Computation and Language
Metric Calculating Benchmark: Code-Verifiable Complicate Instruction Following Benchmark for Large Language Models
Hyeonseok Moon, Seongtae Hong, Jaehyung Seo et al. (+1 more)
Recent frontier-level LLMs have saturated many previously difficult
benchmarks, leaving little room for further differentiation. This progress highlights the need for challenging
benchmarks that provide objective verification. In this paper, we introduce MCBench, a
benchmark designed to evaluate whether LLMs can execute string-matching NLP metrics by strictly following step-by-step instructions. Unlike prior
benchmarks that depend on subjective judgments or general
reasoning, MCBench offers an objective, deterministic and codeverifiable evaluation. This setup allows us to systematically test whether LLMs can maintain accurate step-by-step execution, including instruction adherence, numerical computation, and long-range consistency in handling intermediate results. To ensure objective evaluation of these abilities, we provide a parallel reference code that can evaluate the accuracy of LLM output. We provide three evaluative metrics and three
benchmark variants designed to measure the detailed instruction understanding capability of LLMs. Our analyses show that MCBench serves as an effective and objective tool for evaluating the capabilities of cutting-edge LLMs.
Read the full abstract ▾
All authors Hyeonseok Moon, Seongtae Hong, Jaehyung Seo, Heuiseok Lim
Computation and Language
Standard-to-Dialect Transfer Trends Differ across Text and Speech: A Case Study on Intent and Topic Classification in German Dialects
Verena Blaschke, Miriam Winkler and Barbara Plank
Research on cross-dialectal transfer from a standard to a non-standard dialect variety has typically focused on text data. However, dialects are primarily spoken, and non-standard spellings cause issues in text processing. We compare standard-to-dialect transfer in three settings: text models,
speech models, and cascaded systems where
speech first gets automatically transcribed and then further processed by a text model. We focus on German dialects in the context of written and spoken intent classification – releasing the first dialectal audio intent classification
dataset – with supporting experiments on topic classification. The
speech-only setup provides the best results on the dialect data while the text-only setup works best on the standard data. While the cascaded systems lag behind the text-only models for German, they perform relatively well on the dialectal data if the transcription system generates normalized, standard-like output.
Read the full abstract ▾
Computation and Language
Contrastive Weak-to-strong Generalization
Houcheng Jiang, Junfeng Fang, Jiaxin Wu et al. (+5 more)
Weak-to-strong generalization provides a promising paradigm for scaling large
language models (LLMs) by training stronger models on samples from aligned weaker ones, without requiring human feedback or explicit reward modeling. However, its robustness and generalization are hindered by the noise and biases in weak-model outputs, which limit its applicability in practice. To address this challenge, we leverage implicit rewards, which approximate explicit rewards through log-likelihood ratios, and reveal their structural equivalence with Contrastive Decoding (CD), a decoding strategy shown to reduce noise in LLM generation. Building on this connection, we propose Contrastive Weak-to-Strong Generalization (ConG), a framework that employs contrastive decoding between pre- and post-
alignment weak models to generate higher-quality samples. This approach enables more reliable capability transfer, denoising, and improved robustness, substantially mitigating the limitations of traditional weak-to-strong methods. Empirical results across different model families confirm consistent improvements, demonstrating the generality and effectiveness of ConG. Taken together, our findings highlight the potential of ConG to advance weak-to-strong generalization and provide a promising pathway toward AGI.
Read the full abstract ▾
All authors Houcheng Jiang, Junfeng Fang, Jiaxin Wu, Tianyu Zhang, Chen Gao, Xiang Wang, Xiangnan He, Yang Deng
Computation and Language
CS3-Bench: Evaluating and Enhancing Speech-to-Speech LLMs for Mandarin-English Code-Switching
Heyang Liu, Yuhao Wang, Ziyang Cheng et al. (+4 more)
The advancement of
multimodal large
language models has accelerated the development of
speech-to-
speech interaction systems. While natural monolingual interaction has been achieved, we find existing models exhibit deficiencies in
language alignment. In our proposed Code-Switching
Speech-to-
Speech Benchmark (CS3-Bench), experiments on 7 mainstream models demonstrate a relative performance drop of up to 66% in knowledge-intensive question answering and varying degrees of misunderstanding in open-ended conversations. Starting from a model with severe performance deterioration, we propose both data constructions and training approaches to improve the
language alignment capabilities, specifically employing Chain of Recognition (CoR) to enhance understanding and Keyword Highlighting (KH) to guide generation. Our approach improves the knowledge accuracy from 25.14% to 46.13%, with open-ended understanding rate from 64.5% to 86.5%, and significantly reduces pronunciation errors in the secondary
language. CS3-Bench is available at https://huggingface.co/
datasets/VocalNet/CS3-Bench.
Read the full abstract ▾
All authors Heyang Liu, Yuhao Wang, Ziyang Cheng, Ronghua Wu, Qunshan Gu, Yanfeng Wang, Yu Wang
Computation and Language
Do LLMs Really Need 10+ Thoughts for "Find the Time 1000 Days Later"? Towards Structural Understanding of LLM Overthinking
Xinliang Frederick Zhang, Anhad Mohananey, Alexandra Chronopoulou et al. (+5 more)
Models employing long chain-of-thought (CoT)
reasoning have shown superior performance on complex
reasoning tasks. Yet, this capability introduces a critical and often overlooked inefficiency – overthinking – models often engage in unnecessarily extensive
reasoning even for simple queries, incurring significant computations without accuracy improvements. While prior work has explored solutions to mitigate overthinking, a fundamental gap remains in our understanding of its underlying causes. Most existing analyses are limited to superficial, profiling-based observations, failing to delve into LLMs' inner workings. This study introduces a systematic, fine-grained analyzer of LLMs' thought process to bridge the gap, TRACE. We first
benchmark the overthinking issue, confirming that long-thinking models are five to twenty times slower on simple tasks with no substantial gains. We then use TRACE to first decompose the thought process into minimally complete sub-thoughts. Next, by inferring discourse relationships among sub-thoughts, we construct granular thought progression graphs and subsequently identify common thinking patterns for topically similar queries. Our analysis reveals two major patterns for open-weight thinking models – Explorer and Late Landing. This finding provides evidence that over-verification and over-exploration are the primary drivers of overthinking in LLMs. Grounded in thought structures, we propose a utility-based definition of overthinking, which moves beyond length-based metrics. This revised definition offers a more insightful understanding of LLMs' thought progression, as well as practical guidelines for principled overthinking management.
Read the full abstract ▾
All authors Xinliang Frederick Zhang, Anhad Mohananey, Alexandra Chronopoulou, Pinelopi Papalampidi, Somit Gupta, Tsendsuren Munkhdalai, Lu Wang, Shyam Upadhyay
Computation and Language
Ready to Translate, Not to Represent? Bias and Performance Gaps in Multilingual LLMs Across Language Families and Domains
Md. Faiyaz Abdullah Sayeedi, Md. Mahbub Alam, Subhey Sadi Rahman et al. (+5 more)
The rise of Large
Language Models (LLMs) has redefined Machine Translation (MT), enabling context-aware and fluent translations across hundreds of
languages and textual domains. Despite their remarkable capabilities, LLMs often exhibit uneven performance across
language families and specialized domains. Moreover, recent evidence reveals that these models can encode and amplify different biases present in their training data, posing serious concerns for
fairness, especially in low-resource
languages. To address these gaps, we introduce Translation Tangles, a unified framework and
dataset for evaluating the translation quality and
fairness of open-source LLMs. Our approach
benchmarks 24 bidirectional
language pairs across multiple domains using different metrics. We further propose a hybrid bias detection pipeline that integrates rule-based heuristics, semantic similarity filtering, and LLM-based validation. We also introduce a high-quality, bias-annotated
dataset based on human evaluations of 1,439 translation-reference pairs. The code and
dataset are accessible on GitHub: https://github.com/faiyazabdullah/TranslationTangles
Read the full abstract ▾
All authors Md. Faiyaz Abdullah Sayeedi, Md. Mahbub Alam, Subhey Sadi Rahman, Md. Adnanul Islam, Jannatul Ferdous Deepti, Tasnim Mohiuddin, Md Mofijul Islam, Swakkhar Shatabda
Computation and Language
AdaSwitch: Balancing Exploration and Guidance in Knowledge Distillation via Adaptive Switching
Jingyu Peng, Maolin Wang, Hengyi Cai et al. (+5 more)
Small
language models (SLMs) are crucial for applications with strict latency and computational constraints, yet achieving high performance remains challenging. Knowledge
distillation (KD) can transfer capabilities from large teacher models, but existing methods face a dilemma: off-policy
distillation provides high-quality supervision but suffers from exposure bias (training inference mismatch), while on-policy approaches ensure consistency but are limited by the low quality of student-generated outputs. To address these issues, we propose AdaSwitch, a novel approach that dynamically combines on-policy and off-policy generation via an adaptive switching mechanism. AdaSwitch allows the student to explore its predictions within its capability and selectively integrates teacher guidance only when divergence exceeds a context-aware threshold. This paradigm preserves generation consistency while ensuring high-quality supervision. Experiments on three
datasets demonstrate that AdaSwitch consistently improves accuracy and
reasoning capability with moderate overhead.
Read the full abstract ▾
All authors Jingyu Peng, Maolin Wang, Hengyi Cai, Yuchen Li, Kai Zhang, Shuaiqiang Wang, Dawei Yin, Xiangyu Zhao
Computation and Language · Cross-listed from cs.LG
Self-Improving LLM Agents at Test-Time
Emre Can Acikgoz, Cheng Qian, Heng Ji et al. (+2 more)
One paradigm of
language model (LM) fine-tuning relies on creating large training
datasets, under the assumption that high quantity and diversity will enable models to generalize to novel tasks after post-training. In practice, gathering large sets of data is inefficient, and training on them is prohibitively expensive; worse, there is no guarantee that the resulting model will handle complex scenarios or generalize better. Moreover, existing techniques rarely assess whether a training sample provides novel information or is redundant with the knowledge already acquired by the model, resulting in unnecessary costs. In this work, we explore a new test-time self-improvement method to create more effective and generalizable agentic LMs on-the-fly. The proposed
algorithm can be summarized in three steps: (i) first it identifies the samples that model struggles with (self-awareness), (ii) then generates similar examples from detected uncertain samples (self-data augmentation), and (iii) uses these newly generated samples at test-time fine-tuning (self-improvement). We study two variants of this approach: Test-Time Self-Improvement (TT-SI), where the same model generates additional training examples from its own uncertain cases and then learns from them, and contrast this approach with Test-Time
Distillation (TT-D), where a stronger model generates similar examples for uncertain cases, enabling student to adapt using distilled supervision. Empirical evaluations across different
agent benchmarks demonstrate that TT-SI improves the performance with +5.48% absolute accuracy gain on average across all
benchmarks and surpasses other standard learning methods, yet using 68x less training samples. Our findings highlight the promise of TT-SI, demonstrating the potential of self-improvement
algorithms at test-time as a new paradigm for building more capable
agents toward self-
evolution.
Read the full abstract ▾
All authors Emre Can Acikgoz, Cheng Qian, Heng Ji, Dilek Hakkani-Tür, Gokhan Tur
Computation and Language · Cross-listed from cs.LG
MetaDefense: Defending Finetuning-based Jailbreak Attack Before and During Generation
Weisen Jiang and Sinno Jialin Pan
This paper introduces MetaDefense, a novel framework for defending against finetuning-based jailbreak attacks in large
language models (LLMs). We observe that existing defense mechanisms fail to generalize to harmful queries disguised by unseen attack templates, despite LLMs being capable of distinguishing disguised harmful queries in the
embedding space. Based on these insights, we propose a two-stage defense approach: (i) pre-generation defense that detects harmful queries before response generation begins, and (ii) mid-generation defense that monitors partial responses during generation to prevent outputting more harmful content. Our MetaDefense trains the LLM to predict the harmfulness of both queries and partial responses using specialized
prompts, enabling early termination of potentially harmful interactions. Extensive experiments across multiple LLM architectures (LLaMA-2-7B, Qwen-2.5-3B-Instruct, and LLaMA-3.2-3B-Instruct) demonstrate that MetaDefense significantly outperforms existing defense mechanisms, achieving robust defense against harmful queries with seen and unseen attack templates while maintaining competitive performance on benign tasks. Code is available at https://github.com/ws-jiang/MetaDefense.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.SI
From Keywords to Clusters: AI-Driven Analysis of YouTube Comments to Reveal Election Issue Salience in 2024
Raisa M. Simoes, Timoteo Kelly, Eduardo J. Simoes et al. (+1 more)
This paper aims to explore two competing data science methodologies to attempt answering the question, "Which issues contributed most to voters' choice in the 2024 presidential election?" The methodologies involve novel empirical evidence driven by artificial intelligence (AI) techniques. By using two distinct methods based on natural
language processing and clustering analysis to mine over eight thousand user comments on election-related YouTube
videos from one right leaning journal, Wall Street Journal, and one left leaning journal, New York Times, during pre-election week, we quantify the frequency of selected issue areas among user comments to infer which issues were most salient to potential voters in the seven days preceding the November 5th election. Empirically, we primarily demonstrate that immigration and democracy were the most frequently and consistently invoked issues in user comments on the analyzed YouTube
videos, followed by the issue of identity politics, while inflation was significantly less frequently referenced. These results corroborate certain findings of post-election surveys but also refute the supposed importance of inflation as an election issue. This indicates that variations on opinion mining, with their analysis of raw user data online, can be more revealing than polling and surveys for analyzing election outcomes.
Read the full abstract ▾
All authors Raisa M. Simoes, Timoteo Kelly, Eduardo J. Simoes, Praveen Rao
Computation and Language
Multilingual Generative Retrieval via Cross-lingual Semantic Compression
Yuxin Huang, Simeng Wu, Ran Song et al. (+4 more)
Generative Information
Retrieval is an emerging
retrieval paradigm that exhibits remarkable performance in monolingual scenarios.However, applying these methods to
multilingual retrieval still encounters two primary challenges, cross-lingual identifier misalignment and identifier inflation. To address these limitations, we propose
Multilingual Generative
Retrieval via Cross-lingual Semantic Compression (MGR-CSC), a novel framework that unifies semantically equivalent
multilingual keywords into shared atoms to align semantics and compresses the identifier space, and we propose a dynamic multi-step constrained decoding strategy during
retrieval. MGR-CSC improves cross-lingual
alignment by assigning consistent identifiers and enhances decoding efficiency by reducing redundancy. Experiments demonstrate that MGR-CSC achieves outstanding
retrieval accuracy, improving by 6.83% on mMarco100k and 4.77% on mNQ320k, while reducing document identifiers length by 74.51% and 78.2%, respectively.
Read the full abstract ▾
All authors Yuxin Huang, Simeng Wu, Ran Song, Yan Xiang, Yantuan Xian, Shengxiang Gao, Zhengtao Yu
Computation and Language
Dynamic Generation of Multi-LLM Agents Communication Topologies with Graph Diffusion Models
Eric Hanchen Jiang, Mengting Li, Guancheng Wan et al. (+8 more)
The efficiency of multi-
agent systems driven by large
language models (LLMs) largely hinges on their communication
topology. However, designing an optimal
topology is a non-trivial challenge, as it requires balancing competing objectives such as task performance, communication cost, and robustness. Existing frameworks often rely on static or hand-crafted
topologies, which inherently fail to adapt to diverse task requirements, leading to either excessive
token consumption for simple problems or performance bottlenecks for complex ones. To address this challenge, we introduce a novel generative framework called Guided
Topology Diffusion (GTD). Inspired by conditional discrete graph
diffusion models, GTD formulates
topology synthesis as an iterative construction process. At each step, the generation is steered by a lightweight proxy model that predicts multi-objective rewards (e.g., accuracy, utility, cost), enabling real-time, gradient-free optimization towards task-adaptive
topologies. This iterative, guided synthesis process distinguishes GTD from single-step generative frameworks, enabling it to better navigate complex design trade-offs. We validated GTD across multiple
benchmarks, and experiments show that this framework can generate highly task-adaptive, sparse, and efficient communication
topologies, significantly outperforming existing methods in LLM
agent collaboration.
Read the full abstract ▾
All authors Eric Hanchen Jiang, Mengting Li, Guancheng Wan, Sophia Yin, Yuchen Wu, Xiao Liang, Xinfeng Li, Yizhou Sun, Wei Wang, Kai-Wei Chang, Ying Nian Wu
Computation and Language
HiPRAG: Hierarchical Process Rewards for Efficient Agentic Retrieval Augmented Generation
Peilin Wu, Mian Zhang, Kun Wan et al. (+4 more)
Agentic RAG is a powerful technique for incorporating external information that LLMs lack, enabling better problem solving and question answering. However, suboptimal search behaviors exist widely, such as over-search (retrieving information already known) and under-search (failing to search when necessary), which leads to unnecessary overhead and unreliable outputs. Current training methods, which typically rely on outcome-based rewards in a RL framework, lack the fine-grained control needed to address these inefficiencies. To overcome this, we introduce Hierarchical Process Rewards for Efficient agentic RAG (HiPRAG), a training methodology that incorporates a fine-grained, knowledge-grounded process reward into the RL training. Our approach evaluates the necessity of each search decision on-the-fly by decomposing the
agent's
reasoning trajectory into discrete, parsable steps. We then apply a hierarchical reward function that provides an additional bonus based on the proportion of optimal search and non-search steps, on top of commonly used outcome and format rewards. Experiments on the Qwen2.5 and Llama-3.2 models across seven diverse QA
benchmarks show that our method achieves average accuracies of 65.4% (3B) and 67.2% (7B). This is accomplished while improving search efficiency, reducing the over-search rate to just 2.3% and concurrently lowering the under-search rate. These results demonstrate the efficacy of optimizing the
reasoning process itself, not just the final outcome. Further experiments and analysis demonstrate that HiPRAG shows good generalizability across a wide range of RL
algorithms, model families, sizes, and types. This work demonstrates the importance and potential of fine-grained control through RL, for improving the efficiency and optimality of
reasoning for search
agents.
Read the full abstract ▾
All authors Peilin Wu, Mian Zhang, Kun Wan, Wentian Zhao, Kaiyu He, Xinya Du, Zhiyu Chen
Computation and Language
LLM4Cell: A Survey of Large Language and Agentic Models for Single-Cell Biology
Sajib Acharjee Dip, Adrika Zafor, Bikash Kumar Paul et al. (+4 more)
Large
language models (LLMs) and emerging agentic frameworks are beginning to transform single-
cell biology by enabling natural-
language reasoning, generative annotation, and
multimodal data integration. However, progress remains fragmented across data modalities, architectures, and evaluation standards. LLM4
Cell presents the first unified survey of 58 foundation and agentic models developed for single-
cell research, spanning RNA, ATAC, multi-omic, and spatial modalities. We categorize these methods into five families-foundation, text-bridge, spatial,
multimodal, epigenomic, and agentic-and map them to eight key analytical tasks including annotation, trajectory and perturbation modeling, and drug-response prediction. Drawing on over 40 public
datasets, we analyze
benchmark suitability, data diversity, and ethical or scalability constraints, and evaluate models across 10 domain dimensions covering biological grounding, multi-omics
alignment,
fairness,
privacy, and explainability. By linking
datasets, models, and evaluation domains, LLM4
Cell provides the first integrated view of
language-driven single-
cell intelligence and outlines open challenges in interpretability, standardization, and trustworthy model development.
Read the full abstract ▾
All authors Sajib Acharjee Dip, Adrika Zafor, Bikash Kumar Paul, Uddip Acharjee Shuvo, Muhit Islam Emon, Xuan Wang, Liqing Zhang
Computation and Language
RCPU: Rotation-Constrained Error Compensation for Structured Pruning of Large Language Models
Shuichiro Haruta, Kazunori Matsumoto, Zhi Li et al. (+2 more)
In this paper, we propose a rotation-constrained compensation method to address the errors introduced by structured
pruning of large
language models (LLMs). LLMs are trained on massive
datasets and accumulate rich semantic knowledge in their representation space. In contrast,
pruning is typically carried out with only a small amount of calibration data, which makes output mismatches unavoidable. Although direct least-squares fitting can reduce such errors, it tends to overfit to the limited calibration set, destructively modifying pretrained weights. To overcome this difficulty, we update the pruned parameters under a rotation constraint. This constrained update preserves the geometry of output representations (i.e., norms and inner products) and simultaneously re-aligns the pruned subspace with the original outputs. Furthermore, in rotation-constrained compensation, removing components that strongly contribute to the principal directions of the output makes error recovery difficult. Since input dimensions with large variance strongly affect these principal directions, we design a variance-aware importance score that ensures such dimensions are preferentially kept in the pruned model. By combining this scoring rule with rotation-constrained updates, the proposed method effectively compensates errors while retaining the components likely to be more important in a geometry-preserving manner. In the experiments, we apply the proposed method to Llama-7B and Llama-2-13B, and evaluate it on WikiText2 and multiple
language understanding
benchmarks. The results demonstrate consistently better perplexity and task accuracy compared with existing baselines.
Read the full abstract ▾
All authors Shuichiro Haruta, Kazunori Matsumoto, Zhi Li, Yanan Wang, Mori Kurokawa
Computation and Language
Drift No More? Context Equilibria in Multi-Turn LLM Interactions
Vardhan Dongre, Ryan A. Rossi, Viet Dac Lai et al. (+3 more)
Large
Language Models (LLMs) excel at single-turn tasks such as instruction following and summarization, yet real-world deployments require sustained multi-turn interactions where user goals and conversational context persist and evolve. A recurring challenge in this setting is context drift: the gradual divergence of a model's outputs from goal-consistent behavior across turns. Unlike single-turn errors, drift unfolds temporally and is poorly captured by static evaluation metrics. In this work, we present a study of context drift in multi-turn interactions and propose a simple dynamical framework to interpret its behavior. We formalize drift as the turn-wise KL divergence between the
token-level predictive distributions of the test model and a goal-consistent reference model, and propose a recurrence model that interprets its
evolution as a bounded
stochastic process with restoring forces and controllable interventions. We instantiate this framework in both synthetic long-horizon rewriting tasks and realistic user-
agent simulations such as in $\tau$-Bench, measuring drift for several open-weight LLMs that are used as user simulators. Our experiments consistently reveal stable, noise-limited
equilibria rather than runaway degradation, and demonstrate that simple reminder interventions reliably reduce divergence in line with theoretical predictions. Together, these results suggest that multi-turn drift can be understood as a controllable
equilibrium phenomenon rather than as inevitable decay, providing a foundation for studying and mitigating context drift in extended interactions.
Read the full abstract ▾
All authors Vardhan Dongre, Ryan A. Rossi, Viet Dac Lai, David Seunghyun Yoon, Dilek Hakkani-Tür, Trung Bui
Computation and Language
Instance Relation Learning Network with Label Knowledge Propagation for Few-shot Multi-label Intent Detection
Shiman Zhao, Shangyuan Li, Wei Chen et al. (+4 more)
Few-shot Multi-label Intent Detection (MID) is crucial for dialogue systems, aiming to detect multiple intents of utterances in low-resource dialogue domains. Previous studies focus on a two-stage pipeline. They first learn representations of utterances with multiple labels and then use a threshold-based strategy to identify multi-label results. However, these methods rely on representation classification and ignore instance relations, leading to error propagation. To solve the above issues, we propose a multi-label joint learning method for few-shot MID in an end-to-end manner, which constructs an instance relation learning network with label knowledge propagation to eliminate error propagation. Concretely, we learn the interaction relations between instances with class information to propagate label knowledge between a few labeled (support set) and unlabeled (query set) instances. With label knowledge propagation, the relation strength between instances directly indicates whether two utterances belong to the same intent for multi-label prediction. Besides, a dual relation-enhanced loss is developed to optimize support- and query-level relation strength to improve performance. Experiments show that we outperform strong baselines by an average of 9.54% AUC and 11.19% Macro-F1 in 1-shot scenarios.
Read the full abstract ▾
All authors Shiman Zhao, Shangyuan Li, Wei Chen, Tengjiao Wang, Jiahui Yao, Jiabin Zheng, Kam Fai Wong
Computation and Language
The Unintended Trade-off of AI Alignment:Balancing Hallucination Mitigation and Safety in LLMs
Omar Mahmoud, Ali Khalil, Buddhika Laknath Semage et al. (+2 more)
Hallucination in large
language models (LLMs) has been widely studied in recent years, with progress in both detection and mitigation aimed at improving truthfulness. Yet, a critical side effect remains largely overlooked: enhancing truthfulness can negatively impact safety
alignment. In this paper, we investigate this trade-off and show that increasing factual accuracy often comes at the cost of weakened refusal behavior. Our analysis reveals that this arises from overlapping components in the model that simultaneously encode
hallucination and refusal information, leading
alignment methods to suppress factual knowledge unintentionally. We further examine how fine-tuning on benign
datasets, even when curated for safety, can degrade
alignment for the same reason. To address this, we propose a method that disentangles refusal-related features from
hallucination features using sparse autoencoders, and preserves refusal behavior during fine-tuning through subspace orthogonalization. This approach prevents
hallucinations from increasing while maintaining safety
alignment.We evaluate our method on commonsense
reasoning tasks and harmful
benchmarks (AdvBench and StrongReject). Results demonstrate that our approach preserves refusal behavior and task utility, mitigating the trade-off between truthfulness and safety.
Read the full abstract ▾
All authors Omar Mahmoud, Ali Khalil, Buddhika Laknath Semage, Thommen George Karimpanal, Santu Rana
Computation and Language
Curing Miracle Steps in LLM Mathematical Reasoning with Rubric Rewards
Youliang Yuan, Qiuyang Mang, Jingbang Chen et al. (+7 more)
In this paper, we observe that current models are susceptible to reward hacking, leading to a substantial overestimation of a model's
reasoning ability. This is evidenced by a high incidence of false positives-solutions that reach the correct answer through an unsound process. Through a systematic analysis with human verification, we establish a taxonomy of these failure modes, identifying patterns like Miracle Steps-abrupt jumps to a correct output without a valid preceding derivation. Probing experiments suggest that these Miracle Steps are linked to answer-recall shortcuts, including memorization from pretraining, where the model accesses the correct answer independently of its
reasoning chain. To mitigate this systemic issue, we introduce the Rubric Reward Model (RRM), a process-oriented reward function that evaluates the entire
reasoning trajectory against problem-specific rubrics. The RRM explicitly penalizes logical flaws and encourages rigorous deduction. When integrated into an RL pipeline, RRM-based training consistently outperforms outcome-only supervision across four math
benchmarks. Notably, it boosts Verified Pass@1024 on AIME2024 from 26.7% to 62.6% and reduces the incidence of Miracle Steps by 71%. Our work demonstrates that rewarding the solution process is crucial for building accurate and reliable models.
Read the full abstract ▾
All authors Youliang Yuan, Qiuyang Mang, Jingbang Chen, Hong Wan, Xiaoyuan Liu, Junjielong Xu, Jen-tse Huang, Wenxuan Wang, Wenxiang Jiao, Pinjia He
Computation and Language
ToolLibGen: Scalable Automatic Tool Creation and Aggregation for LLM Reasoning
Murong Yue, Zhiwei Liu, Liangwei Yang et al. (+8 more)
Large
Language Models (LLMs) equipped with external tools have demonstrated enhanced performance on complex
reasoning tasks. The widespread adoption of this tool-augmented
reasoning is hindered by the scarcity of domain-specific tools. For instance, in domains such as physics question answering, suitable and specialized tools are often missing. Recent work has explored automating tool creation by extracting reusable functions from Chain-of-Thought (CoT)
reasoning traces; however, these approaches face a critical scalability bottleneck. As the number of generated tools grows, storing them in an unstructured collection leads to significant
retrieval challenges, including an expanding search space and ambiguity between function-related tools. To address this, we propose a systematic approach to automatically refactor an unstructured collection of tools into a structured tool library. Our system first generates discrete, task-specific tools and clusters them into semantically coherent topics. Within each cluster, we introduce a multi-
agent framework to consolidate scattered functionalities: a code
agent refactors code to extract shared logic and creates versatile, aggregated tools, while a reviewing
agent ensures that these aggregated tools maintain the complete functional capabilities of the original set. This process transforms numerous question-specific tools into a smaller set of powerful, aggregated tools without loss of functionality. Experimental results demonstrate that our approach significantly improves tool
retrieval accuracy and overall
reasoning performance across multiple
reasoning tasks. Furthermore, our method shows enhanced scalability compared with baselines as the number of question-specific increases.
Read the full abstract ▾
All authors Murong Yue, Zhiwei Liu, Liangwei Yang, Jianguo Zhang, Zuxin Liu, Haolin Chen, Ziyu Yao, Silvio Savarese, Caiming Xiong, Shelby Heinecke, Huan Wang
Computation and Language
Test-Time Reasoners Are Strategic Multiple-Choice Test-Takers
Nishant Balepur, Atrey Desai and Rachel Rudinger
Large
language models (LLMs) now give
reasoning before answering, excelling in tasks like multiple-choice question answering (MCQA). Yet, a concern is that LLMs do not solve MCQs as intended, as work finds LLMs sans
reasoning succeed in MCQA without using the question, i.e., choices-only. Such partial-input success is often linked to trivial shortcuts, but
reasoning traces could reveal if choices-only strategies are truly shallow. To examine these strategies, we have
reasoning LLMs solve MCQs in full and choices-only inputs; test-time
reasoning often boosts accuracy in full and in choices-only, half the time. While possibly due to shallow shortcuts, choices-only success is barely affected by the length of
reasoning traces, and after finding traces pass faithfulness tests, we show they use less problematic strategies like inferring missing questions. In all, we challenge claims that partial-input success is always a flaw, so we propose how
reasoning traces could separate problematic data from less problematic
reasoning.
Read the full abstract ▾
Computation and Language
Parallel Test-Time Scaling for Latent Reasoning Models
Runyang You, Yongqi Li, Meng Liu et al. (+3 more)
Parallel test-time scaling (TTS) is a pivotal approach for enhancing large
language models (LLMs), typically by sampling multiple
token-based chains-of-thought in parallel and aggregating outcomes through voting or search. Recent advances in latent
reasoning, where intermediate
reasoning unfolds in continuous vector spaces, offer a more efficient alternative to explicit Chain-of-Thought, yet whether such latent models can similarly benefit from parallel TTS remains open, mainly due to the absence of sampling mechanisms in continuous space, and the lack of probabilistic signals for advanced trajectory aggregation. This work enables parallel TTS for latent
reasoning models by addressing the above issues. For sampling, we introduce two uncertainty-inspired
stochastic strategies: Monte Carlo Dropout and Additive Gaussian Noise. For aggregation, we design a Latent Reward Model (LatentRM) trained with step-wise contrastive objective to score and guide latent
reasoning. Extensive experiments and visualization analyses show that both sampling strategies scale effectively with compute and exhibit distinct exploration dynamics, while LatentRM enables effective trajectory selection. Together, our explorations open a new direction for scalable inference in continuous spaces. Code and checkpoints released at https://github.com/ModalityDance/LatentTTS
Read the full abstract ▾
All authors Runyang You, Yongqi Li, Meng Liu, Wenjie Wang, Liqiang Nie, Wenjie Li
Computation and Language
OpenRubrics: Towards Scalable Synthetic Rubric Generation for Reward Modeling and LLM Alignment
Tianci Liu, Ran Xu, Tony Yu et al. (+4 more)
Reward modeling lies at the core of reinforcement learning from human feedback (RLHF), yet most existing reward models rely on scalar or pairwise judgments that fail to capture the multifaceted nature of human preferences. Recent studies have explored rubrics-as-rewards (RaR) that uses structured criteria to capture multiple dimensions of response quality. However, producing rubrics that are both reliable and scalable remains a key challenge. In this work, we introduce OpenRubrics, a diverse, large-scale collection of (
prompt, rubric) pairs for training rubric-generation and rubric-based reward models. To elicit discriminative and comprehensive evaluation signals, we introduce Contrastive Rubric Generation (CRG), which derives both hard rules (explicit constraints) and principles (implicit qualities) by contrasting preferred and rejected responses. We further remove noisy rubrics via preserving preference-label consistency. Across multiple reward-modeling
benchmarks, our rubric-based reward model, Rubric-RM, surpasses strong size-matched baselines by 8.4%. These gains transfer to policy models on instruction-following and biomedical
benchmarks.
Read the full abstract ▾
All authors Tianci Liu, Ran Xu, Tony Yu, Ilgee Hong, Carl Yang, Tuo Zhao, Haoyu Wang
Computation and Language
ToolExpander: Extending the Frontiers of Tool-Using Reinforcement Learning to Weak LLMs
Fu Chen, Peng Wang, Xiyin Li et al. (+3 more)
Training Large
Language Models (LLMs) with Group Relative Policy Optimization (GRPO) encounters a significant challenge: models often fail to produce accurate responses, particularly in small-scale architectures. This limitation not only diminishes performance improvements and undermines the potential of GRPO but also frequently leads to mid-training collapse, adversely affecting stability and final efficacy. To address these issues, we propose ToolExpander, a novel framework that advances tool-oriented reinforcement learning for resource-constrained LLMs through two key innovations:(1) Dynamic Multi-Round Hard Sampling, which dynamically substitutes challenging samples(those without correct outputs over 10 rollouts) with high-quality few-shot demonstrations during training, coupled with an exponential learning rate decay strategy to mitigate oscillations;(2) Self-Exemplifying Thinking, an enhanced GRPO framework that eliminates KL divergence and incorporates adjusted clipping coefficients, encouraging models to autonomously generate and analyze few-shot examples via a minimal additional reward (0.01).Experimental results demonstrate that ToolExpander significantly enhances tool-using capabilities in LLMs, especially in weaker small-scale models, improving both training stability and overall performance.
Read the full abstract ▾
All authors Fu Chen, Peng Wang, Xiyin Li, Wen Li, Shichi Lei, Dongdong Xiang
Computation and Language
Multilingual Knowledge Graph Completion via Efficient Multilingual Knowledge Sharing
Cunli Mao, Xiaofei Gao, Ran Song et al. (+4 more)
Large
language models (LLMs) based
Multilingual Knowledge Graph Completion (MKGC) aim to predict missing facts by leveraging LLMs'
multilingual understanding capabilities, improving the completeness of
multilingual knowledge graphs (KGs). However, existing MKGC research underutilizes the
multilingual capabilities of LLMs and ignores the shareability of cross-lingual knowledge. In this paper, we propose a novel MKGC framework that leverages
multilingual shared knowledge to significantly enhance performance through two components: Knowledge-level Grouped Mixture of Experts (KL-GMoE) and Iterative Entity Reranking (IER). KL-GMoE efficiently models shared knowledge, while IER significantly enhances its utilization. To evaluate our framework, we constructed a mKG
dataset containing 5
languages and conducted comprehensive comparative experiments with existing state-of-the-art (SOTA) MKGC method. The experimental results demonstrate that our framework achieves improvements of 5.47%, 3.27%, and 1.01% in the Hits@1, Hits@3, and Hits@10 metrics, respectively, compared with SOTA MKGC method. Further experimental analysis revealed the properties of knowledge sharing in settings of unseen and unbalanced
languages. We have released the
dataset and code for our work on https://github.com/gaoxiaofei07/KL-GMoE.
Read the full abstract ▾
All authors Cunli Mao, Xiaofei Gao, Ran Song, Shizhu He, Shengxiang Gao, Kang Liu, Zhengtao Yu
Computation and Language · Cross-listed from cs.AI
oMeBench: Towards Robust Benchmarking of LLMs in Organic Mechanism Elucidation and Reasoning
Ruiling Xu and Yifan Zhang
Organic reaction mechanisms describe the step-wise elementary processes by which reactants transform into intermediates and products, and are fundamental to understanding chemical reactivity and guiding molecular and reaction de-sign. While large
language models (LLMs) have shown promise on chemical tasks such as synthesis design, it remains unclear to what extent this reflects genuine chemical
reasoning capabilities: the ability to generate chemically valid intermediates, maintain consistency across reaction steps, and follow logically coherent multi-step pathways. To investigate this, we introduce oMeBench, the first large-scale, expert-curated
benchmark for organic mechanism
reasoning, comprising over 10,000 annotated mechanistic steps with reaction type labels, intermediate structures, and difficulty ratings. To enable fine-grained evaluation, we further propose oMeS, a dynamic scoring framework that jointly assesses step-level logical consistency and chemical structural similarity. Systematic evaluation of state-of-the-art LLMs reveals that while current models exhibit promising chemical intuition, they struggle to produce correct and consistent
reasoning across multi-step mechanisms. Notably, combining prompting strategies with fine-tuning enables smaller-scale models to achieve performance comparable to closed-source frontier models. We hope oMeBench will serve as a rigorous foundation for advancing AI systems toward genuine chemical
reasoning.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.IR
Who Stole Your Data? A Method for Detecting Unauthorized RAG Theft
Peiyang Liu, Ziqiang Cui, Di Liang et al. (+1 more)
Retrieval-augmented generation (RAG) enhances Large
Language Models (LLMs) by mitigating
hallucinations and outdated information issues, yet simultaneously facilitates unauthorized data appropriation at scale. This paper addresses this challenge through two key contributions. First, we introduce RPD, a novel
dataset specifically designed for RAG plagiarism detection that encompasses diverse professional domains and writing styles, overcoming limitations in existing resources. Second, we develop a dual-layered watermarking system that embeds protection at both semantic and lexical levels, complemented by an interrogator-detective framework that employs statistical hypothesis testing on accumulated evidence. Extensive experimentation demonstrates our approach's effectiveness across varying query volumes, defense
prompts, and
retrieval parameters, while maintaining resilience against
adversarial evasion techniques. This work establishes a foundational framework for intellectual property protection in
retrieval-augmented AI systems.
Read the full abstract ▾
All authors Peiyang Liu, Ziqiang Cui, Di Liang, Wei Ye
Computation and Language
SUBQRAG: Sub-Question Driven Dynamic Graph RAG
Jiaoyang Li, Junhao Ruan, Shengwei Tang et al. (+5 more)
Graph
Retrieval-Augmented Generation (Graph RAG) effectively builds a knowledge graph (KG) to connect disparate facts across a large document corpus. However, this broad-view approach often lacks the deep structured
reasoning needed for complex multi-hop question answering (QA), leading to incomplete evidence and error accumulation. To address these limitations, we propose SubQRAG, a sub-question-driven framework that enhances
reasoning depth. SubQRAG decomposes a complex question into an ordered chain of verifiable sub-questions. For each sub-question, it retrieves relevant triples from the graph. When the existing graph is insufficient, the system dynamically expands it by extracting new triples from source documents in real time. All triples used in the
reasoning process are aggregated into a "graph
memory," forming a structured and traceable evidence path for final answer generation. Experiments on three multi-hop QA
benchmarks demonstrate that SubQRAG achieves consistent and significant improvements, especially in Exact Match scores.
Read the full abstract ▾
All authors Jiaoyang Li, Junhao Ruan, Shengwei Tang, Saihan Chen, Kaiyan Chang, Yuan Ge, Tong Xiao, Jingbo Zhu
Computation and Language
MemWeaver: A Hierarchical Memory from Textual Interactive Behaviors for Personalized Generation
Shuo Yu, Mingyue Cheng, Daoyu Wang et al. (+4 more)
The primary form of user-internet engagement is shifting from leveraging implicit feedback signals, such as browsing and clicks, to harnessing the rich explicit feedback provided by textual interactive behaviors. This shift unlocks a rich source of user textual history, presenting a profound opportunity for a deeper form of personalization. However, prevailing approaches offer only a shallow form of personalization, as they treat user history as a flat list of texts for
retrieval and fail to model the rich temporal and semantic structures reflecting dynamic nature of user interests. In this work, we propose MemWeaver, a framework that weaves the user's entire textual history into a hierarchical
memory to power deeply personalized generation. The core innovation of our
memory lies in its ability to capture both the temporal
evolution of interests and the semantic relationships between different activities. To achieve this, MemWeaver builds two complementary
memory components that both integrate temporal and semantic information, but at different levels of abstraction: behavioral
memory, which captures specific user actions, and cognitive
memory, which represents long-term preferences. This dual-component
memory serves as a comprehensive representation of the user, allowing large
language models (LLMs) to reason over both concrete behaviors and abstracted cognitive traits. This leads to content generation that is deeply aligned with their latent preferences. Experiments on the six
datasets of the
Language Model Personalization (LaMP)
benchmark validate the efficacy of MemWeaver. Our code is available.
Read the full abstract ▾
All authors Shuo Yu, Mingyue Cheng, Daoyu Wang, Qi Liu, Zirui Liu, Ze Guo, Xiaoyu Tao
Computation and Language · Cross-listed from cs.AI
Multimodal Safety Evaluation in Generative Agent Social Simulations
Alhim Vera, Karen Sanchez, Carlos Hinojosa et al. (+3 more)
Can generative
agents be trusted in
multimodal environments? Despite advances in large
language and
vision-
language models that enable
agents to act autonomously and pursue goals in rich settings, their ability to reason about safety, coherence, and trust across modalities remains limited. We introduce a reproducible simulation framework for evaluating
agents along three dimensions: (1) safety improvement over time, including iterative plan revisions in text-visual scenarios; (2) detection of unsafe activities across multiple categories of social situations; and (3) social dynamics, measured as interaction counts and acceptance ratios of social exchanges.
Agents are equipped with layered
memory, dynamic planning,
multimodal perception, and are instrumented with SocialMetrics, a suite of behavioral and structural metrics that quantifies plan revisions, unsafe-to-safe conversions, and information
diffusion across networks. Experiments show that while
agents can detect direct
multimodal contradictions, they often fail to align local revisions with global safety, reaching only a 55 percent success rate in correcting unsafe plans. Across eight simulation runs with three models - Claude, GPT-4o mini, and Qwen-VL - five
agents achieved average unsafe-to-safe conversion rates of 75, 55, and 58 percent, respectively. Overall performance ranged from 20 percent in multi-risk scenarios with GPT-4o mini to 98 percent in localized contexts such as fire/heat with Claude. Notably, 45 percent of unsafe actions were accepted when paired with misleading visuals, showing a strong tendency to overtrust images. These findings expose critical limitations in current architectures and provide a reproducible platform for studying
multimodal safety, coherence, and social dynamics.
Read the full abstract ▾
All authors Alhim Vera, Karen Sanchez, Carlos Hinojosa, Haidar Bin Hamid, Donghoon Kim, Bernard Ghanem