Computation and Language · Cross-listed from cs.LG
Financial Instruction Following Evaluation (FIFE)
Glenn Matlin, Siddharth, Anirudh JM et al. (+3 more)
Language Models (LMs) struggle with complex, interdependent instructions, particularly in high-stakes domains like finance where precision is critical. We introduce FIFE, a novel, high-difficulty
benchmark designed to assess LM instruction-following capabilities for financial analysis tasks. FIFE comprises 88 human-authored
prompts and employs a verification system with chainable, verifiable constraints for fine-grained reward signals. We evaluate 53 models (proprietary, open-weight, open-source) in a zero-shot setting. Our key findings reveal a clear performance hierarchy: the top open-weight model (76.1 strict / 79.5 loose) surpasses the leading proprietary system (65.9 strict / 70.5 loose), while the best open-source models lag significantly (45.5 strict / 48.9 loose). However, even top-performing models struggle with FIFE's complex requirements, failing to achieve perfect compliance. We release our
dataset and code as an open-source resource to promote research in Reinforcement Learning for the financial domain.
Read the full abstract ▾
All authors Glenn Matlin, Siddharth, Anirudh JM, Aditya Shukla, Yahya Hassan, Sudheer Chava
Computation and Language · Cross-listed from eess.AS
KidSpeak: A General Multi-purpose LLM for Kids' Speech Recognition and Screening
Rohan Sharma, Dancheng Liu, Jingchen Sun et al. (+4 more)
With the rapid advancement of conversational and
diffusion-based AI, there is a growing adoption of AI in educational services, ranging from grading and assessment tools to personalized learning systems that provide targeted support for students. However, this adaptability has yet to fully extend to the domain of children's
speech, where existing models often fail due to their reliance on
datasets designed for clear, articulate adult
speech. Children, particularly those in early developmental stages or with
speech and
language pathologies, present unique challenges that current AI models and
datasets are ill-equipped to handle. To address this, we introduce KidSpeak, a multi-task
speech-enhanced Foundation Model capable of both generative and discriminative tasks specifically tailored to children's
speech patterns. Our framework employs a two-stage training process that incorporates phonetic knowledge into the
speech encoder, achieving an average accuracy of 87% across four separate tasks. Furthermore, recognizing the limitations of scalable human annotation and existing
speech alignment tools, we propose the Flexible and Automatic
Speech Aligner (FASA) and leverage the method to construct high quality
datasets for training and evaluation. This novel
alignment tool significantly improves the quality of aligned children's
speech from noisy data, enhancing data quality by 13.6x compared to human annotations, as demonstrated on the CHILDES
dataset. To the best of our knowledge, KidSpeak and FASA represent the first comprehensive solution designed for
speech and
language therapy in children, offering both a multi-purpose
speech LLM and a robust
alignment tool.
Read the full abstract ▾
All authors Rohan Sharma, Dancheng Liu, Jingchen Sun, Shijie Zhou, Jiayu Qin, Jinjun Xiong, Changyou Chen
Computation and Language
CAIRNS: Balancing Readability and Scientific Accuracy in Climate Adaptation Question Answering
Liangji Kong, Aditya Joshi and Sarvnaz Karimi
Climate adaptation strategies are proposed in response to
climate change. They are practised in agriculture to sustain food production. These strategies can be found in unstructured data (for example, scientific literature from the Elsevier website) or structured (heterogeneous
climate data via government APIs). We present
Climate Adaptation question-answering with Improved Readability and Noted Sources (CAIRNS), a framework that enables experts – farmer advisors – to obtain credible preliminary answers from complex evidence sources from the web. It enhances readability and citation reliability through a structured ScholarGuide
prompt and achieves robust evaluation via a consistency-weighted hybrid evaluator that leverages inter-model agreement with experts. Together, these components enable readable, verifiable, and domain-grounded question-answering without fine-tuning or reinforcement learning. Using a previously reported
dataset of expert-curated question-answers, we show that CAIRNS outperforms the baselines on most of the metrics. Our thorough ablation study confirms the results on all metrics. To validate our LLM-based evaluation, we also report an analysis of correlations against human judgment.
Read the full abstract ▾
Computation and Language
DETAIL Matters: Measuring the Impact of Prompt Specificity on Reasoning in Large Language Models
Olivia Kim
Prompt design plays a critical role in the
reasoning performance of large
language models (LLMs), yet the impact of
prompt specificity - how detailed or vague a
prompt is - remains understudied. This paper introduces DETAIL, a framework for evaluating LLM performance across varying levels of
prompt specificity. We generate multi-level
prompts using GPT-4, quantify specificity via perplexity, and assess correctness using GPT-based semantic equivalence. Experiments on 30 novel
reasoning tasks across GPT-4 and O3-mini reveal that specificity improves accuracy, especially for smaller models and procedural tasks. Our results highlight the need for adaptive prompting strategies and provide tools and data to support further research.
Read the full abstract ▾
Computation and Language
Lightweight Latent Reasoning for Narrative Tasks
Alexander Gurung, Esmeralda S. Whitammer and Mirella Lapata
Large
language models (LLMs) tackle complex tasks by generating long chains of thought or "
reasoning traces" that act as latent variables in the generation of an output given a query. A model's ability to generate such traces can be optimized with reinforcement learning (RL) to improve their utility in predicting an answer. This optimization comes at a high computational cost, especially for narrative-related tasks that involve retrieving and processing many
tokens. To this end, we propose LiteReason, a latent
reasoning method that can be interleaved with standard
token sampling and easily combined with RL techniques. LiteReason employs a lightweight
Reasoning Projector module, trained to produce continuous latent
tokens that help the model 'skip'
reasoning steps. During RL, the policy model decides when to activate the projector, switching between latent and discrete
reasoning as needed. Experimental results on plot hole detection and book chapter generation show that our method outperforms latent
reasoning baselines and comes close to matching non-latent RL training, while reducing final
reasoning length by 77-92%. Overall, LiteReason guides RL training to a more efficient part of the performance-computation tradeoff curve.
Read the full abstract ▾
Computation and Language
Vukosi Marivate, Kayode Olaleye, Sitwala Mundia et al. (+19 more)
This paper introduces Swivuriso, a 3000-hour
multilingual speech dataset developed as part of the African Next Voices project, to support the development and benchmarking of automatic
speech recognition (ASR) technologies in seven South African
languages. Covering agriculture, healthcare, and general domain topics, Swivuriso addresses significant gaps in existing ASR
datasets. We describe the design principles, ethical considerations, and data collection procedures that guided the
dataset creation. We present baseline results of training/finetuning ASR models with this data and compare to other ASR
datasets for the langauges concerned.
Read the full abstract ▾
All authors Vukosi Marivate, Kayode Olaleye, Sitwala Mundia, Andinda Bakainga, Unarine Netshifhefhe, Mahmooda Milanzie, Tsholofelo Hope Mogale, Thapelo Sindane, Zainab Abdulrasaq, Kesego Mokgosi, Chijioke Okorie, Nia Zion Van Wyk, Graham Morrissey, Dale Dunbar, Francois Smit, Tsosheletso Chidi, Rooweither Mabuya, Andiswa Bukula, Respect Mlambo, Tebogo Macucwa, Idris Abdulmumin, and Seani Rananga
Computation and Language
A Knowledge-Based Language Model: Deducing Grammatical Knowledge in a Multi-Agent Language Acquisition Simulation
David Ph. Shakouri, Crit Cremers and Niels O. Schiller
This paper presents an initial study performed by the MODOMA system. The MODOMA is a computational multi-
agent laboratory environment for unsupervised
language acquisition experiments such that acquisition is based on the interaction between two
language models, an adult and a child
agent. Although this framework employs statistical as well as rule-based procedures, the result of
language acquisition is a knowledge-based
language model, which can be used to generate and parse new utterances of the target
language. This system is fully parametrized and researchers can control all aspects of the experiments while the results of
language acquisition, that is, the acquired grammatical knowledge, are explicitly represented and can be consulted. Thus, this system introduces novel possibilities for conducting computational
language acquisition experiments. The experiments presented by this paper demonstrate that functional and content categories can be acquired and represented by the daughter
agent based on training and test data containing different amounts of exemplars generated by the adult
agent. Interestingly, similar patterns, which are well-established for human-generated data, are also found for these machine-generated data. As the procedures resulted in the successful acquisition of discrete grammatical categories by the child
agent, these experiments substantiate the validity of the MODOMA approach to modelling
language acquisition.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.SD
Story2MIDI: Emotionally Aligned Music Generation from Text
Mohammad Shokri, Alexandra C. Salem, Gabriel Levine et al. (+2 more)
In this paper, we introduce Story2MIDI, a sequence-to-sequence
Transformer-based model for generating
emotion-aligned music from a given piece of text. To develop this model, we construct the Story2MIDI
dataset by merging existing
datasets for sentiment analysis from text and
emotion classification in music. The resulting
dataset contains pairs of text blurbs and music pieces that evoke the same
emotions in the reader or listener. Despite the small scale of our
dataset and limited computational resources, our results indicate that our model effectively learns
emotion-relevant features in music and incorporates them into its generation process, producing samples with diverse emotional responses. We evaluate the generated outputs using objective musical metrics and a human listening study, confirming the model's ability to capture intended emotional cues.
Read the full abstract ▾
All authors Mohammad Shokri, Alexandra C. Salem, Gabriel Levine, Johanna Devaney, Sarah Ita Levitan
Computation and Language
Think Before You Prune: Self-Reflective Structured Pruning for Reasoning Language Models
Ziyan Wang, Enmao Diao, Qi Le et al. (+5 more)
Reasoning LLMs (RLMs) such as OpenAI o1, DeepSeek-R1, and Qwen3 deliver strong multi-step
reasoning through chain-of-thought generation, but their large model sizes and lengthy decode-time outputs make them costly to deploy and unsuitable for resource-constrained settings. To reduce computing and
memory cost,
pruning offers a promising solution by removing unimportant parameters. However, despite their success on standard LLMs, existing
pruning methods severely damage RLMs, as even moderate sparsity (e.g., 20%) can collapse accuracy and completely disrupt the model's
reasoning coherence. We begin by analyzing why existing
pruning pipelines fail on
reasoning LLMs and find that their brittleness largely stems from a mismatch between the calibration data, the
pruning objective, and the model's decode-time
reasoning behavior. Our study further shows that the most reliable calibration signal comes not from human-written labels but from the model's own self-generated
reasoning traces, which more accurately reflect its inference distribution. Guided by these insights, we introduce RESP, a self-reflective structured
pruning framework that aligns
pruning decisions with the model's
reasoning dynamics through self-generated calibration, decode-only gradient-based importance estimation, and progressive regeneration that maintains calibration fidelity as sparsity increases. Experiments on Qwen3-8B demonstrate that RESP markedly outperforms existing structured
pruning methods on both GSM8K and MathQA, preserving near-dense accuracy at 20-30% sparsity and substantially mitigating performance collapse at higher sparsity levels. At 40% sparsity, RESP attains 81.3% accuracy on GSM8K and 59.6% on MathQA, surpassing the strongest baselines by 66.87% and 47%, respectively.
Read the full abstract ▾
All authors Ziyan Wang, Enmao Diao, Qi Le, Pu Wang, Guanchu Wang, Minwoo Lee, Shu-ping Yeh, Li Yang
Computation and Language · Cross-listed from cs.CR
Factor(T,U): Factored Cognition Strengthens Monitoring of Untrusted AI
Aaron Sandoval and Cody Rushing
The field of AI Control seeks to develop robust control protocols, deployment safeguards for untrusted AI which may be intentionally subversive. However, existing protocols that rely on weaker monitors to detect unsafe behavior often fail on complex tasks beyond the monitor's comprehension. We develop control protocols based on factored cognition, in which a trusted model decomposes a task, an untrusted model solves each resultant child task in isolation, and the results are reassembled into a full solution. These protocols may improve safety by several means, such as by simplifying the context for monitors, or by obscuring vulnerabilities in the environment. We implement our protocols in the APPS coding setting and red team them against
backdoor attempts from an
adversarial GPT-4.1 Nano. We find that: (i) Adding factored cognition to a trusted monitoring protocol can boost safety from 41% to 63%; (ii) Safety improves because monitor performance improves; (iii) Factored cognition makes it no harder for capable LLMs to write
backdoors in APPS. While our protocols show low usefulness in APPS, they hold promise for more complex tasks.
Read the full abstract ▾
Computation and Language
Feature Selection Empowered BERT for Detection of Hate Speech with Vocabulary Augmentation
Pritish N. Desai, Tanay Kewalramani and Srimanta Mandal
Abusive
speech on social media poses a persistent and evolving challenge, driven by the continuous emergence of novel slang and obfuscated terms designed to circumvent detection systems. In this work, we present a data efficient strategy for fine tuning BERT on hate
speech classification by significantly reducing training set size without compromising performance. Our approach employs a TF IDF-based sample selection mechanism to retain only the most informative 75 percent of examples, thereby minimizing training overhead. To address the limitations of BERT's native vocabulary in capturing evolving hate
speech terminology, we augment the tokenizer with domain-specific slang and lexical variants commonly found in abusive contexts. Experimental results on a widely used hate
speech dataset demonstrate that our method achieves competitive performance while improving computational efficiency, highlighting its potential for scalable and adaptive abusive content moderation.
Read the full abstract ▾
Computation and Language
Four Over Six: More Accurate NVFP4 Quantization with Adaptive Block Scaling
Jack Cook, Junxian Guo, Guangxuan Xiao et al. (+7 more)
As large
language models have grown larger, interest has grown in low-precision numerical formats such as NVFP4 as a way to improve speed and reduce
memory usage. However, quantizing models to NVFP4 remains challenging as the lack of precision generally degrades model performance. In this work, we address this issue with Four Over Six (4/6), a modification to the block-scaled NVFP4 quantization
algorithm that yields reduced quantization error. Unlike integer formats, floating point formats have non-uniform step sizes which create larger quantization error on larger values. 4/6 takes advantage of this by adaptively scaling some blocks to smaller FP4 values, making the distribution of representable values more uniform and reducing quantization error for near-maximal values. We show that 4/6 can be implemented efficiently on modern hardware accelerators, resulting in performance gains during both pre-training and inference with minimal computational overhead. In pre-training experiments with the Nemotron 3 Nano 30B-A3B model architecture, we find that 4/6 brings training loss closer to BF16 compared to models trained with current state-of-the-art NVFP4 training recipes. Our code is available at https://github.com/mit-han-lab/fouroversix.
Read the full abstract ▾
All authors Jack Cook, Junxian Guo, Guangxuan Xiao, Yujun Lin, Keith Wyss, Mahdi Nazemi, Asit Mishra, Carlo del Mundo, Tijmen Blankevoort, Song Han
Computation and Language
The Art of Scaling Test-Time Compute for Large Language Models
Aradhye Agarwal, Ayan Sengupta and Tanmoy Chakraborty
Test-time scaling (TTS) – the dynamic allocation of compute during inference – is a promising direction for improving
reasoning in large
language models (LLMs). However, a systematic comparison of well-known TTS strategies under identical conditions is missing, and the influence of model type and problem difficulty on performance remains unclear. To address these gaps, we conduct the first large-scale study of TTS, spanning over thirty billion
tokens generated using eight open-source LLMs (7B to 235B parameters), across four
reasoning datasets. We observe three consistent trends: (1) no single TTS strategy universally dominates; (2)
reasoning models exhibit distinct trace-quality patterns across problem difficulty and trace length, forming short-horizon and long-horizon categories; and (3) for a given model type, the optimal TTS performance scales monotonically with compute budget. Based on these insights, we provide a practical recipe for selecting the best TTS strategy, considering problem difficulty, model type, and compute budget, providing a practical guide to effective inference-time scaling.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.LG
AlignSAE: Concept-Aligned Sparse Autoencoders
Minglai Yang, Xinyu Guo, Zhengliang Shi et al. (+4 more)
Large
Language Models (LLMs) encode factual knowledge within hidden parametric spaces that are difficult to inspect or control. While Sparse Autoencoders (SAEs) can decompose hidden activations into more fine-grained, interpretable features, they often struggle to reliably align these features with human-defined concepts, resulting in entangled and distributed feature representations. To address this, we introduce AlignSAE, a method that aligns SAE features with a predefined ontology through a "pre-train, then post-train" curriculum. After an initial unsupervised training phase, we apply supervised post-training to bind specific concepts to dedicated latent slots while preserving the remaining capacity for general reconstruction. This separation creates an interpretable interface where specific concepts can be inspected and controlled without interference from unrelated features. Empirical results on GPT-2 across synthetic biographical QA (1-hop, 6 concepts) and compositional relational
reasoning (2-hop, 20 concepts) demonstrate that AlignSAE enables precise causal interventions, such as reliable concept swaps (94.8% success in 1-hop), by targeting single, semantically aligned slots. AlignSAE achieves disentanglement improvements of $9$–$\infty\times$ over standard SAEs (measured via Mutual Information Gap), and further supports multi-hop
reasoning and a mechanistic probe of grokking-like generalization dynamics. Although our evaluation focuses on GPT-2, AlignSAE itself is model-agnostic.
Read the full abstract ▾
All authors Minglai Yang, Xinyu Guo, Zhengliang Shi, Jinhe Bi, Steven Bethard, Mihai Surdeanu, Liangming Pan
Computation and Language · Cross-listed from cs.AI
LLM CHESS: Benchmarking Reasoning and Instruction-Following in LLMs through Chess
Sai Kolasani, Maxim Saplin, Nicholas Crispino et al. (+5 more)
We introduce LLM CHESS, an evaluation framework designed to probe the generalization of
reasoning and instruction-following abilities in large
language models (LLMs) through extended agentic interaction in the domain of chess. We rank over 50 open and closed source models by playing against a random opponent using a range of behavioral metrics, including win and loss rates, move quality, move legality, hallucinated actions, and game duration. For a subset of top
reasoning models, we derive an Elo estimate by playing against a chess engine with variably configured skill, which allows for comparisons between models in an easily understandable way. Despite the simplicity of the instruction-following task and the weakness of the opponent, many state-of-the-art models struggle to complete games or achieve consistent wins. Similar to other
benchmarks on complex
reasoning tasks, our experiments reveal a clear separation between
reasoning and non-
reasoning models. However, unlike existing static
benchmarks, the
stochastic and dynamic nature of LLM CHESS uniquely reduces overfitting and memorization while preventing
benchmark saturation, proving difficult even for top
reasoning models. To support future work on evaluating
reasoning and instruction-following in LLMs, we release our experimental framework, a public leaderboard, and a
dataset of associated games.
Read the full abstract ▾
All authors Sai Kolasani, Maxim Saplin, Nicholas Crispino, Kyle Montgomery, Jared Quincy Davis, Matei Zaharia, Chi Wang, Chenguang Wang
Computation and Language · Cross-listed from cs.AI
Chain-of-Ground: Improving GUI Grounding via Iterative Reasoning and Reference Feedback
Aiden Yiliu Li, Bizhi Yu, Daoan Lei et al. (+2 more)
GUI grounding aims to align natural
language instructions with precise regions in complex user interfaces. Advanced
multimodal large
language models show strong ability in visual GUI grounding but still struggle with small or visually similar targets and ambiguity in real world layouts. These limitations arise from limited grounding capacity and from underuse of existing
reasoning potential. We present Chain of Ground CoG a training free multi step grounding framework that uses
multimodal large
language models for iterative visual
reasoning and refinement. Instead of direct prediction the model progressively reflects and adjusts its hypotheses leading to more accurate and interpretable localization. Our approach achieves 68.4 accuracy on the ScreenSpot Pro
benchmark an improvement of 4.8 points. To measure real world generalization we introduce TPanel UI a
dataset of 420 labeled industrial control panels with visual distortions such as blur and masking. On TPanel UI Chain of Ground improves over the strong baseline Qwen3 VL 235B by 6.9 points showing the effectiveness of multi step training free grounding across real world and digital interfaces. These results highlight a direction for unlocking grounding potential through structured iterative refinement instead of additional training.
Read the full abstract ▾
All authors Aiden Yiliu Li, Bizhi Yu, Daoan Lei, Tianhe Ren, Shilong Liu
Computation and Language · Cross-listed from cs.AI
Atomic Skills are the Prerequisite: When Reinforcement Learning Synthesizes Compositional Reasoning, and When It Only Amplifies
Sitao Cheng, Xunjian Yin, Ruiwen Zhou et al. (+5 more)
Does Reinforcement Learning (RL) merely amplify existing skills, or synthesize novel skills? We investigate this question through the lens of Complementary
Reasoning: the critical practical capability of integrating internal knowledge with external context, a prerequisite for reliable Continual Learning and
Retrieval-Augmented Generation. To avoid pre-training contamination, we construct a controlled semanticsynthetic
dataset of biographies and decompose this capability into two atomic skills: Parametric
Reasoning (retrieving facts encoded in model weights) and Contextual
Reasoning (processing novel in-context information). We present two findings. First, models supervised directly on the composite task reach high accuracy on seen facts and
reasoning paths (90%) but collapse on novel facts and
reasoning paths (18%), indicating that Supervised Fine-Tuning (SFT) relies on rote memorization rather than genuine skill integration. Second, RL bridges this generalization gap, acting as a skill synthesizer rather than a mere amplifier–but only under a strict prerequisite: it synthesizes new composite strategies only when the base model has first mastered the independent atomic skills via SFT. These results suggest that decoupled atomic training followed by RL offers a scalable path to complex novel
reasoning.
Read the full abstract ▾
All authors Sitao Cheng, Xunjian Yin, Ruiwen Zhou, Yuxuan Li, Xinyi Wang, Liangming Pan, William Yang Wang, Victor Zhong
Computation and Language
How Far Are We from Genuinely Useful Deep Research Agents?
Dingling Zhang, He Zhu, Jincheng Ren et al. (+15 more)
Deep Research
Agents (DRAs) aim to automatically produce analyst-level reports through iterative information
retrieval and synthesis. However, most existing DRAs were validated on question-answering
benchmarks, while research on generating comprehensive reports remains overlooked. Worse, current
benchmarks for report synthesis suffer from task complexity and subjective metrics – this fails to reflect user demands and limits the practical utility of generated reports. To address these gaps, we present Fine-grained DEepResearch bench (FINDER), an enhanced
benchmark consisting of 100 human-curated research tasks with 419 structured checklist items that standardize report structure, analytical depth, and factual grounding. Based on approximately 1,000 reports produced by mainstream DRAs, we further propose Deep rEsearch Failure Taxonomy (DEFT), the first failure taxonomy for deep research
agents. DEFT contains 14 fine-grained failure modes across
reasoning,
retrieval, and generation, and is built upon grounded theory with human-LLM co-annotating and inter-annotator reliability validation. Our experimental findings reveal that current DRAs struggle not with task comprehension but with evidence integration, verification, and
reasoning-resilient planning.
Read the full abstract ▾
All authors Dingling Zhang, He Zhu, Jincheng Ren, Kangqi Song, Xinran Zhou, Boyu Feng, Shudong Liu, Jiabin Luo, Weihao Xie, Zhaohui Wang, Tianrui Qin, King Zhu, Yuqing Wang, Qianben Chen, Yuchen Eleanor Jiang, Wei Wang, Jiaheng Liu, Wangchunshu Zhou
Computation and Language · Cross-listed from cs.LG
Agentic Policy Optimization via Instruction-Policy Co-Evolution
Han Zhou, Xingchen Wan, Ivan Vulić et al. (+1 more)
Reinforcement Learning with Verifiable Rewards (RLVR) has advanced the
reasoning capability of large
language models (LLMs), enabling autonomous
agents that can conduct effective multi-turn and tool-integrated
reasoning. While instructions serve as the primary protocol for defining
agents, RLVR typically relies on static and manually designed instructions. However, those instructions may be suboptimal for the base model, and the optimal instruction may change as the
agent's policy improves and explores the interaction with the environment. To bridge the gap, we introduce INSPO, a novel Instruction-Policy co-
evolution framework that integrates instruction optimization as a dynamic component of the reinforcement learning (RL) loop. INSPO maintains a dynamic population of instruction candidates that are sampled with questions, where reward signals in RL loops are automatically attributed to each instruction, and low performers are periodically pruned. New instructions are generated and verified through an on-policy reflection mechanism, where an LLM-based optimizer analyzes past experience from a replay buffer and evolves more effective strategies given the current policy. We conduct extensive experiments on multi-turn
retrieval and
reasoning tasks, demonstrating that INSPO substantially outperforms strong baselines relying on static instructions. INSPO discovers innovative instructions that guide the
agent toward more strategic
reasoning paths, achieving substantial performance gains with only a marginal increase in computational overhead.
Read the full abstract ▾
All authors Han Zhou, Xingchen Wan, Ivan Vulić, Anna Korhonen
Computation and Language
Rectifying LLM Thought from Lens of Optimization
Junnan Liu, Hongwei Liu, Songyang Zhang et al. (+1 more)
Recent advancements in large
language models (LLMs) have been driven by their emergent
reasoning capabilities, particularly through long chain-of-thought (CoT) prompting, which enables thorough exploration and deliberation. Despite these advances, long-CoT LLMs often exhibit suboptimal
reasoning behaviors, such as overthinking and excessively protracted
reasoning chains, which can impair performance. In this paper, we analyze
reasoning processes through an optimization lens, framing CoT as a gradient descent procedure where each
reasoning step constitutes an update toward problem resolution. Building on this perspective, we introduce RePro (Rectifying Process-level Reward), a novel approach to refine LLM
reasoning during post-training. RePro defines a surrogate objective function to assess the optimization process underlying CoT, utilizing a dual scoring mechanism to quantify its intensity and stability. These scores are aggregated into a composite process-level reward, seamlessly integrated into reinforcement learning with verifiable rewards (RLVR) pipelines to optimize LLMs. Extensive experiments across multiple reinforcement learning
algorithms and diverse LLMs, evaluated on
benchmarks spanning mathematics, science, and coding, demonstrate that RePro consistently enhances
reasoning performance and mitigates suboptimal
reasoning behaviors.
Read the full abstract ▾
All authors Junnan Liu, Hongwei Liu, Songyang Zhang, Kai Chen
Computation and Language
Latent Debate: A Surrogate Framework for Interpreting LLM Thinking towards Binary Decisions
Lihu Chen, Xiang Yin and Francesca Toni
Understanding the internal `thinking' process of Large
Language Models (LLMs) and %how this relates their connections to
hallucinations remains a key challenge. To this end, we introduce latent debate, a structured surrogate framework for interpreting model outputs on True/False prediction tasks through the lens of internal latent arguments and interactions amongst them. Unlike human debates, latent debate captures the hidden supporting and attacking signals that arise within a model during a single inference. We first present a model- and task-agnostic conceptual framework, and then instantiate it symbolically to approximate the thinking process of LLMs towards binary decisions. Empirical studies demonstrate that our latent debate is a faithful structured surrogate model that has highly consistent predictions with the original LLM, while providing a form of interpretability. We also demonstrate that our latent debate provides a strong baseline for
hallucination detection. Specifically, we identify strong correlations between debate patterns and
hallucinations, such as a high degree of disagreement in the middle layers of the latent debate surrogate is linked to a higher risk of
hallucinations. Our findings suggest that latent debate provides a new framework for understanding the internal processes of LLMs.
Read the full abstract ▾
Computation and Language
OPOR-Bench: Evaluating Large Language Models on Online Public Opinion Report Generation
Jinzheng Yu, Yang Xu, Haozhen Li et al. (+5 more)
Online Public Opinion Reports consolidate news and social media for timely crisis management by governments and enterprises. While large
language models have made automated report generation technically feasible, systematic research in this specific area remains notably absent, particularly lacking formal task definitions and corresponding
benchmarks. To bridge this gap, we define the Automated Online Public Opinion Report Generation (OPOR-GEN) task and construct OPOR-BENCH, an event-centric
dataset covering 463 crisis events with their corresponding news articles, social media posts, and a reference summary. To evaluate report quality, we propose OPOR-EVAL, a novel
agent-based framework that simulates human expert evaluation by analyzing generated reports in context. Experiments with frontier models demonstrate that our framework achieves high correlation with human judgments. Our comprehensive task definition,
benchmark dataset, and evaluation framework provide a solid foundation for future research in this critical domain.
Read the full abstract ▾
All authors Jinzheng Yu, Yang Xu, Haozhen Li, Junqi Li, Yifan Feng, Ligu Zhu, Hao Shen, Lei Shi
Computation and Language
Exploring Human Perceptions of AI Responses: Insights from a Mixed-Methods Study on Risk Mitigation in Generative Models
Heloisa Candello, Muneeza Azmat, Uma Sushmitha Gunturi et al. (+7 more)
With the rapid uptake of generative AI, investigating human perceptions of generated responses has become crucial. A major challenge is their `aptitude' for hallucinating and generating harmful contents. Despite major efforts for implementing guardrails, human perceptions of these mitigation strategies are largely unknown. We conducted a mixed-method experiment for evaluating the responses of a mitigation strategy across multiple-dimensions: faithfulness,
fairness, harm-removal capacity, and relevance. In a within-subject study design, 57 participants assessed the responses under two conditions: harmful response plus its mitigation and solely mitigated response. Results revealed that participants' native
language, AI work experience, and annotation familiarity significantly influenced evaluations. Participants showed high sensitivity to linguistic and contextual attributes, penalizing minor grammar errors while rewarding preserved semantic contexts. This contrasts with how
language is often treated in the quantitative evaluation of LLMs. We also introduced new metrics for training and evaluating mitigation strategies and insights for human-AI evaluation studies.
Read the full abstract ▾
All authors Heloisa Candello, Muneeza Azmat, Uma Sushmitha Gunturi, Raya Horesh, Rogerio Abreu de Paula, Heloisa Pimentel, Marcelo Carpinette Grave, Aminat Adebiyi, Tiago Machado, Maysa Malfiza Garcia de Macedo
Computation and Language
Cross-Lingual Interleaving for Speech Language Models
Adel Moumen, Guangzhi Sun and Philip C. Woodland
Spoken
Language Models (SLMs) aim to learn linguistic competence directly from
speech using discrete units, widening access to Natural
Language Processing (NLP) technologies for
languages with limited written resources. However, progress has been largely English-centric due to scarce spoken evaluation
benchmarks and training data, making cross-lingual learning difficult. We present a cross-lingual interleaving method that mixes
speech tokens across
languages without textual supervision. We also release an EN-FR training
dataset, TinyStories (~42k hours), together with EN-FR spoken StoryCloze and TopicCloze
benchmarks for cross-lingual semantic evaluation, both synthetically generated using GPT-4. On 360M and 1B SLMs under matched training-
token budgets, interleaving improves monolingual semantic accuracy, enables robust cross-lingual continuation, and strengthens cross-lingual hidden-state
alignment. Taken together, these results indicate that cross-lingual interleaving is a simple, scalable route to building
multilingual SLMs that understand and converse across
languages. All resources will be made open-source to support reproducibility.
Read the full abstract ▾
Computation and Language
BHRAM-IL: A Benchmark for Hallucination Recognition and Assessment in Multiple Indian Languages
Hrishikesh Terdalkar, Kirtan Bhojani, Aryan Dongare et al. (+1 more)
Large
language models (LLMs) are increasingly deployed in
multilingual applications but often generate plausible yet incorrect or misleading outputs, known as
hallucinations. While
hallucination detection has been studied extensively in English, under-resourced Indian
languages remain largely unexplored. We present BHRAM-IL, a
benchmark for
hallucination recognition and assessment in multiple Indian
languages, covering Hindi, Gujarati, Marathi, Odia, along with English. The
benchmark comprises 36,047 curated questions across nine categories spanning factual, numerical,
reasoning, and linguistic tasks. We evaluate 14 state-of-the-art
multilingual LLMs on a
benchmark subset of 10,265 questions, analyzing cross-lingual and factual
hallucinations across
languages, models, scales, categories, and domains using category-specific metrics normalized to (0,1) range. Aggregation over all categories and models yields a primary score of 0.23 and a
language-corrected fuzzy score of 0.385, demonstrating the usefulness of BHRAM-IL for
hallucination-focused evaluation. The
dataset, and the code for generation and evaluation are available on GitHub (https://github.com/sambhashana/BHRAM-IL/) and HuggingFace (https://huggingface.co/
datasets/sambhashana/BHRAM-IL/) to support future research in
multilingual hallucination detection and mitigation.
Read the full abstract ▾
All authors Hrishikesh Terdalkar, Kirtan Bhojani, Aryan Dongare, Omm Aditya Behera
Computation and Language
Beyond SFT: Reinforcement Learning for Safer Large Reasoning Models with Better Reasoning Ability
Jinghan Jia, Nathalie Baracaldo and Sijia Liu
Large
reasoning models (LRMs) extend large
language models by generating explicit chain-of-thought (CoT)
reasoning, significantly improving mathematical and logical problem solving. However, this explicit
reasoning process also introduces new safety risks, as unsafe behaviors often emerge within intermediate
reasoning trajectories, even when final answers appear harmless. Existing safety
alignment approaches primarily rely on supervised fine-tuning (SFT) over safety-oriented long CoT
datasets. While intuitive, we find that SFT produces inconsistent safety improvements, degrades
reasoning ability, and generalizes poorly across model families. These limitations suggest that purely supervised approaches are insufficient for robust safety
alignment in LRMs. To address this, we investigate reinforcement learning (RL) as a complementary optimization framework for LRM safety training. Unlike SFT, RL directly optimizes model policies with reward feedback, enabling more adaptive and stable
alignment. Extensive experiments across multiple model families and
benchmarks show that RL achieves stronger and more consistent safety gains while maintaining
reasoning competence. Further analysis of reflection dynamics and
token-level
entropy reveals that RL suppresses unsafe exploratory
reasoning while preserving reflective depth, leading to safer and more reliable
reasoning processes.
Read the full abstract ▾
Computation and Language
InnoGym: Benchmarking the Innovation Potential of AI Agents
Jintian Zhang, Kewei Xu, Jingsheng Zheng et al. (+10 more)
LLMs and
Agents have achieved impressive progress in code generation, mathematical
reasoning, and scientific discovery. However, existing
benchmarks primarily measure correctness, overlooking the diversity of methods behind solutions. True innovation depends not only on producing correct answers but also on the originality of the approach. We present InnoGym, the first
benchmark and framework designed to systematically evaluate the innovation potential of AI
agents. InnoGym introduces two complementary metrics: performance gain, which measures improvement over the best-known solutions, and novelty, which captures methodological differences from prior approaches. The
benchmark includes 18 carefully curated tasks from real-world engineering and scientific domains, each standardized through resource filtering, evaluator validation, and solution collection. In addition, we provide iGym, a unified execution environment for reproducible and long-horizon evaluations. Extensive experiments show that while some
agents produce novel approaches, their lack of robustness limits performance gains. These results highlight a key gap between creativity and effectiveness, underscoring the need for
benchmarks that evaluate both.
Read the full abstract ▾
All authors Jintian Zhang, Kewei Xu, Jingsheng Zheng, Zhuoyun Yu, Yuqi Zhu, Yujie Luo, Lanning Wei, Shuofei Qiao, Lun Du, Da Zheng, Shumin Deng, Huajun Chen, Ningyu Zhang
Computation and Language · Cross-listed from cs.AI
H-Neurons: On the Existence, Impact, and Origin of Hallucination-Associated Neurons in LLMs
Cheng Gao, Huimin Chen, Chaojun Xiao et al. (+3 more)
Large
language models (LLMs) frequently generate
hallucinations – plausible but factually incorrect outputs – undermining their reliability. While prior work has examined
hallucinations from macroscopic perspectives such as training data and objectives, the underlying
neuron-level mechanisms remain largely unexplored. In this paper, we conduct a systematic investigation into
hallucination-associated
neurons (H-
Neurons) in LLMs from three perspectives: identification, behavioral impact, and origins. Regarding their identification, we demonstrate that a remarkably sparse subset of
neurons (less than $0.1\%$ of total
neurons) can reliably predict
hallucination occurrences, with strong generalization across diverse scenarios. In terms of behavioral impact, controlled interventions reveal that these
neurons are causally linked to over-compliance behaviors. Concerning their origins, we trace these
neurons back to the pre-trained base models and find that these
neurons remain predictive for
hallucination detection, indicating they emerge during pre-training. Our findings bridge macroscopic behavioral patterns with microscopic neural mechanisms, offering insights for developing more reliable LLMs.
Read the full abstract ▾
All authors Cheng Gao, Huimin Chen, Chaojun Xiao, Zhiyi Chen, Zhiyuan Liu, Maosong Sun
Computation and Language
Reasoning About the Unsaid: Misinformation Detection with Omission-Aware Graph Inference
Zhengjia Wang, Danding Wang, Qiang Sheng et al. (+2 more)
This paper investigates the detection of misinformation, which deceives readers by explicitly fabricating misleading content or implicitly omitting important information necessary for informed judgment. While the former has been extensively studied, omission-based deception remains largely overlooked, even though it can subtly guide readers toward false conclusions under the illusion of completeness. To pioneer in this direction, this paper presents OmiGraph, the first omission-aware framework for misinformation detection. Specifically, OmiGraph constructs an omission-aware graph for the target news by utilizing a contextual environment that captures complementary perspectives of the same event, thereby surfacing potentially omitted contents. Based on this graph, omission-oriented relation modeling is then proposed to identify the internal contextual dependencies, as well as the dynamic omission intents, formulating a comprehensive omission relation representation. Finally, to extract omission patterns for detection, OmiGraph introduces omission-aware message-passing and aggregation that establishes holistic deception perception by integrating the omission contents and relations. Experiments show that, by considering the omission perspective, our approach attains remarkable performance, achieving average improvements of +5.4% F1 and +5.3% ACC on two large-scale
benchmarks.
Read the full abstract ▾
All authors Zhengjia Wang, Danding Wang, Qiang Sheng, Jiaying Wu, Juan Cao
Computation and Language
Beware of Reasoning Overconfidence: Pitfalls in the Reasoning Process for Multi-solution Tasks
Jiannan Guan, Qiguang Chen, Libo Qin et al. (+5 more)
Large
Language Models (LLMs) excel in
reasoning tasks requiring a single correct answer, but they perform poorly in multi-solution tasks that require generating comprehensive and diverse answers. We attribute this limitation to
reasoning overconfidence: a tendency to express undue certainty in an incomplete solution set. To examine the effect, we introduce MuSoBench, a
benchmark of multi-solution problems. Experiments show that the conventional short chain-of-thought (Short-CoT) prompting paradigm exhibits pronounced overconfidence, whereas the emerging long chain-of-thought (Long-CoT) approach mitigates it through iterative exploration and self-reflection. We further characterise observable behaviours and influential factors. To probe the underlying cause, we propose the cognitive-
rigidity hypothesis, which posits that overconfidence arises when the
reasoning process prematurely converges on a narrow set of thought paths. An
attention-
entropy analysis offers preliminary support for this view. These findings provide tools for assessing the completeness of LLM
reasoning and highlight the need to move evaluation beyond single-answer accuracy toward comprehensive exploration.
Read the full abstract ▾
All authors Jiannan Guan, Qiguang Chen, Libo Qin, Dengyun Peng, Jinhao Liu, Liangyu Huo, Jian Xie, Wanxiang Che