a framework showing remarkable agreement with experimental data from pretrained model distributions and ideal ones. Overall, we show that the anomaly is driven by the interplay between narrow tensor distributions and the limited dynamic range of the quantized scales. Based on these insights, we propose the use of FP8 unsigned E5M3 (UE5M3) as a novel hardware-friendly format for the scales in FP4 microscaling data types. We demonstrate that UE5M3 achieves comparable performance to the conventional FP8 unsigned E4M3 scales while obviating the need of global scaling operations on weights and activations.
Read the full abstract

All authors Andrea Fasoli, Monodeep Kar, Chi-Chun Liu, Swagath Venkataramani, Viji Srinivasan, Leland Chang, Naigang Wang

Computation and Language

FROST: Filtering Reasoning Outliers with Attention for Efficient Reasoning

Haozheng Luo, Zhuolin Jiang, Md Zahid Hasan et al. (+2 more)

We propose FROST, an attention-aware method for efficient reasoning. Unlike traditional approaches, FROST leverages attention weights to prune uncritical reasoning paths, yielding shorter and more reliable reasoning trajectories. Methodologically, we introduce the concept of reasoning outliers and design an attention-based mechanism to remove them. Theoretically, FROST preserves and enhances the model's reasoning capacity while eliminating outliers at the sentence level. Empirically, we validate FROST on four benchmarks using two strong reasoning models (Phi-4-Reasoning and GPT-OSS-20B), outperforming state-of-the-art methods such as TALE and ThinkLess. Notably, FROST achieves an average 69.68% reduction in token usage and a 26.70% improvement in accuracy over the base model. Furthermore, in evaluations of attention outlier metrics, FROST reduces the maximum infinity norm by 15.97% and the average kurtosis by 91.09% compared to the base model. Code is available at https://github.com/robinzixuan/FROST
Read the full abstract

All authors Haozheng Luo, Zhuolin Jiang, Md Zahid Hasan, Yan Chen, Soumalya Sarkar