论文检索

Chenglong Wang, Yifu Huo, Yang Gan, Yongyu Mu, Qiaozhi He, Murun Yang, Bei Li, Chunliang Zhang, Tongran Liu, Anxiang Ma 等

Previous methods evaluate reward models by testing them on a fixed pairwise ranking test set, but they typically do not provide performance information on each preference dimension. In this work, we address the evaluation challenge of reward models by probing preference representations. To confirm the effectiveness of this evaluation method, we construct a Multi-dimensional Reward Model Benchmark (MRMBench), a collection of six probing tasks for different preference dimensions. We design it to favor and encourage reward models that better capture preferences across different dimensions. Furthermore, we introduce an analysis method, inference-time probing, which identifies the dimensions used during the reward prediction and enhances its interpretability. Through extensive experiments, we find that MRMBench strongly correlates with LLM alignment performance, supporting it as a reliable reference for developing advanced reward models. By analyzing the evaluation results on MRMBench, we reveal that reward models struggle to simultaneously capture preferences across multiple dimensions, highlighting the potential of multi-objective optimization in reward modeling. Furthermore, our results demonstrate that the proposed inference-time probing method provides a reliable metric for assessing the confidence of reward predictions, leading to improved alignment of large language models.

Chenglong Wang, Yongyu Mu, Hang Zhou, Yifu Huo, Ziming Zhu, Jiali Zeng, Murun Yang, Bei Li, Xiaoyang Hao, Chunliang Zhang 等

Major progress in reward modeling over recent years has been driven by a paradigm shift from task-specific designs to generalist reward models. Despite this trend, developing effective reward models remains a fundamental challenge: the heavy reliance on large-scale labeled preference data. Pre-training on abundant unlabeled data offers a promising direction, but existing approaches fall short in instilling explicit reasoning capabilities into reward models. To bridge this gap, we propose a self-training approach that can leverage unlabeled data to scale up reward reasoning in reward models. Based on this approach, we develop GRAM-R² a generative reward model trained to produce not only preference labels but also accompanying reward rationales. GRAM-R² can serve as a foundation model for reward reasoning and can be applied to a wide range of tasks with minimal or no additional fine-tuning. It can support downstream applications such as policy optimization and task-specific reward tuning. Experiments on response ranking, task adaptation, and reinforcement learning from human feedback demonstrate that GRAM-R² consistently delivers strong performance, outperforming several strong discriminative and generative baselines.

Chenan Wang, Daniel H. Shi, Haipeng Chen

Inference time latency has remained an open challenge for real world applications of large language models (LLMs). State-of-the-art (SOTA) speculative sampling (SpS) methods for LLMs, like EAGLE-3, use tree-based drafting to explore multiple candidate continuations in parallel. However, the hyperparameters controlling the tree structure are static, which limits flexibility and efficiency across diverse contexts and domains. We introduce Reinforcement learning for Speculative Sampling (Re-SpS), the first reinforcement learning (RL)-based framework for draft tree hyperparameter optimization. Re-SpS dynamically adjusts draft tree hyperparameters in real-time, learning context-aware policies that maximize generation speed by balancing speculative aggression with computational overhead. It leverages efficient state representations from target model hidden states and introduces multi-step action persistence for better context modeling. Evaluation results across five diverse benchmarks demonstrate consistent improvements over the SOTA method EAGLE-3, achieving up to 5.45x speedup over the backbone LLM and up to 1.12x speedup compared to EAGLE-3 across five diverse benchmarks, with no loss in output fidelity.

Changyue Wang, Weihang Su, Qingyao Ai, Yiqun Liu

Large Reasoning Models (LRMs) extend large language models with explicit, multi-step reasoning traces to enhance transparency and performance on complex tasks. However, these reasoning traces can be redundant or logically inconsistent, becoming a new and hard-to-detect source of hallucination. Existing hallucination detection methods focus primarily on answer-level uncertainty and often fail to detect hallucinations or logical inconsistencies arising from the model’s reasoning trace. This oversight is particularly problematic for LRMs, where the explicit thinking trace is not only an important support to the model's decision-making process but also a key source of potential hallucination. To this end, we propose RACE (Reasoning and Answer Consistency Evaluation), a novel framework specifically tailored for hallucination detection in LRMs. RACE operates by extracting essential reasoning steps and computing four diagnostic signals: inter-sample consistency of reasoning traces, entropy-based answer uncertainty, semantic alignment between reasoning and answers, and internal coherence of reasoning. This joint analysis enables fine-grained hallucination detection even when the final answer appears correct. Experiments across datasets and different LLMs demonstrate that RACE outperforms existing hallucination detection baselines, offering a robust and generalizable solution for evaluating LRMs.

Changjian Wang, Weihong Deng, Weili Guan, Quan Lu, Ning Jiang

Multi-hop question answering (MHQA) requires integrating knowledge scattered across multiple passages to derive the correct answer. Traditional retrieval-augmented generation (RAG) methods primarily focus on coarse-grained textual semantic similarity and ignore structural associations among dispersed knowledge, which limits their effectiveness in MHQA tasks. GraphRAG methods address this by leveraging knowledge graphs (KGs) to capture structural associations, but they tend to overly rely on structural information and fine-grained word- or phrase-level retrieval, resulting in an underutilization of textual semantics. In this paper, we propose a novel RAG approach called HGRAG for MHQA that achieves cross-granularity integration of structural and semantic information via hypergraphs. Structurally, we construct an entity hypergraph where fine-grained entities serve as nodes and coarse-grained passages as hyperedges, and establish knowledge association through shared entities. Semantically, we design a hypergraph retrieval method that integrates fine-grained entity similarity and coarse-grained passage similarity via hypergraph diffusion. Finally, we employ a retrieval enhancement module, which further refines the retrieved results both semantically and structurally, to obtain the most relevant passages as context for answer generation with the LLM. Experimental results on benchmark datasets demonstrate that our approach outperforms state-of-the-art methods in QA performance, and achieves a 6× speedup in retrieval efficiency.

Bo Wang, Junzhuo Li, Hong Chen, Yuanlin Chu, Yuxuan Fan, Xuming Hu

Mixture-of-Experts (MoE) architectures decouple model capacity from per-token computation, enabling scaling beyond the computational limits imposed by dense scaling laws. Yet how MoE architectures shape knowledge acquisition during pre-training—and how this process differs from dense architectures—remains unknown. To address this issue, we introduce Gated-LPI (Log-Probability Increase), a neuron-level attribution metric that decomposes log-probability increase across neurons. We present a time-resolved comparison of knowledge acquisition dynamics in MoE and dense architectures, tracking checkpoints over 1.2M (~ 5.0T tokens) and 600K (~ 2.5T tokens) training steps, respectively. Our experiments uncover three patterns: (1) Low-entropy backbone. The top approximately 1% of MoE neurons capture over 45% of positive updates, forming a high-utility core, which is absent in the dense baseline. (2) Early consolidation. The MoE model locks into a stable importance profile within < 100K steps, whereas the dense model remains volatile throughout training. (3) Functional robustness. Masking the ten most important MoE attention heads reduces relational HIT@10 by < 10%, compared with > 50% for the dense model, showing that sparsity fosters distributed—rather than brittle—knowledge storage. These patterns collectively demonstrate that sparsity fosters an intrinsically stable and distributed computational backbone from early in training, helping bridge the gap between sparse architectures and training-time interpretability.

Ante Wang, Yujie Lin, Jingyao Liu, Suhang Wu, Hao Liu, Xinyan Xiao, Jinsong Su

Critical thinking is essential for building robust AI systems, preventing them from blindly accepting flawed data or biased reasoning. However, prior work has primarily focused on passive critical thinking, where models simply reject problematic queries without taking constructive steps to address user requests. In this work, we introduce proactive critical thinking, a paradigm where models actively seek missing or clarifying information from users to resolve their queries better. To evaluate this capability, we present GSM-MC and GSM-MCE, two novel benchmarks based on GSM8K for assessing mathematical reasoning under incomplete or misleading conditions. Experiments on Qwen3 and Llama series models show that, while these models excel in traditional reasoning tasks, they struggle with proactive critical thinking, especially smaller ones. However, we demonstrate that reinforcement learning (RL) can significantly improve this ability. By incorporating heuristic information into the reward function, we achieve substantial gains, boosting the Qwen3-1.7B's accuracy from 0.15% to 73.98% on GSM-MC. We hope this work advances models that collaborate more effectively with users in problem-solving through proactive critical thinking.

Haiyuan Wan, Chen Yang, Junchi Yu, Meiqi Tu, Jiaxuan Lu, Di Yu, Jianbao Cao, Ben Gao, Jiaqing Xie, Aoran Wang 等

Deep research agents have attracted growing attention for their potential to orchestrate multi-stage research workflows, spanning literature synthesis, methodological design, and empirical verification. Despite these strides, evaluating their research capability faithfully is rather challenging due to the difficulty of collecting frontier research questions that genuinely capture researchers’ attention and intellectual curiosity. To address this gap, we introduce DeepResearch Arena, a benchmark grounded in academic seminars that capture rich expert discourse and interaction, better reflecting real-world research environments and reducing the risk of data leakage. To automatically construct DeepResearch Arena, we propose a Multi-Agent Hierarchical Task Generation (MAHTG) system that extracts research-worthy inspirations from seminar transcripts. The MAHTG system further translates research-worthy inspirations into high-quality research tasks, ensuring the traceability of research task formulation while filtering noise. With the MAHTG system, we curate DeepResearch Arena with over 10,000 high-quality research tasks from over 200 academic seminars, spanning 12 disciplines, such as literature, history, and science. Our extensive evaluation shows that DeepResearch Arena presents substantial challenges for current state-of-the-art agents, with clear performance gaps observed across different models.

Hoang Tran Vuong, Tue Le, Quyen Tran, Linh Ngo Van, Trung Le

Knowledge distillation (KD) is widely recognized as an effective approach for compressing large language models (LLMs). However, standard KD methods often falter when confronted with architectural or tokenization heterogeneity between teacher and student models, which creates a mismatch in their representations. While Optimal Transport (OT) provides a promising solution to align these representations, most OT-based methods rely on a single cost function, which isn’t enough to capture the multifaceted discrepancies between models with distinct designs. To address this limitation, we introduce Multi-Cost Wasserstein Knowledge Distillation (MCW-KD), a novel framework that enhances KD by simultaneously optimizing several cost functions within a unified OT formulation. MCW-KD employs specific cost matrices to effectively align both the final hidden states and the output distributions of the models. We also provide a rigorous theoretical foundation for the proposed Multi-Cost Wasserstein Distance, ensuring both mathematical validity and computational ability. Extensive experiments on instruction-following datasets demonstrate that MCW-KD significantly improves student model performance compared to state-of-the-art KD baselines, especially when teacher and student models have different tokenizers.

Sachin Vashistha, Aryan Bibhuti, Atharva Naik, Martin Tutek, Somak Aditya

Real-world conversations are rich with pragmatic elements, such as entity mentions, references, and implicatures. Understanding such nuances is a requirement for successful natural communication, and often requires building a local _world model_ which encodes such elements and captures the dynamics of their evolving states. However, it is not well-understood whether language models (LMs) construct or maintain a robust implicit representation of conversations. In this work, we evaluate the ability of LMs to encode and update their internal world model in dyadic conversations and test their _malleability_ under linguistic alterations. To facilitate this, we apply seven minimal linguistic alterations to conversations sourced from popular conversational QA datasets and construct a benchmark with two variants (i.e., Manual and Synthetic) comprising yes-no questions. We evaluate nine open and one closed source LMs and observe that they struggle to maintain robust accuracy. Our analysis unveils that LMs struggle to memorize crucial details, such as tracking entities under linguistic alterations to conversations. We then propose a dual-perspective interpretability framework which identifies transformer layers that are _useful_ or _harmful_ and highlights linguistic alterations most influenced by harmful layers, typically due to encoding spurious signals or relying on shortcuts. Inspired by these insights, we propose two layer-regularization based fine-tuning strategies that suppress the effect of the harmful layers.

Marco Valentino, Geonhee Kim, Dhairya Dalal, Zhixue Zhao, André Freitas

Large language models (LLMs) exhibit reasoning biases, often conflating content plausibility with formal logical validity. This can lead to wrong inferences in critical domains, where plausible arguments are incorrectly deemed logically valid or vice versa. This paper investigates how content biases on reasoning can be mitigated through activation steering, an inference-time technique that modulates internal activations. Specifically, after localising the layers responsible for formal and plausible inference, we investigate activation steering on a controlled syllogistic reasoning task, designed to disentangle formal validity from content plausibility. An extensive empirical analysis reveals that contrastive steering methods consistently support linear control over content biases. However, a static approach is insufficient to debias all the tested models. We then investigate how to control content effects by dynamically determining the steering parameters through fine-grained conditional methods. By introducing a novel kNN-based conditional approach (K-CAST), we demonstrate that conditional steering can effectively reduce biases on unresponsive models, achieving up to 15% absolute improvement in formal reasoning accuracy. Finally, we found that steering for content effects is robust to prompt variations, incurs minimal side effects on multilingual language modeling capabilities, and can partially generalize to different reasoning tasks. In practice, we demonstrate that activation-level interventions offer a scalable inference-time strategy for enhancing the robustness of LLMs, contributing towards more systematic and unbiased reasoning capabilities

Thomas Jean-Michel Valentin, Ardi Madadi, Gaetano Sapia, Marcel Böhme

Generating code from a natural language programming task is one of the most successful applications of Large Language Models (LLMs). Yet, the generated program may be buggy. Without an oracle, such as an existing, correct implementation or a formal specification, can we somehow estimate how likely the generated program is correct? In this paper, we propose a measure of incorrectness, called incoherence, that can be estimated efficiently in the absence of an oracle and allows us to establish a lower bound on the error, i.e., the probability that the LLM-generated program for that specification is incorrect. In our experiments, our incoherence-based methodology can automatically identify about two-thirds of incorrect programs without reports of false positives for the average task. In fact, an oracle-based evaluation of LLMs can be reliably replaced by an incoherence-based evaluation. In particular, we find a very strong agreement between the ranking of LLMs by the number of programs deemed correct via an oracle (pass@1) and the ranking of LLMs by the number of programs deemed correct via incoherence.

Omri Uzan, Yuval Pinter

Tasks that require character-level reasoning, such as counting or locating characters within words, remain challenging for contemporary language models. A common conjecture is that language models' reliance on subword units, rather than characters, contributes to their struggles with character-level tasks, yet recent studies offer conflicting conclusions about the role of tokenization, leaving its impact unclear. To address this gap, we introduce CharBench, a comprehensive benchmark of character-level tasks that is two orders of magnitude larger than existing alternatives. We evaluate a diverse range of leading open-weight and proprietary models on CharBench and find that it presents a significant challenge to modern LLMs, with average accuracies of 43.6% and 32.3% on some tasks. We present an in-depth analysis of how intrinsic properties of words and their segmentations into tokens correspond to model performance. For counting tasks, we find that tokenization properties are weakly correlated with correctness, while the length of the queried word and the actual character count play a more significant part. In contrast, for tasks requiring intra-word positional understanding, performance is negatively correlated with the length of the token containing the queried character, suggesting that longer tokens obscure information on character position for LLMs. We encourage future work to build on the benchmark and evaluation methodology introduced here as tools for improving model performance on these tasks.

Kumud Tripathi, Aditya Srinivas Menon, Aman Gaurav, Raj Prakash Gohil, Pankaj Wasnik

The Whisper model, an open-source automatic speech recognition system, is widely adopted for its strong performance across multilingual and zero-shot settings. However, it frequently suffers from hallucination errors, especially under noisy acoustic conditions. Previous works to reduce hallucinations in Whisper-style ASR systems have primarily focused on audio preprocessing or post-processing of transcriptions to filter out erroneous content. However, modifications to the Whisper model itself remain largely unexplored to mitigate hallucinations directly. To address this challenge, we present a two-stage architecture that first enhances encoder robustness through Adaptive Layer Attention (ALA) and further suppresses hallucinations using a multi-objective knowledge distillation (KD) framework. In the first stage, ALA groups encoder layers into semantically coherent blocks via inter-layer correlation analysis. A learnable multi-head attention module then fuses these block representations, enabling the model to jointly exploit low- and high-level features for more robust encoding. In the second stage, our KD framework trains the student model on noisy audio to align its semantic and attention distributions with a teacher model processing clean inputs. Our experiments on noisy speech benchmarks show notable reductions in hallucinations and word error rates, while preserving performance on clean speech. Together, ALA and KD offer a principled strategy to improve Whisper’s reliability under real-world noisy conditions.

Khanh-Tung Tran, Barry O'Sullivan, Hoang D. Nguyen

Recent advances have enabled Large Language Models (LLMs) to tackle reasoning tasks by generating chain-of-thought (CoT) rationales, yet these gains have largely applied to high-resource languages, leaving low-resource languages underperformed. In this work, we first investigate CoT techniques in extremely low-resource scenarios through previous prompting, model editing, and fine-tuning approaches. We introduce \emph{English-Pivoted CoT Training}, leveraging the insight that LLMs internally operate in a latent space aligned toward the dominant language. Given input in a low-resource language, we perform supervised fine-tuning to generate CoT in English and output the final response in the target language. Across mathematical reasoning benchmarks, our approach outperforms other baselines with up to 28.33% improvement in low-resource scenarios. Our analyses and additional experiments, including Mixed-Language CoT and Two-Stage Training, show that explicitly separating language understanding from reasoning enhances crosslingual reasoning abilities. To facilitate future work, we also release LC2024, the first benchmark for mathematical task in Irish, an extremely low-resource and endangered language. Our results and resources highlight a practical pathway to multilingual reasoning without extensive retraining in every extremely low-resource language, despite data scarcity.

Hieu Tran, Zonghai Yao, Nguyen Luong Tran, Zhichao Yang, Feiyun Ouyang, Shuo Han, Razieh Rahimi, Hong Yu

Inspired by the dual-process theory of human cognition from Thinking, Fast and Slow, we introduce PRIME (Planning and Retrieval-Integrated Memory for Enhanced Reasoning), a multi-agent reasoning framework that dynamically integrates System 1 (fast, intuitive thinking) and System 2 (slow, deliberate thinking). PRIME first employs a Quick Thinking Agent to generate a rapid answer; if uncertainty is detected, it then triggers a structured System 2 reasoning pipeline composed of specialized agents for planning, hypothesis generation, retrieval, information integration, and decision-making. This multi-agent design mimics human cognitive processes faithfully and enhances both efficiency and accuracy. Experimental results with LLaMA 3 models demonstrate that PRIME enables open-source LLMs to perform competitively with state-of-the-art closed-source models like GPT-4 and GPT-4o on benchmarks requiring multi-hop and knowledge-grounded reasoning. This research establishes PRIME as a scalable solution for improving LLMs in domains requiring complex, knowledge-intensive reasoning.

Yuxuan Tian, Zihan Wang, Yebo Peng, Aomufei Yuan, Zhiming Wang, Bairen Yi, Xin Liu, Yong Cui, Tong Yang

Efficient inference of large language models (LLMs) is hindered by an ever-growing key-value (KV) cache, making KV cache compression a critical research direction. Traditional methods selectively evict less important KV cache entries, which leads to information loss and hallucinations. Recently, merging-based strategies have been explored to retain more information by merging KV pairs that would be discarded; however, these existing approaches inevitably introduce inconsistencies in attention distributions before and after merging, causing degraded generation quality. To overcome this challenge, we propose KeepKV , a novel adaptive KV cache merging method designed to preserve performance under strict memory constraints, achieving single-step lossless compression and providing error bounds for multi-step compression. KeepKV introduces the Electoral Votes mechanism that records merging history and adaptively adjusts attention scores. Moreover, it further leverages a novel Zero Inference-Perturbation Merging method, compensating for attention loss resulting from cache merging. Extensive experiments on various benchmarks and LLM architectures demonstrate that KeepKV substantially reduces memory usage while successfully retaining essential context information, achieving over 2 times inference throughput improvement and maintaining superior generation quality even with only 10% KV cache budgets.

Ye Tian, Chengcheng Wang, Jing Han, Yehui Tang, Kai Han

As Large Language Models (LLMs) continue to grow in size, storing and transmitting them on edge devices becomes increasingly challenging. Traditional methods like quantization and pruning struggle to achieve extreme compression of LLMs without sacrificing accuracy. In this paper, we introduce PocketLLM, a novel approach to compress LLMs in a latent space via meta-networks. A simple encoder network is proposed to project the weights of LLMs into discrete latent vectors, which are then represented using a compact codebook. A lightweight decoder network is employed to map the codebook's representative vectors back to the original weight space. This method allows for significant compression of the large weights in LLMs, consisting solely of a small decoder, a concise codebook, and an index. Extensive experiments show that PocketLLM achieves superior performance even at significantly high compression ratios, e.g., compressing Llama 2-7B by 10x with a negligible drop in accuracy.

Changyuan Tian, Zhicong Lu, Shuang Qian, Nayu Liu, Peiguang Li, Li Jin, Leiyi Hu, Zhizhao Zeng, Sirui Wang, Ke Zeng 等

To improve Multi-step Mathematical Reasoning (MsMR) of Large Language Models (LLMs), it is crucial to obtain scalable supervision from the corpus by automatically critiquing mistakes in the reasoning process of MsMR and rendering a final verdict of the problem-solution. Most existing methods rely on crafting high-quality supervised fine-tuning demonstrations for critiquing capability enhancement and pay little attention to delving into the underlying reason for the poor critiquing performance of LLMs. In this paper, we orthogonally quantify and investigate the potential reason — imbalanced evaluation preference, and conduct a statistical preference analysis. Motivated by the analysis of the reason, a novel perplexity-aware reinforcement learning algorithm is proposed to rectify the evaluation preference, elevating the critiquing capability. Specifically, to probe into LLMs' critiquing characteristics, a One-to-many Problem-Solution (OPS) benchmark is meticulously constructed to quantify the behavior difference of LLMs when evaluating the problem solutions generated by itself and others. Then, to investigate the behavior difference in depth, we conduct a statistical preference analysis oriented on perplexity and find an intriguing phenomenon — "LLMs incline to judge solutions with lower perplexity as correct", which is dubbed as imbalanced evaluation preference. To rectify this preference, we regard perplexity as the baton in the algorithm of Group Relative Policy Optimization, supporting the LLMs to explore trajectories that judge lower perplexity as wrong and higher perplexity as correct. Extensive experimental results on our built OPS and existing available critic benchmarks demonstrate the validity of our method.

Pengwei Tang, Xiaolin Hu, Yong Liu, Lizhong Ding, Dongjie Zhang, Xing Wu, Debing Zhang

Low-Rank Adaptation (LoRA) is the leading parameter-efficient fine-tuning method for Large Language Models (LLMs), but it still suffers from catastrophic forgetting. Recent work has shown that specialized LoRA initialization can alleviate catastrophic forgetting. There are currently two approaches to LoRA initialization aimed at preventing knowledge forgetting during fine-tuning: (1) making residual weights close to pre-trained weights, and (2) ensuring the space of LoRA initialization is orthogonal to pre-trained knowledge. The former is what current methods strive to achieve, while the importance of the latter is not sufficiently recognized. We find that the space of LoRA initialization is the key to preserving pre-trained knowledge rather than the residual weights. Existing methods like MiLoRA propose making the LoRA initialization space orthogonal to pre-trained weights. However, MiLoRA utilizes the null space of pre-trained weights. Compared to pre-trained weights, the input activations of pre-trained knowledge take into account the parameters of all previous layers as well as the input data, while pre-trained weights only contain information from the current layer. Moreover, we find that the effective ranks of input activations are much smaller than those of pre-trained weights. Thus, the null space of activations is more accurate and contains less pre-trained knowledge information compared to that of weights. Based on these, we introduce LoRA-Null, our proposed method that initializes LoRA in the null space of activations. Experimental results show that LoRA-Null effectively preserves the pre-trained world knowledge of LLMs while achieving good fine-tuning performance, as evidenced by extensive experiments.