Peer Revenue Impact factor: volatile

Papers

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

Computer Science Computation and Language

cs.CL Watch Computation and Language

Saturday 8 November

Computation and Language

MIND Your Reasoning: A Meta-Cognitive Intuitive-Reflective Network for Dual-Reasoning in Multimodal Stance Detection

Bingbing Wang, Zhengda Jin, Bin Liang et al. (+4 more)

Multimodal Stance Detection (MSD) is a crucial task for understanding public opinion on social media. Existing methods predominantly operate by learning to fuse modalities. They lack an explicit reasoning process to discern how inter-modal dynamics, such as irony or conflict, collectively shape the user's final stance, leading to frequent misjudgments. To address this, we advocate for a paradigm shift from *learning to fuse* to *learning to reason*. We introduce **MIND**, a **M**eta-cognitive **I**ntuitive-reflective **N**etwork for **D**ual-reasoning. Inspired by the dual-process theory of human cognition, MIND operationalizes a self-improving loop. It first generates a rapid, intuitive hypothesis by querying evolving Modality and Semantic Experience Pools. Subsequently, a meta-cognitive reflective stage uses Modality-CoT and Semantic-CoT to scrutinize this initial judgment, distill superior adaptive strategies, and evolve the experience pools themselves. These dual experience structures are continuously refined during training and recalled at inference to guide robust and context-aware stance decisions. Extensive experiments on the MMSD benchmark demonstrate that our MIND significantly outperforms most baseline models and exhibits strong generalization.
Read the full abstract

All authors Bingbing Wang, Zhengda Jin, Bin Liang, Wenjie Li, Jing Li, Ruifeng Xu, Min Zhang

Computation and Language

Efficient Hate Speech Detection: A Three-Layer LoRA-Tuned BERTweet Framework

Mahmoud El-Bahnasawi

This paper addresses the critical challenge of developing computationally efficient hate speech detection systems that maintain competitive performance while being practical for real-time deployment. We propose a novel three-layer framework that combines rule-based pre-filtering with a parameter-efficient LoRA-tuned BERTweet model and continuous learning capabilities. Our approach achieves 0.85 macro F1 score - representing 94% of the performance of state-of-the-art large language models like SafePhi (Phi-4 based) while using a base model that is 100x smaller (134M vs 14B parameters). Compared to traditional BERT-based approaches with similar computational requirements, our method demonstrates superior performance through strategic dataset unification and optimized fine-tuning. The system requires only 1.87M trainable parameters (1.37% of full fine-tuning) and trains in approximately 2 hours on a single T4 GPU, making robust hate speech detection accessible in resource-constrained environments while maintaining competitive accuracy for real-world deployment.
Read the full abstract

Computation and Language

Visual Exploration of Feature Relationships in Sparse Autoencoders with Curated Concepts

Xinyuan Yan, Shusen Liu, Kowshik Thopalli et al. (+1 more)

Sparse autoencoders (SAEs) have emerged as a powerful tool for uncovering interpretable features in large language models (LLMs) through the sparse directions they learn. However, the sheer number of extracted directions makes comprehensive exploration intractable. While conventional embedding techniques such as UMAP can reveal global structure, they suffer from limitations including high-dimensional compression artifacts, overplotting, and misleading neighborhood distortions. In this work, we propose a focused exploration framework that prioritizes curated concepts and their corresponding SAE features over attempts to visualize all available features simultaneously. We present an interactive visualization system that combines topology-based visual encoding with dimensionality reduction to faithfully represent both local and global relationships among selected features. This hybrid approach enables users to investigate SAE behavior through targeted, interpretable subsets, facilitating deeper and more nuanced analysis of concept representation in latent space.
Read the full abstract

All authors Xinyuan Yan, Shusen Liu, Kowshik Thopalli, Bei Wang

Computation and Language

Multi-Reward GRPO Fine-Tuning for De-biasing Large Language Models: A Study Based on Chinese-Context Discrimination Data

Deng Yixuan and Ji Xiaoqiang

Large Language Models (LLMs) often exhibit implicit biases and discriminatory tendencies that reflect underlying social stereotypes. While recent alignment techniques such as RLHF and DPO have mitigated some of these issues, they remain limited in addressing culturally specific and multi-dimensional forms of discrimination. This paper proposes a Multi-Reward Group Relative Policy Optimization (GRPO) framework to fine-tune LLMs toward ethical and bias-free behavior. Our approach constructs a synthetic English-language dataset derived from Chinese-context discrimination categories, including regional, ethnic, and occupational biases. Each instance is paired with both neutral and biased responses to train a reward model based on DeBERTa-v3, which provides multi-dimensional reward signals capturing fairness, neutrality, and linguistic quality. The trained reward model then guides GRPO fine-tuning to optimize model outputs along these ethical dimensions. Experimental results demonstrate significant reductions in bias intensity and improved alignment with non-discriminatory standards without compromising fluency or informativeness. This study highlights the effectiveness of GRPO-based multi-reward optimization for de-biasing LLMs and offers a replicable framework for cultural-contextual ethical alignment.
Read the full abstract

Computation and Language

LLMs Do Not See Age: Assessing Demographic Bias in Automated Systematic Review Synthesis

Favour Yahdii Aghaebe, Tanefa Apekey, Elizabeth Williams et al. (+1 more)

Clinical interventions often hinge on age: medications and procedures safe for adults may be harmful to children or ineffective for older adults. However, as language models are increasingly integrated into biomedical evidence synthesis workflows, it remains uncertain whether these systems preserve such crucial demographic distinctions. To address this gap, we evaluate how well state-of-the-art language models retain age-related information when generating abstractive summaries of biomedical studies. We construct DemogSummary, a novel age-stratified dataset of systematic review primary studies, covering child, adult, and older adult populations. We evaluate three prominent summarisation-capable LLMs, Qwen (open-source), Longformer (open-source) and GPT-4.1 Nano (proprietary), using both standard metrics and a newly proposed Demographic Salience Score (DSS), which quantifies age-related entity retention and hallucination. Our results reveal systematic disparities across models and age groups: demographic fidelity is lowest for adult-focused summaries, and under-represented populations are more prone to hallucinations. These findings highlight the limitations of current LLMs in faithful and bias-free summarisation and point to the need for fairness-aware evaluation frameworks and summarisation pipelines in biomedical NLP.
Read the full abstract

All authors Favour Yahdii Aghaebe, Tanefa Apekey, Elizabeth Williams, Nafise Sadat Moosavi

Computation and Language

Revisiting Entropy in Reinforcement Learning for Large Reasoning Models

Renren Jin, Pengzhi Gao, Yuqi Ren et al. (+6 more)

Reinforcement learning with verifiable rewards (RLVR) has emerged as a prominent paradigm for enhancing the reasoning capabilities of large language models (LLMs). However, the entropy of LLMs usually collapses during RLVR training, leading to premature convergence to suboptimal local minima and hindering further performance improvement. Although various approaches have been proposed to mitigate entropy collapse, a comprehensive study of entropy in RLVR remains lacking. To bridge this gap, we conduct extensive experiments to investigate the entropy dynamics of LLMs trained with RLVR and analyze how model entropy correlates with response diversity, calibration, and performance across various benchmarks. Our results identify three key factors that influence entropy: the clipping thresholds in the optimization objective, the number of off-policy updates, and the diversity of the training data. Furthermore, through both theoretical analysis and empirical validation, we demonstrate that tokens with positive advantages are the primary drivers of entropy collapse. Motivated by this insight, we propose Positive-Advantage Reweighting, a simple yet effective approach that regulates model entropy by adjusting the loss weights assigned to tokens with positive advantages during RLVR training, while maintaining competitive performance.
Read the full abstract

All authors Renren Jin, Pengzhi Gao, Yuqi Ren, Zhuowen Han, Tongxuan Zhang, Wuwei Huang, Wei Liu, Jian Luan, Deyi Xiong

Computation and Language

Interpretable Recognition of Cognitive Distortions in Natural Language Texts

Anton Kolonin and Anna Arinicheva

We propose a new approach to multi-factor classification of natural language texts based on weighted structured patterns such as N-grams, taking into account the heterarchical relationships between them, applied to solve such a socially impactful problem as the automation of detection of specific cognitive distortions in psychological care, relying on an interpretable, robust and transparent artificial intelligence model. The proposed recognition and learning algorithms improve the current state of the art in this field. The improvement is tested on two publicly available datasets, with significant improvements over literature-known F1 scores for the task, with optimal hyper-parameters determined, having code and models available for future use by the community.
Read the full abstract

Computation and Language

Reinforcement Learning Improves Traversal of Parametric Knowledge in LLMs

Renfei Zhang, Manasa Kaniselvan, Rylan Schaeffer et al. (+1 more)

Reinforcement learning (RL) is often credited with improving reasoning at the expense of factual knowledge. We instead find that reasoning models outperform their instruction-tuned versions on factual recall by accessing existing parametric knowledge more effectively. Across five model families, structured prompting, which explicitly guides models through hierarchical traversal, recovers most of this gap, suggesting that much of the missing knowledge is latent rather than absent. Controlled RL experiments further support this: training on unseen, non-extractable facts improves recall of held-out, frequent but previously inaccessible facts, ruling out simple data exposure. Decomposing the training objective further attributes this gain to iterated on-policy exploration. The same mechanism appears behaviorally and internally: the reasoning advantage grows with retrieval depth, while layerwise analysis finds similar factual representations but divergent query representations. Distilled models, in contrast, often imitate self-correction without acquiring the exploration needed for navigation. Together, these findings suggest that improving factual recall in LLMs depends not only on expanding what models know but also on teaching them to navigate it, motivating future post-training methods that optimize traversal.
Read the full abstract

All authors Renfei Zhang, Manasa Kaniselvan, Rylan Schaeffer, Niloofar Mireshghallah

Computation and Language

IDALC: A Semi-Supervised Framework for Intent Detection and Active Learning based Correction

Ankan Mullick, Sukannya Purkayastha, Saransh Sharma et al. (+2 more)

Voice-controlled dialog systems have become immensely popular due to their ability to perform a wide range of actions in response to diverse user queries. These agents possess a predefined set of skills or intents to fulfill specific user tasks. But every system has its own limitations. There are instances where, even for known intents, if any model exhibits low confidence, it results in rejection of utterances that necessitate manual annotation. Additionally, as time progresses, there may be a need to retrain these agents with new intents from the system-rejected queries to carry out additional tasks. Labeling all these emerging intents and rejected utterances over time is impractical, thus calling for an efficient mechanism to reduce annotation costs. In this paper, we introduce IDALC (Intent Detection and Active Learning based Correction), a semi-supervised framework designed to detect user intents and rectify system-rejected utterances while minimizing the need for human annotation. Empirical findings on various benchmark datasets demonstrate that our system surpasses baseline methods, achieving a 5-10% higher accuracy and a 4-8% improvement in macro-F1. Remarkably, we maintain the overall annotation cost at just 6-10% of the unlabelled data available to the system. The overall framework of IDALC is shown in Fig. 1
Read the full abstract

All authors Ankan Mullick, Sukannya Purkayastha, Saransh Sharma, Pawan Goyal, Niloy Ganguly

Computation and Language · Cross-listed from cs.CR

Injecting Falsehoods: Adversarial Man-in-the-Middle Attacks Undermining Factual Recall in LLMs

Alina Fastowski, Bardh Prenkaj, Yuxiao Li et al. (+1 more)

LLMs are now an integral part of information retrieval. As such, their role as question answering chatbots raises significant concerns due to their shown vulnerability to adversarial man-in-the-middle (MitM) attacks. Here, we propose the first principled attack evaluation on LLM factual memory under prompt injection via Xmera, our novel, theory-grounded MitM framework. By perturbing the input given to "victim" LLMs in three closed-book and fact-based QA settings, we undermine the correctness of the responses and assess the uncertainty of their generation process. Surprisingly, trivial instruction-based attacks report the highest success rate (up to ~85.3%) while simultaneously having a high uncertainty for incorrectly answered questions. To provide a simple defense mechanism against Xmera, we train Random Forest classifiers on the response uncertainty levels to distinguish between attacked and unattacked queries (average AUC of up to ~94.8%). We believe that signaling users to be cautious about the answers they receive from black-box and potentially corrupt LLMs is a first checkpoint toward user cyberspace safety.
Read the full abstract

All authors Alina Fastowski, Bardh Prenkaj, Yuxiao Li, Gjergji Kasneci

Computation and Language

NILC: Discovering New Intents with LLM-assisted Clustering

Hongtao Wang, Renchi Yang and Wenqing Lin

New intent discovery (NID) seeks to recognize both new and known intents from unlabeled user utterances, which finds prevalent use in practical dialogue systems. Existing works towards NID mainly adopt a cascaded architecture, wherein the first stage focuses on encoding the utterances into informative text embeddings beforehand, while the latter is to group similar embeddings into clusters (i.e., intents), typically by K-Means. However, such a cascaded pipeline fails to leverage the feedback from both steps for mutual refinement, and, meanwhile, the embedding-only clustering overlooks nuanced textual semantics, leading to suboptimal performance. To bridge this gap, this paper proposes NILC, a novel clustering framework specially catered for effective NID. Particularly, NILC follows an iterative workflow, in which clustering assignments are judiciously updated by carefully refining cluster centroids and text embeddings of uncertain utterances with the aid of large language models (LLMs). Specifically, NILC first taps into LLMs to create additional semantic centroids for clusters, thereby enriching the contextual semantics of the Euclidean centroids of embeddings. Moreover, LLMs are then harnessed to augment hard samples (ambiguous or terse utterances) identified from clusters via rewriting for subsequent cluster correction. Further, we inject supervision signals through non-trivial techniques seeding and soft must links for more accurate NID in the semi-supervised setting. Extensive experiments comparing NILC against multiple recent baselines under both unsupervised and semi-supervised settings showcase that NILC can achieve significant performance improvements over six benchmark datasets of diverse domains consistently.
Read the full abstract

Computation and Language · Cross-listed from cs.CY

The Imperfect Learner: Incorporating Developmental Trajectories in Memory-based Student Simulation

Zhengyuan Liu, Stella Xin Yin, Bryan Chen Zhengyu Tan et al. (+5 more)

User simulation is important for developing and evaluating human-centered AI, yet current student simulation in educational applications has significant limitations. Existing approaches focus on single learning experiences and do not account for students' gradual knowledge construction and evolving skill sets. Moreover, large language models are optimized to produce direct and accurate responses, making it challenging to represent the incomplete understanding and developmental constraints that characterize real learners. In this paper, we introduce a novel framework for memory-based student simulation that incorporates developmental trajectories through a hierarchical memory mechanism with structured knowledge representation. The framework also integrates metacognitive processes and personality traits to enrich the individual learner profiling, through dynamical consolidation of both cognitive development and personal learning characteristics. In practice, we implement a curriculum-aligned simulator grounded on the Next Generation Science Standards. Experimental results show that our approach can effectively reflect the gradual nature of knowledge development and the characteristic difficulties students face, providing a more accurate representation of learning processes.
Read the full abstract

All authors Zhengyuan Liu, Stella Xin Yin, Bryan Chen Zhengyu Tan, Roy Ka-Wei Lee, Guimei Liu, Dion Hoe-Lian Goh, Wenya Wang, Nancy F. Chen

Computation and Language

Retrieval-Augmented Generation in Medicine: A Scoping Review of Technical Implementations, Clinical Applications, and Ethical Considerations

Rui Yang, Matthew Yu Heng Wong, Huitao Li et al. (+13 more)

The rapid growth of medical knowledge and increasing complexity of clinical practice pose challenges. In this context, large language models (LLMs) have demonstrated value; however, inherent limitations remain. Retrieval-augmented generation (RAG) technologies show potential to enhance their clinical applicability. This study reviewed RAG applications in medicine. We found that research primarily relied on publicly available data, with limited application in private data. For retrieval, approaches commonly relied on English-centric embedding models, while LLMs were mostly generic, with limited use of medical-specific LLMs. For evaluation, automated metrics evaluated generation quality and task performance, whereas human evaluation focused on accuracy, completeness, relevance, and fluency, with insufficient attention to bias and safety. RAG applications were concentrated on question answering, report generation, text summarization, and information extraction. Overall, medical RAG remains at an early stage, requiring advances in clinical validation, cross-linguistic adaptation, and support for low-resource settings to enable trustworthy and responsible global use.
Read the full abstract

All authors Rui Yang, Matthew Yu Heng Wong, Huitao Li, Xin Li, Wentao Zhu, Jingchi Liao, Kunyu Yu, Jonathan Chong Kai Liew, Weihao Xuan, Yingjian Chen, Yuhe Ke, Jasmine Chiat Ling Ong, Douglas Teodoro, Chuan Hong, Daniel Shi Wei Ting, Nan Liu

Computation and Language · Cross-listed from cs.CR

Can LLM Infer Risk Information From MCP Server System Logs?

Jiayi Fu, Yuansen Zhang and Yinggui Wang

Large Language Models (LLMs) demonstrate strong capabilities in solving complex tasks when integrated with external tools. The Model Context Protocol (MCP) has become a standard interface for enabling such tool-based interactions. However, these interactions introduce substantial security concerns, particularly when the MCP server is compromised or untrustworthy. While prior benchmarks primarily focus on prompt injection attacks or analyze the vulnerabilities of LLM-MCP interaction trajectories, limited attention has been given to the underlying system logs associated with malicious MCP servers. To address this gap, we present the first synthetic benchmark for evaluating LLMs' ability to identify security risks from system logs. We define nine categories of MCP server risks and generate 1,800 synthetic system logs using ten state-of-the-art LLMs. These logs are embedded in the return values of 243 curated MCP servers, yielding a dataset of 2,421 chat histories for training and 471 queries for evaluation. Our pilot experiments reveal that smaller models often fail to detect risky system logs, leading to high false negatives. While models trained with supervised fine-tuning (SFT) tend to over-flag benign logs, resulting in elevated false positives, Reinforcement Learning with Verifiable Reward (RLVR) offers a better precision-recall balance. In particular, after training with Group Relative Policy Optimization (GRPO), Llama3.1-8B-Instruct achieves 83 percent accuracy, surpassing the best-performing large remote model by 9 percentage points. Fine-grained, per-category analysis further underscores the effectiveness of reinforcement learning in enhancing LLM safety within the MCP framework. Code and data are available at https://github.com/PorUna-byte/MCP-RiskCue.
Read the full abstract

Computation and Language

Can Fine-Tuning Erase Your Edits? On the Fragile Coexistence of Knowledge Editing and Adaptation

Yinjie Cheng, Paul Youssef, Christin Seifert et al. (+2 more)

Knowledge editing (KE) offers a lightweight alternative to retraining for updating large language models (LLMs). Meanwhile, fine-tuning remains the default operation for adapting LLMs to new domains and tasks. Despite their widespread adoption, these two post-training interventions have been studied in isolation, leaving open a crucial question: if we fine-tune an edited model, do the edits survive? This question is motivated by practical objectives: removing covert or malicious edits, and preserving beneficial edits. If fine-tuning impairs edits (Fig.1), current KE methods become less efficient, as a newly fine-tuned model requires re-editing; if edits persist, fine-tuned models risk propagating hidden malicious edits, raising serious safety concerns. To this end, we systematically quantify edit decay after fine-tuning across 254 experimental configurations. Our results show that in general, edits decay substantially after subsequent fine-tuning. AlphaEdit exhibits the greatest decay on the zsRE benchmark when applied to GPT-J, where 25.27% of previously successful edits become unsuccessful after fine-tuning. We further find that fine-tuning only the edited layers is sufficient to effectively remove edits, while incurring only modest degradation in downstream performance. Surprisingly, fine-tuning non-edited layers leads to greater edit decay than all-layer fine-tuning. Besides, our activation space analysis reveals that fine-tuning produces a larger and more coherent representational shift, both in magnitude and direction, than KE. Overall, our study underscores the necessity of evaluating KE within the broader LLM application pipeline.
Read the full abstract

All authors Yinjie Cheng, Paul Youssef, Christin Seifert, Jörg Schlötterer, Zhixue Zhao

Computation and Language · Cross-listed from cs.AI

DiagnoLLM: A Hybrid Bayesian Neural Language Framework for Interpretable Disease Diagnosis

Bowen Xu, Xinyue Zeng, Jiazhen Hu et al. (+2 more)

Building trustworthy clinical AI systems requires not only accurate predictions but also transparent, biologically grounded explanations. We present DiagnoLLM, a hybrid framework that integrates Bayesian deconvolution, eQTL-guided deep learning, and LLM-based narrative generation for interpretable disease diagnosis. DiagnoLLM begins with GP-unmix, a Gaussian Process-based hierarchical model that infers cell-type-specific gene expression profiles from bulk and single-cell RNA-seq data while modeling biological uncertainty. These features, combined with regulatory priors from eQTL analysis, power a neural classifier that achieves high predictive performance in Alzheimer's Disease (AD) detection (88.0% accuracy). To support human understanding and trust, we introduce an LLM-based reasoning module that translates model outputs into audience-specific diagnostic reports, grounded in clinical features, attribution signals, and domain knowledge. Human evaluations confirm that these reports are accurate, actionable, and appropriately tailored for both physicians and patients. Our findings show that LLMs, when deployed as post-hoc reasoners rather than end-to-end predictors, can serve as effective communicators within hybrid diagnostic pipelines.
Read the full abstract

All authors Bowen Xu, Xinyue Zeng, Jiazhen Hu, Tuo Wang, Adithya Kulkarni

Computation and Language

DRAGON: Guard LLM Unlearning in Context via Negative Detection and Reasoning

Yaxuan Wang, Chris Yuhao Liu, Quan Liu et al. (+4 more)

Unlearning in Large Language Models (LLMs) is crucial for protecting private data and removing harmful knowledge. Most existing approaches rely on fine-tuning to balance unlearning efficiency with general language capabilities. However, these methods typically require training or access to retain data, which is often unavailable in real world scenarios. Although these methods can perform well when both forget and retain data are available, few works have demonstrated equivalent capability in more practical, data-limited scenarios. To overcome these limitations, we propose Detect-Reasoning Augmented GeneratiON (DRAGON), a systematic, reasoning-based framework that utilizes in-context chain-of-thought (CoT) instructions to guard deployed LLMs before inference. Instead of modifying the base model, DRAGON leverages the inherent instruction-following ability of LLMs and introduces a lightweight detection module to identify forget-worthy prompts without any retain data. These are then routed through a dedicated CoT guard model to enforce safe and accurate in-context intervention. To robustly evaluate unlearning performance, we introduce novel metrics for unlearning performance and the continual unlearning setting. Extensive experiments across three representative unlearning tasks validate the effectiveness of DRAGON, demonstrating its strong unlearning capability, scalability, and applicability in practical scenarios.
Read the full abstract

All authors Yaxuan Wang, Chris Yuhao Liu, Quan Liu, Jinglong Pang, Wei Wei, Yujia Bao, Yang Liu

Friday 7 November

Computation and Language · Cross-listed from cs.DC

Characterizing and Understanding Energy Footprint and Efficiency of Small Language Model on Edges

Md Romyull Islam, Bobin Deng, Nobel Dhar et al. (+4 more)

Cloud-based large language models (LLMs) and their variants have significantly influenced real-world applications. Deploying smaller models (i.e., small language models (SLMs)) on edge devices offers additional advantages, such as reduced latency and independence from network connectivity. However, edge devices' limited computing resources and constrained energy budgets challenge efficient deployment. This study evaluates the power efficiency of five representative SLMs - Llama 3.2, Phi-3 Mini, TinyLlama, and Gemma 2 on Raspberry Pi 5, Jetson Nano, and Jetson Orin Nano (CPU and GPU configurations). Results show that Jetson Orin Nano with GPU acceleration achieves the highest energy-to-performance ratio, significantly outperforming CPU-based setups. Llama 3.2 provides the best balance of accuracy and power efficiency, while TinyLlama is well-suited for low-power environments at the cost of reduced accuracy. In contrast, Phi-3 Mini consumes the most energy despite its high accuracy. In addition, GPU acceleration, memory bandwidth, and model architecture are key in optimizing inference energy efficiency. Our empirical analysis offers practical insights for AI, smart systems, and mobile ad-hoc platforms to leverage tradeoffs from accuracy, inference latency, and power efficiency in energy-constrained environments.
Read the full abstract

All authors Md Romyull Islam, Bobin Deng, Nobel Dhar, Tu N. Nguyen, Selena He, Yong Shi, Kun Suo

Computation and Language

Evaluating LLM Understanding via Structured Tabular Decision Simulations

Sichao Li, Xinyue Xu and Xiaomeng Li

Large language models (LLMs) often achieve impressive predictive accuracy, yet correctness alone does not imply genuine understanding. True LLM understanding, analogous to human expertise, requires making consistent, well-founded decisions across multiple instances and diverse domains, relying on relevant and domain-grounded decision factors. We introduce Structured Tabular Decision Simulations (STaDS), a suite of expert-like decision settings that evaluate LLMs as if they were professionals undertaking structured decision “exams”. In this context, understanding is defined as the ability to identify and rely on the correct decision factors, features that determine outcomes within a domain. STaDS jointly assesses understanding through: (i) question and instruction comprehension, (ii) knowledge-based prediction, and (iii) reliance on relevant decision factors. By analyzing 9 frontier LLMs across 15 diverse decision settings, we find that (a) most models struggle to achieve consistently strong accuracy across diverse domains; (b) models can be accurate yet globally unfaithful, and there are frequent mismatches between stated rationales and factors driving predictions. Our findings highlight the need for global-level understanding evaluation protocols and advocate for novel frameworks that go beyond accuracy to enhance LLMs' understanding ability.
Read the full abstract

Computation and Language

Learn More, Forget Less: A Gradient-Aware Data Selection Approach for LLM

Yibai Liu, Shihang Wang, Zeming Liu et al. (+5 more)

Despite large language models (LLMs) have achieved impressive achievements across numerous tasks, supervised fine-tuning (SFT) remains essential for adapting these models to specialized domains. However, SFT for domain specialization can be resource-intensive and sometimes leads to a deterioration in performance over general capabilities due to catastrophic forgetting (CF). To address these issues, we propose a self-adaptive gradient-aware data selection approach (GrADS) for supervised fine-tuning of LLMs, which identifies effective subsets of training data by analyzing gradients obtained from a preliminary training phase. Specifically, we design self-guided criteria that leverage the magnitude and statistical distribution of gradients to prioritize examples that contribute the most to the model's learning process. This approach enables the acquisition of representative samples that enhance LLMs understanding of domain-specific tasks. Through extensive experimentation with various LLMs across diverse domains such as medicine, law, and finance, GrADS has demonstrated significant efficiency and cost-effectiveness. Remarkably, utilizing merely 5% of the selected GrADS data, LLMs already surpass the performance of those fine-tuned on the entire dataset, and increasing to 50% of the data results in significant improvements! With catastrophic forgetting substantially mitigated simultaneously. We will release our code for GrADS later.
Read the full abstract

All authors Yibai Liu, Shihang Wang, Zeming Liu, Zheming Song, Junzhe Wang, Jingjing Liu, Qingjie Liu, Yunhong Wang

Computation and Language

Motif 2 12.7B technical report

Junghwan Lim, Sungmin Lee, Dongseok Kim et al. (+22 more)

We introduce Motif-2-12.7B, a new open-weight foundation model that pushes the efficiency frontier of large language models by combining architectural innovation with system-level optimization. Designed for scalable language understanding and robust instruction generalization under constrained compute budgets, Motif-2-12.7B builds upon Motif-2.6B with the integration of Grouped Differential Attention (GDA), which improves representational efficiency by disentangling signal and noise-control attention pathways. The model is pre-trained on 5.5 trillion tokens spanning diverse linguistic, mathematical, scientific, and programming domains using a curriculum-driven data scheduler that gradually changes the data composition ratio. The training system leverages the MuonClip optimizer alongside custom high-performance kernels, including fused PolyNorm activations and the Parallel Muon algorithm, yielding significant throughput and memory efficiency gains in large-scale distributed environments. Post-training employs a three-stage supervised fine-tuning pipeline that successively enhances general instruction adherence, compositional understanding, and linguistic precision. Motif-2-12.7B demonstrates competitive performance across diverse benchmarks, showing that thoughtful architectural scaling and optimized training design can rival the capabilities of much larger models.
Read the full abstract

All authors Junghwan Lim, Sungmin Lee, Dongseok Kim, Taehyun Kim, Eunhwan Park, Jeesoo Lee, Jeongdoo Lee, Junhyeok Lee, Wai Ting Cheung, Dahye Choi, Jaeheui Her, Jaeyeon Huh, Hanbin Jung, Changjin Kang, Beomgyu Kim, Minjae Kim, Taewhan Kim, Youngrok Kim, Hyukjin Kweon, Haesol Lee, Kungyu Lee, Dongpin Oh, Yeongjae Park, Bokki Ryu, Dongjoo Weon

Computation and Language

It Takes Two: A Dual Stage Approach for Terminology-Aware Translation

Akshat Singh Jaswal

This paper introduces DuTerm, a novel two-stage architecture for terminology-constrained machine translation. Our system combines a terminology-aware NMT model, adapted via fine-tuning on large-scale synthetic data, with a prompt-based LLM for post-editing. The LLM stage refines NMT output and enforces terminology adherence. We evaluate DuTerm on English-to German, English-to-Spanish, and English-to-Russian with the WMT 2025 Terminology Shared Task corpus. We demonstrate that flexible, context-driven terminology handling by the LLM consistently yields higher quality translations than strict constraint enforcement. Our results highlight a critical trade-off, revealing that an LLM's work best for high-quality translation as context-driven mutators rather than generators.
Read the full abstract

Computation and Language · Cross-listed from cs.AI

Anchors in the Machine: Behavioral and Attributional Evidence of Anchoring Bias in LLMs

Felipe Valencia-Clavijo

Large language models (LLMs) are increasingly studied as behavioral subjects (e.g., synthetic participants or users) and decision systems (e.g., LLM-as-a-judge), yet it remains unclear how observed cognitive biases relate to changes in model output probabilities. Anchoring bias, a classic human judgment bias, provides a useful test case. Prior research largely relies on generated responses, leaving distributional effects and input-field contributions less explored. This paper advances the study of anchoring in LLMs through three contributions: (1) a log-probability-based behavioral analysis examining how numerical anchors shift distributions over candidate outputs, using alternative questions and anchor pairs alongside a positive control to reduce reliance on potentially memorized stimuli; (2) exact Banzhaf-value attribution over structured prompt fields to quantify anchor contributions to log-probabilities; and (3) an Anchoring Bias Sensitivity Score (ABSS) integrating behavioral and attributional evidence across six open-source models. Results reveal predominantly positive behavioral anchoring effects, varying substantially across models. Gemma-2B, Phi-2, and Llama-2-7B achieve the highest ABSS rankings, whereas other models exhibit weaker behavioral-attributional alignment despite sometimes showing substantial behavioral shifts. Attributional effects also vary across prompts, indicating that behavioral anchoring does not necessarily imply stronger anchor-field contributions. These findings raise questions about LLM reliability as synthetic participants and potential vulnerabilities to numerical cues in automated evaluation and decision-making. More broadly, the framework connects behavioral science with explainable AI, particularly interpretability, and AI safety, providing a reproducible approach to investigating cognitive-bias-like effects in LLMs.
Read the full abstract

Computation and Language

Language Generation: Complexity Barriers and Implications for Learning

Marcelo Arenas, Pablo Barceló, Luis Cofré et al. (+1 more)

Kleinberg and Mullainathan showed that language generation in the limit is always possible at the level of computability: given enough positive examples, a learner can eventually generate data indistinguishable from a target language. However, such existence results do not address feasibility. We study the sample complexity of language generation in the limit for several canonical classes of formal languages. Our results show that infeasibility already appears for context-free and regular languages, and persists even for strict subclasses such as locally threshold testable languages, as well as for incomparable classes such as non-erasing pattern languages, a well-studied class in the theory of language identification. Overall, our results establish a clear gap between the theoretical possibility of language generation in the limit and its computational feasibility.
Read the full abstract

All authors Marcelo Arenas, Pablo Barceló, Luis Cofré, Alexander Kozachinskiy

Computation and Language

Multi-Scale Feature Fusion and Graph Neural Network Integration for Text Classification with Large Language Models

Xiangchen Song, Yulin Huang, Jinxu Guo et al. (+2 more)

This study investigates a hybrid method for text classification that integrates deep feature extraction from large language models, multi-scale fusion through feature pyramids, and structured modeling with graph neural networks to enhance performance in complex semantic contexts. First, the large language model captures contextual dependencies and deep semantic representations of the input text, providing a rich feature foundation for subsequent modeling. Then, based on multi-level feature representations, the feature pyramid mechanism effectively integrates semantic features of different scales, balancing global information and local details to construct hierarchical semantic expressions. Furthermore, the fused features are transformed into graph representations, and graph neural networks are employed to capture latent semantic relations and logical dependencies in the text, enabling comprehensive modeling of complex interactions among semantic units. On this basis, the readout and classification modules generate the final category predictions. The proposed method demonstrates significant advantages in robustness alignment experiments, outperforming existing models on ACC, F1-Score, AUC, and Precision, which verifies the effectiveness and stability of the framework. This study not only constructs an integrated framework that balances global and local information as well as semantics and structure, but also provides a new perspective for multi-scale feature fusion and structured semantic modeling in text classification tasks.
Read the full abstract

All authors Xiangchen Song, Yulin Huang, Jinxu Guo, Yuchen Liu, Yaxuan Luan

Computation and Language

In-Context Learning Beyond Copying: A Training-Time Analysis of Abstractive ICL

Kerem Sahin, Sheridan Feucht, Adam Belfki et al. (+4 more)

The sudden emergence of induction circuits and the ability to perform in-context copying in LLMs coincides with a sharp drop in training loss over a short span of the training process. This co-occurrence has been cited as evidence that it marks the point at which a wide range of in-context learning (ICL) capabilities begin to emerge. To analyze the training dynamics of few-shot abstractive ICL capabilities (i.e., tasks where the answer is not contained in the input context), we propose Hapax, a training regime that omits the loss contribution of tokens predictable by the induction circuit to control training pressure for copying. Using this intervention, we find that abstractive ICL capabilities can still emerge, despite a substantial reduction in inductive copying and the absence of the characteristic phase changes observed during Vanilla training. Analysis of attention heads shows that Hapax-trained models develop fewer and weaker induction heads. Together, these results suggest that the developmental link between emergent phenomena associated with induction circuits and the emergence of abstractive ICL is weaker than previously hypothesized. More broadly, Hapax provides a controlled framework for analyzing how copying-related training signals shape the development of in-context learning.
Read the full abstract

All authors Kerem Sahin, Sheridan Feucht, Adam Belfki, Jannik Brinkmann, Aaron Mueller, David Bau, Chris Wendler

Computation and Language

OckBench: Measuring the Efficiency of LLM Reasoning

Zheng Du, Hao Kang, Song Han et al. (+2 more)

Large language models (LLMs) such as GPT-5 and Gemini 3 have pushed the frontier of automated reasoning and code generation. Yet current benchmarks emphasize accuracy and output quality, neglecting a critical dimension: efficiency of token usage. The token efficiency is highly variable in practical. Models solving the same problem with similar accuracy can exhibit up to a \textbf{5.0$\times$} difference in token length, leading to massive gap of model reasoning ability. Such variance exposes significant redundancy, highlighting the critical need for a standardized benchmark to quantify the gap of token efficiency. Thus, we introduce OckBench, the first benchmark that jointly measures accuracy and token efficiency across reasoning and coding tasks. Our evaluation reveals that token efficiency remains largely unoptimized across current models, significantly inflating serving costs and latency. These findings provide a concrete roadmap for the community to optimize the latent reasoning ability, token efficiency. Ultimately, we argue for an evaluation paradigm shift: tokens must not be multiplied beyond necessity. Our benchmarks are available at https://ockbench.github.io/.
Read the full abstract

All authors Zheng Du, Hao Kang, Song Han, Tushar Krishna, Ligeng Zhu

Computation and Language · Cross-listed from cs.SD

Persian Musical Instruments Classification Using Polyphonic Data Augmentation

Diba Hadi Esfangereh, Mohammad Hossein Sameti, Sepehr Harfi Moridani et al. (+2 more)

Musical instrument classification is essential for music information retrieval (MIR) and generative music systems. However, research on non-Western traditions, particularly Persian music, remains limited. We address this gap by introducing a new dataset of isolated recordings covering seven traditional Persian instruments, two common but originally non-Persian instruments (i.e., violin, piano), and vocals. We propose a culturally informed data augmentation strategy that generates realistic polyphonic mixtures from monophonic samples. Using the MERT model (Music undERstanding with large-scale self-supervised Training) with a classification head, we evaluate our approach with out-of-distribution data which was obtained by manually labeling segments of traditional songs. On real-world polyphonic Persian music, the proposed method yielded the best ROC-AUC (0.795), highlighting complementary benefits of tonal and temporal coherence. These results demonstrate the effectiveness of culturally grounded augmentation for robust Persian instrument recognition and provide a foundation for culturally inclusive MIR and diverse music generation systems.
Read the full abstract

All authors Diba Hadi Esfangereh, Mohammad Hossein Sameti, Sepehr Harfi Moridani, Leili Javidpour, Mahdieh Soleymani Baghshah

Computation and Language · Cross-listed from cs.CV

Long Grounded Thoughts: Synthesizing Visual Problems and Reasoning Chains at Scale

David Acuna, Chao-Han Huck Yang, Yuntian Deng et al. (+6 more)

Despite rapid progress, multimodal reasoning still lacks a systematic approach to synthesize large-scale vision-centric datasets beyond visual math. We introduce a framework able to synthesize vision-centric problems spanning diverse levels of complexity, and the resulting dataset with over 1M high-quality problems including: reasoning traces, preference data, and instruction prompts supporting SFT, offline and online RL. Our vision-centric synthesis framework uses a two-stage process focusing on: (1) generating diverse verifiable questions from existing images at scale, and (2) creating complex compositional visual problems by merging simpler questions. Remarkably, finetuning Qwen2.5-VL-7B on our data outperforms existing open-data baselines across evaluated vision-centric benchmarks, and our best configurations match or surpass strong closed-data models such as MiMo-VL-7B-RL on Vstar Bench, CV-Bench and MMStar-V. Notably, despite being entirely vision-centric, our data transfers positively to text-only reasoning (MMLU-Pro, +3.7%) and audio reasoning (MMAU, +1.32%), demonstrating its effectiveness. Similarly, despite containing no embodied visual data, we observe notable gains (NiEH, +8.8%) when evaluating open-ended embodied QA. Lastly, we use our data to comprehensively analyze at scale (1M+) the entire VLM post-training pipeline showing that (i) SFT on high-quality data with cognitive behaviors on reasoning traces is essential to scale online RL, (ii) offline RL could match online RL's performance while disaggregating compute demands, and, (iii) SFT on high quality data also improve out-of-domain, cross-modality transfer.
Read the full abstract

All authors David Acuna, Chao-Han Huck Yang, Yuntian Deng, Jaehun Jung, Ximing Lu, Prithviraj Ammanabrolu, Hyunwoo Kim, Yuan-Hong Liao, Yejin Choi

Computation and Language · Cross-listed from cs.LG

TabDistill: Distilling Transformers into Neural Nets for Few-Shot Tabular Classification

Pasan Dissanayake and Sanghamitra Dutta

Transformer-based models have shown promising performance on tabular data compared to their classical counterparts such as neural networks and Gradient Boosted Decision Trees (GBDTs) in scenarios with limited training data. They utilize their pre-trained knowledge to adapt to new domains, achieving commendable performance with only a few training examples, also called the few-shot regime. However, the performance gain in the few-shot regime comes at the expense of significantly increased complexity and number of parameters. To circumvent this trade-off, we introduce TabDistill, a new strategy to distill the pre-trained knowledge in complex transformer-based models into simpler neural networks for effectively classifying tabular data. Our framework yields the best of both worlds: being parameter-efficient while performing well with limited training data. The distilled neural networks surpass classical baselines such as regular neural networks, XGBoost and logistic regression under equal training data, and in some cases, even the original transformer-based models that they were distilled from.
Read the full abstract