论文检索

Kaiyu He, Mian Zhang, Peilin Wu, Xinya Du, Zhiyu Chen

While Large Language Models (LLMs) excel at factual retrieval, they often struggle with the "curse of two-hop reasoning" in compositional tasks. Recent research suggests that parameter-sharing transformers can bridge this gap by forming a "Generalization Circuit" during a prolonged "grokking" phase. A fundamental question arises: Is a grokked model truly superior to its non-grokked counterparts? Furthermore, is the extensive computational cost of waiting for the grokking phase worthwhile? In this work, we conduct a mechanistic study to evaluate the Generalization Circuit’s role in knowledge assimilation and transfer. We demonstrate that: (i) The inference paths established by non-grokked and grokked models for in-distribution compositional queries are identical. This suggests that the "Generalization Circuit" does not represent the sudden acquisition of a new reasoning paradigm. Instead, we argue that grokking is the process of integrating memorized atomic facts into an easy-acquire, naturally established reasoning path. (ii) Achieving high accuracy on unseen cases after prolonged training and the formation of a certain reasoning path are not bound; they can occur independently under specific data regimes. (iii) Even a mature circuit exhibits limited transferability when integrating new knowledge, suggesting that "grokked" Transformers do not achieve a full mastery of compositional logic.

Sihong Wu, Yiling Ma, Yilun Zhao, Tiansheng Hu, Owen Jiang, Manasi Patwardhan, Arman Cohan

Large language models (LLMs) are increasingly used across the scientific workflow, including to draft peer-review reports. However, many AI-generated reviews are superficial and insufficiently actionable, leaving authors without concrete, implementable guidance and motivating the gap this work addresses. We propose RbtAct, which targets actionable review feedback generation and places existing peer review rebuttal at the center of learning. Rebuttals show which reviewer comments led to concrete revisions or specific plans, and which were only defended. Building on this insight, we leverage rebuttal as implicit supervision to directly optimize a feedback generator for actionability. To support this objective, we propose a new task called perspective-conditioned segment-level review feedback generation, in which the model is required to produce a single focused comment based on the complete paper and a specified perspective such as experiments and writing. We also build a large dataset named RMR-75K that maps review segments to the rebuttal segments that address them, with perspective labels and impact categories that order author uptake. We then train the Llama-3.1-8B-Instruct model with supervised fine-tuning on review segments followed by preference optimization using rebuttal derived pairs. Experiments with human experts and LLM-as-a-judge show consistent gains in actionability and specificity over strong baselines while maintaining grounding and relevance.

Jisu Shin, Hoyun Song, Juhyun Oh, Changgeon Ko, Eunsu Kim, Chani Jung, Alice Oh

People often encounter role conflicts—social dilemmas where the expectations of multiple roles clash and cannot be simultaneously fulfilled. As large language models (LLMs) increasingly navigate these social dynamics, a critical research question emerges. When faced with such dilemmas, do LLMs prioritize dynamic contextual cues or the learned preferences? To address this, we introduce RoleConflictBench, a novel benchmark designed to measure the contextual sensitivity of LLMs in role conflict scenarios. To enable objective evaluation within this subjective domain, we employ situational urgency as a constraint for decision-making. We construct the dataset through a three-stage pipeline that generates over 13,000 realistic scenarios across 65 roles in five social domains by systematically varying the urgency of competing situations. This controlled setup enables us to quantitatively measure contextual sensitivity, determining whether model decisions align with the situational contexts or are overridden by the learned role preferences. Our analysis of 10 LLMs reveals that models substantially deviate from this objective baseline. Instead of responding to dynamic contextual cues, their decisions are predominantly governed by the preferences toward specific social roles.

Xiaochen Zhu, Caiqi Zhang, Yizhou Chi, Tom Stafford, Nigel Collier, Andreas Vlachos

Multi-agent debate (MAD) is widely used to improve large language model (LLM) performance through test-time scaling, yet recent work shows that vanilla MAD often underperforms simple majority vote despite higher computational cost. Studies show that, under homogeneous agents and uniform belief updates, debate preserves expected correctness and therefore cannot reliably improve outcomes. Drawing on findings from human deliberation and collective decision-making, we identify two key mechanisms missing from vanilla MAD: (i) diversity of initial viewpoints and (ii) explicit, calibrated confidence communication. We propose two lightweight interventions. First, a diversity-aware initialisation that selects a more diverse pool of candidate answers, increasing the likelihood that a correct hypothesis is present at the start of debate. Second, a confidence-modulated debate protocol in which agents express calibrated confidence and condition their updates on others’ confidence. We show theoretically that diversity-aware initialisation improves the prior probability of MAD success without changing the underlying update dynamics, while confidence-modulated updates enable debate to systematically drift to the correct hypothesis. Empirically, across six reasoning-oriented QA benchmarks, our methods consistently outperform vanilla MAD and majority vote. Our results connect human deliberation with LLM-based debate and demonstrate that simple, principled modifications can substantially enhance debate effectiveness.

Changjun Park, Sejong Yoon, Jaekwang Kim

We propose DPLoRA (Dual-Pruning Low-Rank Adaptation), a framework that optimizes rank allocation via two stages: (1) an initial pruning stage (OPLoRA; Optimal Pruning LoRA) that uses Integer Linear Programming (ILP) to determine optimal layer-wise ranks without manual tuning; and (2) a progressive pruning stage that further reduces ranks adaptively during training using importance scores smoothed by Exponential Moving Average (EMA). Experiments demonstrate that OPLoRA consistently outperforms existing PEFT baselines on GLUE and instruction-following tasks, while the full DPLoRA framework establishes a new state-of-the-art among compared PEFT baselines on GLUE at high-performance settings (p=0.4). At efficiency-focused settings (p=0.8), our method reduces trainable parameters by over 80% and training time by 46% compared to standard LoRA, offering a highly efficient solution for deploying large-scale models in resource-constrained environments.

Shuyang Zhang, Zhixuan Liu, Zhichen Dong, Hao Zhang, Chaochao Lu, Chao Yang

Prompt optimizers are widely used to create high-quality prompts for Large Language Models (LLMs), but their effectiveness remains unstable in practice. This instability is caused by the misalignment between conservative needs (e.g., safety compliance) and open-ended goals (e.g., creative writing). To address this, we propose a semantic-entropy-based method, using task uncertainty to guide prompt optimization. Specifically, we measure the task’s uncertainty level with pre-defined templates, then use this measure to direct prompt optimization: selecting high-entropy prompt candidates for creative tasks and low-entropy candidates for conservative ones. Extensive experiments across various model families demonstrate that our method consistently outperforms baselines by effectively adjusting entropy levels. Our approach requires no training, works with black-box models, and integrates easily into existing prompt optimizers.

Zoya Volovikova, Nikita Sorokin, Dmitriy Lukashevskiy, Aleksandr Panov, Alexey Skrynnik

We introduce SuperIgor, a framework for instruction-following tasks. Unlike prior methods that rely on predefined subtasks, SuperIgor enables a language model to generate and refine high-level plans through a self-learning mechanism, reducing the need for manual dataset annotation. Our approach involves iterative co-training: an RL agent is trained to follow the generated plans, while the language model adapts and modifies these plans based on RL feedback and preferences. This creates a feedback loop where both the agent and the planner improve jointly. We validate our framework in environments with rich dynamics and stochasticity. Results show that SuperIgor agents adhere to instructions more strictly than baseline methods, while also demonstrating strong generalization to previously unseen instructions.

Shuqi Cao, Jingyi He, Fei Tan

Long-term conversational large language model (LLM) agents require memory systems that can recover relevant evidence from historical interactions without overwhelming the answer stage with irrelevant context. However, existing memory systems, including hierarchical ones, still often rely solely on vector similarity for retrieval. It tends to produce bloated evidence sets: adding many superficially similar dialogue turns yields little additional recall, but lowers retrieval precision, increases answer-stage context cost, and makes retrieved memories harder to inspect and manage. To address this, we propose HiGMem (Hierarchical and LLM-Guided Memory System), a two-level event-turn memory system that allows LLMs to use event summaries as semantic anchors to predict which related turns are worth reading. This allows the model to inspect high-level event summaries first and then focus on a smaller set of potentially useful turns, providing a concise and reliable evidence set through reasoning, while avoiding the retrieval overhead that would be excessively high compared to vector retrieval.On the LoCoMo10 benchmark, HiGMem achieves the best F1 on four of five question categories and improves adversarial F1 from 0.54 to 0.78 over A-Mem, while retrieving an order of magnitude fewer turns. Code is publicly available at https://github.com/ZeroLoss-Lab/HiGMem.

Akram Elbouanani, Aboubacar Tuo, Adrian Popescu

Existing approaches to bias evaluation in large language models (LLMs) trade ecological validity for statistical control, relying on artificial prompts that poorly reflect real-world use, or on naturalistic tasks that lack scale and rigor. We introduce a scalable bias-auditing framework using named entities as probes to measure structural disparities in model behavior. We show that synthetic data reliably reproduces bias patterns observed in natural text, enabling large-scale analysis. Using this approach, we conduct the largest bias audit to date, comprising 1.9 billion data points across multiple entity types, tasks, languages, models, and prompting strategies. Our results reveal systematic biases: models penalize right-wing politicians, favor left-wing politicians, prefer Western and wealthy nations over the Global South, favor Western companies, and penalize firms in the defense and pharmaceutical sectors. While instruction tuning reduces bias, increasing model scale amplifies it, and prompting in Chinese or Russian does not attenuate Western-aligned preferences. These results indicate that LLMs should undergo rigorous auditing before deployment in high-stakes applications.

Hongzhi Zhang, Yuanze Hu, Tinghai Zhang, Jia Fu, Tao Wang, Junwei Jing, Zhaoxin Fan, Wei Bi, Ruiming Tang, Han Li 等

The evolution of Large Language Models (LLMs) towards autonomous agents has catalyzed progress in Deep Research. While retrieval capabilities are well-benchmarked, the post-retrieval synthesis stage—where agents must digest massive amounts of context and consolidate fragmented evidence into coherent, long-form reports—remains under-evaluated due to the subjectivity of open-ended writing.To bridge this gap, we introduce DeepSynth-Eval, a benchmark designed to objectively evaluate information consolidation capabilities. We leverage high-quality survey papers as gold standards, reverse-engineer research requests, and construct Oracle Contexts from their bibliographies to isolate synthesis from retrieval noise. We propose a fine-grained evaluation protocol using General Checklists (for factual coverage) and Constraint Checklists (for structural organization), transforming subjective judgment into verifiable metrics. Experiments across 96 tasks reveal that synthesizing information from hundreds of references remains a significant challenge. Our results demonstrate that agentic "plan-then-write" workflows significantly outperform single-turn generation, effectively reducing hallucinations and improving adherence to complex structural constraints.

Yanshan Liu, Hongbo Zhang, Zhen Sun, Jiaheng Wei, Kaishun Wu

In complex domains like interior design, user requests are often ambiguous and multimodal. Professional designers address this by asking strategic clarification questions based on hierarchical priorities, a capability lacking in current Vision-Language Models (VLMs). When fine-tuned on dialogue data, existing models often exhibit modality forgetting, overfitting to textual patterns while neglecting visual cues and thus producing hallucinated or visually irrelevant questions. To bridge this gap, we introduce VIDA (Visual Intent-driven Design Assistant), an assistant designed to generate proactive, visually grounded, and strategically prioritized clarification questions. Instead of standard fine-tuning, we propose a strategy-aware alignment framework that evolves from imitation learning to value-driven reinforcement. We utilize Group Sequence Policy Optimization to strictly enforce expert protocols, ensuring the model not only mimics fluent speech but also adheres to optimal inquiry strategies. Crucially, we design a novel hierarchical reward mechanism with Dynamic Intent Binding to align the assistant with professional prioritization standards. To facilitate this research, we construct and release InteriorClarify, a multimodal benchmark dataset comprising 1,016 real-world consultation cases annotated with this three-tier intent hierarchy. Extensive experiments demonstrate that VIDA sets a new state-of-the-art, improving the Strategic Alignment Score (SAS) by 20.59% over SFT baselines and effectively restoring visual grounding capabilities lost during standard fine-tuning.

Weiwen SU, Naoki Yoshinaga, Masashi Toyoda

Large language models (LLMs) enable simulating individual responses from persona information, supporting applications such as opinion elicitation and virtual character creation. However, existing approaches typically assume rich persona profiles, which are often unavailable in practice. In this work, motivated by recent findings that LLMs can identify query-relevant persona dimensions (e.g., whether a user is price-sensitive), we study query-focused individual simulation under cold-start settings, where relevant persona information is identified and requested on demand for each query. To solve this task while minimizing the number of persona requests, we explore a progressive method that iteratively predicts the most critical relevant persona dimension and uses self-reported confidence as a stopping signal to determine when sufficient information has been collected. Experiments on two dialogue datasets show that this query-driven paradigm achieves simulation performance comparable to approaches that rely on rich persona information extracted from dialogue history, using only a few persona dimensions (up to five per query), and this number is further reduced by our progressive method while maintaining or improving simulation quality.

Jiaye Lin, Mengdi Li, Xufeng Zhao, Wenhao Lu, Peilin Zhao, Stefan Wermter, Di Wang

Reward models trained through Reinforcement Learning from AI Feedback (RLAIF) methods frequently suffer from limited generalizability, which hinders the alignment performance of policy models. This challenge stems from various issues, including distribution shift, preference label noise, and mismatch of overly challenging samples with model capacity. In this paper, we aim to enhance the generalizability of reward models through a data-centric approach, driven by the insight that these issues are inherently intertwined from a uniform perspective of data difficulty. Accordingly, we propose a novel framework, Curriculum-RLAIF, which constructs preference pairs with varying difficulty levels and then produces a specific curriculum for reward model training. Comprehensive experimental results suggest that reward models trained with Curriculum-RLAIF achieve improved generalizability, boosting the alignment performance of policy models by a significant margin without incurring additional inference costs compared to various existing non-curriculum baselines. Further analysis and comparison with alternative strategies highlight the superiority of Curriculum-RLAIF in simplicity, efficiency, and effectiveness.

Han Liu, Shuotian Ma, Hui Li, Xiaotong Zhang, Fenglong Ma, Hong Yu

Knowledge distillation has emerged as a pivotal paradigm for transferring the superior reasoning capabilities of Large Reasoning Models (LRMs) to efficient student models. However, the raw Chain-of-Thought (CoT) trajectories are often verbose and redundant, which dilutes the underlying logic and hinders effective knowledge distillation for student models. Although recent work has focused on pruning CoT to streamline these reasoning paths, existing local heuristic methods often fail to capture global causal logic due to rigid rules and limited search spaces, while global heuristic approaches incur substantial computational costs. To address these issues, we propose Pru-CoT (Pruning Chain-of-Thought), a framework that aims to extract the essential logical structure from reasoning chains. Pru-CoT implements a step-level importance assessment via global optimization on a frozen student large language model (LLM), quantifying the gradient-based causal contribution of each component. Guided by these important signals, the framework performs fidelity-constrained pruning, utilizing an LLM-driven process to synthesize concise, logically coherent narratives. Extensive experiments on mathematical reasoning benchmarks demonstrate that models trained with Pru-CoT not only achieve superior accuracy but also generate significantly more compact reasoning paths compared to those trained on raw verbose data.

Chuan Li, Ye Lyu, Chengyu Wang, Mingyuan Fan, Cen Chen

Despite the advanced capabilities of Large Language Models (LLMs), training specialized reasoning models for the medical domain remains a significant challenge due to the scarcity of high-quality, large-scale Chain-of-Thought (CoT) data. Moreover, the intermediate reasoning steps in teacher-generated CoT data can be redundant and noisy, leading models to acquire spurious patterns and resulting in suboptimal performance. To address these issues, we propose MedCoach, a novel framework that introduces a dedicated coach role to guide the student model through question decomposition, thereby smoothing its learning curve in medical reasoning. The framework employs a curriculum-oriented warm-up on simplified sub-questions, facilitating domain adaptation before advancing to complex long-chain reasoning. To ensure the fidelity of the intermediate chain-of-thought signals, we augment this phase with medical knowledge graphs to suppress factual drift and mitigate reasoning noise at a granular level.Subsequently, we introduce a targeted factual perturbation mechanism to foster fine-grained discrimination between valid fact utilization and subtle factual misapplications. Extensive experiments across diverse benchmarks demonstrate notable improvements over existing methods, validating the effectiveness of MedCoach.

Yifeng Liu, Siqi Ouyang, Yatish H R, Lei Li

Large Language Models (LLMs) have demonstrated remarkable capability in machine translation on high-resource language pairs, yet their performance on low-resource translation still lags behind. Existing post-training methods rely heavily on high-quality parallel data, which are often scarce or unavailable for low-resource languages. In this paper, we introduce WALAR, a reinforcement training method using only monolingual text to elevate LLMs’ translation capabilities on massive low-resource languages while retaining their performance on high-resource languages. Our key insight is based on the observation of failure modes (or "holes") in existing source-based multilingual quality estimation (QE) models. Reinforcement learning (RL) using these QE models tends to amplify such holes, resulting in poorer multilingual LLMs. We develop techniques including word alignment and language alignment to mitigate such holes in WALAR’s reward for RL training. We continually trained an LLM supporting translation of 101 languages using WALAR. The experiments show that our new model outperforms LLaMAX, one of the strongest open-source multilingual LLMs by a large margin on 1,414 language directions on Flores-101 dataset.

David Hartmann, Lena Pohlmann, Lelia Hanslik, Noah Gießing, Bettina Berendt, Pieter Delobelle

Large Language Models (LLMs) exhibit systematic biases across demographic groups. Auditing is proposed as an accountability tool for black-box LLM applications, but suffers from resource-intensive query access. We conceptualise auditing as uncertainty estimation over a target fairness metric and introduce BAFA, the Bounded Active Fairness Auditor for query-efficient auditing of black-box LLMs. BAFA maintains a version space of surrogate models consistent with queried scores and computes uncertainty intervals for fairness metrics (e.g., \Delta AUC) via constrained empirical risk minimisation. Active query selection narrows these intervals to reduce estimation error. We evaluate BAFA on two standard fairness dataset case studies: CivilComments and Bias-in-Bios, comparing against stratified sampling, power sampling, and ablations. BAFA achieves target error thresholds with up to 40\times fewer queries than stratified sampling (e.g., 144 vs 5,956 queries at \varepsilon=0.02 for CivilComments) for tight thresholds, demonstrates substantially better performance over time, and shows lower variance across runs. These results suggest that active sampling can reduce resources needed for independent fairness auditing with LLMs, supporting continuous model evaluations.

Bhrij Patel, Davide Belli, Amir Jalalirad, Maximilian Arnold, Aleksandr Ermolov, Bence Major

Function calling agents powered by Large Language Models (LLMs) select external tools to automate complex tasks. On-device agents typically use a retrieval module to select relevant tools, improving performance and reducing context length. However, existing retrieval methods rely on static and limited inputs, failing to capture multi-step tool dependencies and evolving task context. This limitation often introduces irrelevant tools that mislead the agent, degrading efficiency and accuracy. We propose Dynamic Tool Dependency Retrieval (DTDR), a lightweight retrieval method that conditions on both the initial query and the evolving tool calling plan. DTDR models tool dependencies from function calling demonstrations, enabling adaptive retrieval as plans unfold. We benchmark DTDR against state-of-the-art retrieval methods across multiple datasets and LLM backbones, evaluating retrieval precision, downstream task accuracy, and computational efficiency. Additionally, we explore strategies to integrate retrieved tools into prompts. Our results show that DTDR improves function calling success rates between 23% and 104% compared to state-of-the-art static retrievers.

Tong Xu, Xinzhe Cao, Zhihui Zhu, Keyan Ding, Huajun Chen

Current molecular generation benchmarks emphasize task complexity, molecule novelty, and property alignment; they largely overlook a critical concern: the potential safety risks of AI-generated molecules. In practice, many generative models may produce molecules with toxic, reactive, or otherwise hazardous characteristics—posing hidden dangers that remain insufficiently addressed. To address this gap, we introduce MolSafeEval, a benchmark dedicated to evaluating and analyzing the safety risks of molecular generation. Unlike prior approaches that rely on narrow toxicity predictors, MolSafeEval integrates heterogeneous safety knowledge—ranging from toxicological databases to hazard rules—into a structured molecular safety knowledge graph. This graph serves as a foundation for large language model–based reasoning, enabling systematic detection and explanation of unsafe features in generated compounds. We further categorize molecular generative models into four representative task types—unconditional generation, property optimization, target protein–based design, and text-based generation—and provide standardized datasets and safety evaluation protocols for each.

Felix Herron, Solange Rossato, Alexandre Allauzen, François Portet

Modern automatic speech recognition (ASR) systems have been observed to function better for certain speaker groups (SGs) than others, despite recent gains in overall performance. One potential impediment to progress towards fairer ASR is a more nuanced understanding of the types of modeling errors that speech encoder models make, and in particular the difference between the structure of embeddings for high-performance and low-performance SGs. This paper proposes a framework typifying two types of error that can occur in modeling phonemes in ASR systems: random error/high variance in phoneme embedding, vs systematic error/embedding bias. We find that training phoneme classification probes only on a single, typically disadvantaged SG, sometimes improves performance for that SG, which is evidence for the existence of SG-level bias in phoneme embeddings. On the other hand, we find that speakers and SGs with higher levels of phoneme variance are the same as those with worse phoneme prediction accuracy. We conclude that both types of error are present in phoneme embeddings and both are candidate causes for SG-level unfairness in ASR, though random error is likely a greater hindrance to fairness than systematic error. Furthermore, we find that finetuning encoder models using a fairness-enhancing algorithm (domain enhancing and adversarial training) changes neither the benefits of in-domain phoneme classification probe training, nor measured levels of random embedding error.