论文检索

Yang Li, Yajiao Wang, Yu Zhang, Yuanzhe Zhang, Maodi Hu, Mengting Zhang, Xi Sun, Hua Yue, Zhixiong Zhang

Scientific information extraction (SciIE) is a key bottleneck for turning unstructured papers into computable knowledge bases, yet most existing systems still follow a “local extraction then global assembly” paradigm. This workflow is inherently lossy: by extracting fields in isolation, it breaks global correlations and discards high-confidence signals that could otherwise be reused as internal supervision, forcing systems to repeatedly restart from scratch, especially in long, multimodal scientific documents. In this paper, We propose a different view: SciIE should be solved as a progressive filling problem, similar to solving a Sudoku,once a field is filled with high confidence, it should act as a constraint that guides the remaining uncertain fields. Based on this idea, we introduce SudokuFill, a multi-agent framework that maintains a Global Filling State and performs priority scheduling to establish reliable anchors first, then reuses them as internal supervision for iterative deliberation over harder fields. Evaluated on a specialized document-level adjuvant dataset, our framework achieves a SOTA score of 51.83% on our benchmark. Crucially, SudokuFill enables a 7B model to outperform the vanilla GPT-4o, proving that structured architectural reasoning can effectively compensate for parameter scale.

Dongling Li, Qi Chen, Jianxing Yu, Hanjiang Lai, Yanghui Rao, Wenqing Chen, Jian Yin

This paper focuses on the task of answering complex visual questions that involve cross-dimensional (like 2D to 3D) spatial reasoning. This task (called SpatialQA) can enhance the machine’s spatial cognitive abilities in "plane representation - space reconstruction - semantic inference," having great application value. Existing methods often only recognize 1-D visual objects and relations, but they lack the ability to represent in a cross-dimensional space and fail to grasp structured geometric knowledge such as face-face topology and texture details. That would cause problems such as texture misalignment and topological confusion, leading to error accumulation and incorrect answers. To address this problem, we propose a new method with good cross-dimensional reasoning capabilities. In detail, we first analyze the input image, capturing its relations in the 2D plane. To derive the topological relations in the 3D space, we employ a dual-channel augmentation technique to retrieve topological isomorphic examples and geometric rules, supplementing the missing but crucial reasoning clues. We then design a multi-perspective verifier to find the inconsistencies of the macroscopic outlines, eliminating incorrect options. Based on visual clues, we develop a question-guided detector to analyze the texture details and relations of each surface finely, capturing inconsistencies in a micro level. That can correct the reasoning bias to derive the right answer. Moreover, we create a large-scale dataset with 22,483 samples to conduct evaluations. The results show the effectiveness of our method.

Tianyu Liu, Qitan Lv, Hao Li, Xing Gao, Xiao Sun, Xiaoyan Sun

Speculative decoding (SD), where a small draft model is employed to propose *draft* tokens in advance and then the target model validates them in parallel, has emerged as a promising technique for LLM inference acceleration. Many endeavors to improve SD are to eliminate the need for a draft model and generate draft tokens in a retrieval-based manner in order to further alleviate the drafting overhead and significantly reduce the difficulty in deployment and applications. However, retrieval-based SD relies on a matching paradigm to retrieve the most relevant reference as the draft tokens, where these methods often fail to find matched and accurate draft tokens. To address this challenge, we propose *LogitSpec* to effectively expand the retrieval range and find the most relevant reference as drafts. *LogitSpec* is motivated by the observation that the logit of the last token can not only predict **the next token**, but also speculate **the next next token**. Specifically, *LogitSpec* generates draft tokens in two steps: (1) utilizing the last logit to speculate the next next token; (2) retrieving relevant reference for both the next token and the next next token. *LogitSpec* is training-free and plug-and-play, which can be easily integrated into existing LLM inference frameworks. Extensive experiments on a wide range of text generation benchmarks demonstrate that *LogitSpec* can achieve up to 2.61× speedup and 3.28 mean accepted tokens per decoding step.

Linzhuang Sun, Tianyu Guo, Hao Liang, Ruitong Liu, Yuying Li, Qifeng Cai, Jingxuan Wei, Yuchen Wu, Bihui Yu, Xiangxiang Zhang 等

Recent advancements in Large Language Models (LLMs) have revolutionized Text-to-SQL parsing, achieving remarkable success in static, single-turn query generation. However, a significant disparity remains between these academic benchmarks and real-world utility. In practical applications, such as financial auditing or business analytics, user intents are rarely static; they evolve dynamically through iterative refinement, necessitating not just information retrieval (SELECT) but continuous state manipulation (INSERT, UPDATE, DELETE). To bridge this gap, we introduce DySQL-Bench, a novel benchmark designed to rigorously evaluate LLMs within a dynamic interaction framework. Unlike varying manual curation efforts, DySQL-Bench employs a two-stage automated synthesis pipeline: transforming raw relational schemas into hierarchical logic trees to generate user-database interactions, followed by a rigorous verify-and-refine protocol that ensures 100% distinct correctness via human expert validation. We further propose an interactive evaluation environment simulating a triadic workflow involving an LLM-simulated user, the agent under test, and an executable database system. Spanning 13 diverse domains with 1,072 complex tasks, our experiments reveal that current powerful models struggle in this realistic setting. Notably, GPT-4o achieves only 58.34% overall accuracy and a meager 23.81% on the strict Pass^5 metric, highlighting the substantial challenges DySQL-Bench poses for the future of database agents.

Mingzi Cao, Xingwei Tan, Mahmud Elahi Akhter, Marco Valentino, Maria Liakata, Xi Wang, Nikolaos Aletras

Deduction, induction, and abduction are fundamental reasoning paradigms, core for human logical thinking. Although improving Large Language Model (LLM) reasoning has attracted significant research efforts, the extent to which the fundamental paradigms induce generalization has yet to be systematically explored. In this study, we shed light on how the interplay between these core paradigms influences LLMs’ reasoning behavior. To this end, we first collect a new dataset of reasoning trajectories from symbolic tasks, each targeting one of the three fundamental paradigms, to abstract from concrete world knowledge. Then, we investigate effective ways for inducing these skills into LLMs.We experiment with a battery of methods including simple fine-tuning, and more complex approaches to increase model depth, or transform a dense model to a mixture-of-experts. We comprehensively evaluate induced models on realistic out-of-domain tasks, that are entirely formulated in natural language and contain real-world knowledge. Our results reveal that our approach yields strong generalizability with substantial performance gains (up to 14.60) across realistic tasks.

Yaoqi Guo, Ying Xiao, Jie M. Zhang, Mark Harman, Yiling Lou, Yang Liu, Zhenpeng Chen

Software engineering (SE) agents powered by large language models are increasingly adopted in practice, yet they often incur substantial monetary cost. We introduce EET, an experience-driven early termination approach that reduces the cost of SE agents while preserving task performance. EET extracts structured experience from prior issue-resolution executions and leverages it to guide early termination during patch generation and selection, reducing unproductive iterations. We evaluate EET on the SWE-bench Verified benchmark across three representative SE agents. EET consistently reduces total cost by 19%–55% (32% on average), with negligible loss in resolution rate (at most 0.2%). These efficiency gains are achieved, on average, by identifying early-termination opportunities for 11% of issues and reducing API calls, input tokens, and output tokens by 21%, 30%, and 25%, respectively. We release the code, prompts, and data at https://github.com/IanWalls/EET.

Sunguk Shin, Hayeong Lee, Jun Ho Seo, Jinho Lee, Myunsoo Kim, Minsuk Chang, Byung-Jun Lee

While the reasoning capabilities of large language models (LLMs) have advanced considerably, efficiently internalizing and leveraging new information in dynamically interactive environments remains a significant challenge. This limitation is particularly pronounced in partially observable environments, which require agents to manage long-term memory and perform effective exploration under incomplete information. To address this, we propose an LLM agent architecture that integrates a knowledge graph as a graph-based memory module. The agent incrementally constructs the knowledge graph through environmental interactions and retrieves relevant information to generate efficient plans. We evaluate our approach in complex navigation tasks specifically designed to present long-horizon and partially observable challenges. Experimental results demonstrate that incorporating a self-extending memory module significantly enhances the performance and efficiency of the LLM’s planning capabilities.

Xinzhu Chen, Xuesheng Li, Zhongxiang Sun, Weijie Yu

Reinforcement Learning with Verifiable Rewards (RLVR) has become a central approach for improving the reasoning ability of large language models. Recent work studies RLVR through token entropy, arguing that high-entropy tokens drive exploration and should receive stronger updates. However, they overlook the fact that most of a reasoning trajectory consists of low-entropy segments that encode stable and reusable structural patterns. Through qualitative and quantitative analyses, we find that the overlap of low-entropy segments across correct responses strongly correlates with model accuracy, while overlaps involving incorrect responses exhibit stable but unproductive patterns. Motivated by these findings, we propose LESS, a correctness-aware reinforcement framework that performs fine-grained advantage modulation over low-entropy segments. LESS amplifies segments unique to correct responses, suppresses those unique to incorrect ones, and neutralizes segments shared by both, while preserving high-entropy exploration in the underlying RL algorithm. Instantiated on top of GRPO and GSPO, LESS not only improves accuracy over strong RL baselines across three backbones and six math benchmarks, but also achieves stronger robustness of the performance floor.

Hyungjun Yoon, Mohammad Malekzadeh, Sung-Ju Lee, Fahim Kawsar, Lorena Qendro

Large language models (LLMs) are increasingly grounded in sensor data to perceive and reason about human physiology and the physical world. However, accurately interpreting heterogeneous multimodal sensor data remains a fundamental challenge. We show that a single monolithic LLM often fails to reason coherently across modalities, leading to incomplete interpretations and prior-knowledge bias. We introduce ConSensus, a training-free multi-agent collaboration framework that decomposes multimodal sensing tasks into specialized, modality-aware agents. To aggregate agent-level interpretations, we propose a hybrid fusion mechanism that balances semantic aggregation, which enables cross-modal reasoning and contextual understanding, with statistical consensus, which provides robustness through agreement across modalities. While each approach has complementary failure modes, their combination enables reliable inference under sensor noise and missing data. We evaluate ConSensus on five diverse multimodal sensing benchmarks, demonstrating an average accuracy improvement of 7.1% over the single-agent baseline. Furthermore, ConSensus matches or exceeds the performance of iterative multi-agent debate methods while achieving a 12.7 times reduction in average fusion token cost through a single-round hybrid fusion protocol, yielding a robust and efficient solution for real-world multimodal sensing tasks. The source code is available at https://github.com/nokia/multi-agent-collaboration-for-multimodal-sensing.

Hao Zhang, Zhenjia Li, Yifan Gao, Xi Xiao, Heng Zhang, Shuyang Zhang, Xiaoxincc, Bo Huang, Yuhang Wu, Tianyang Wang 等

Parameter-Efficient Fine-Tuning (PEFT), especially Low-Rank Adaptation (LoRA), has emerged as a promising approach to fine-tuning large language models(LLMs) while reducing computational and memory overhead. However, LoRA assumes a uniform rank r for each incremental matrix, not accounting for the varying significance of weight matrices across different modules and layers. AdaLoRA leverages Singular Value Decomposition (SVD) to parameterize updates and employs pruning of singular values to introduce dynamic rank allocation, thereby enhancing adaptability. However, during the training process, it often encounters issues of slow convergence speed and high computational overhead. To address this issue, we propose HyperAdaLoRA, a novel framework that accelerates the convergence of AdaLoRA by leveraging a hypernetwork. Instead of directly optimizing the components of Singular Value Decomposition (P, \Lambda, Q), HyperAdaLoRA employs a hypernetwork based on attention mechanisms to dynamically generate these parameters. By pruning the outputs of the hypernetwork that generates the singular values, dynamic rank allocation is achieved. Comprehensive experiments on various datasets and models demonstrate that our method achieves faster convergence without sacrificing performance. Moreover, our method generalizes well to other LoRA-based approaches, highlighting its strong generalization capability.

Linzhuang Sun, Mingyang Chen, Hao Liang, Tianpeng Li, Zhou Yijie, Chenzheng Zhu, Tianyu Guo, Huanyao Zhang, Jingxuan Wei, Bihui Yu 等

Training effective AI agents for real-world tool-use interactions requires data that faithfully captures the dynamics of human–agent collaboration. However, such data is scarce, and existing methods often resort to synthetic data generation. The inherently dynamic and complex nature of user–agent interactions makes ensuring data quality particularly challenging. Current verification approaches are typically entangled with the synthesis process itself, resulting in complicated implementations that undermine both reproducibility and scalability. To address this, we introduce Tool-Verifier-7B, a plug-and-play framework for data quality control in tool-use scenarios. Building on this verifier and our data synthesis strategy, we construct the Tool-Verify dataset, which contains 3,295 curated samples. To directly assess verifier performance, we further release Tool-V-Bench, a benchmark of 165 human-validated trajectories spanning diverse interaction complexities. Comprehensive experiments show that Tool-Verifier-7B surpasses Qwen2.5-72B-Instruct on Tool-V-Bench. Moreover, the Tool-Verify dataset achieves superior performance compared to the previous APIGen-MT dataset.

Yimin Deng, Zhenxi Lin, Yejing Wang, Guoshuai Zhao, Pengyue Jia, Zichuan Fu, Derong Xu, Yefeng Zheng, Xiangyu Zhao, Li Zhu 等

Diagnostic prediction and clinical reasoning are critical tasks in healthcare applications. While large language models have shown strong capabilities in commonsense reasoning, they still struggle with diagnostic reasoning due to limited domain knowledge. Existing approaches often rely on internal model knowledge or static knowledge bases, which are insufficient to support the knowledge demands of diagnostic reasoning. Moreover, these methods focus solely on the accuracy of final predictions, overlooking alignment with standard clinical reasoning trajectories. To this end, we propose MultiDx, a two-stage diagnostic reasoning framework that performs differential diagnosis by analyzing evidence collected from multiple knowledge sources. Specifically, it first generates suspected diagnoses and reasoning traces by leveraging knowledge from web search, SOAP-formatted case, and clinical case database. Then it integrates multi-perspective evidence through matching, voting, and differential diagnosis to generate the final prediction. Extensive experiments demonstrate the effectiveness of our approach.

Biswesh Mohapatra, Théo Charlot, Giovanni Duca, Mayank Palan, Laurent Romary, Justine Cassell

Common ground plays a critical role in situated spoken dialogues, where interlocutors must establish and maintain shared references to entities, events, and relations to sustain coherent interaction. For dialog systems, the ability to correctly ground conversational content in order to refer back to it later is particularly important. Prior studies have demonstrated that LLMs are capable of performing grounding acts such as requesting clarification or producing acknowledgments, yet relatively little work has investigated how common ground can be explicitly represented and stored for later use. Without such mechanisms, it remains unclear whether acknowledgment or clarification behaviors truly reflect a grounded understanding. In this work, we evaluate a model’s ability to establish and exploit common ground through relational references to entities within the shared context in a situational dialogue. We test multiple methods for representing common ground in situated dialogues and further propose approaches to improve both the establishment of common ground and its subsequent use in the conversation.

Niklas Mellgren, Peter Schneider-Kamp, Lukas Galke Poech

Language models frequently produce plausible yet incorrect reasoning traces that are difficult to verify. We investigate fine-tuning models to use Prolog as an external symbolic reasoning tool, training Qwen2.5-3B-Instruct with Group Relative Policy Optimization (GRPO) on a cleaned version of GSM8K (which we release as gsm8k-prolog-prover). We systematically vary prompt structure, reward composition (execution, syntax, semantics, structure), and inference protocol (single-try, multiple-try, and two agentic modes). Our reinforcement learning approach outperforms supervised fine-tuning on GSM8K, and the resulting 3B model achieves zero-shot performance on MMLU-STEM and MMLU-Pro competitive with 7B few-shot baselines. Most importantly, we identify an accuracy–auditability trade-off: configurations tuned for correctness alone learn to delegate reasoning to natural language and use Prolog only for the final computation, while configurations rewarded for symbolic structure produce fully auditable programs at a cost in accuracy. We interpret this trade-off as a form of reward hacking and discuss its implications for deploying neurosymbolic systems in safety-critical domains.

Xianming Hu, Jingyang Chen, Bin Tang, Yihe Liu, Yihong Huang, Hongbo Zhao, Nuoyi Chen, Jie Zhang, Ping Li, Kai Zhang

Retrieval-Augmented Generation is a powerful tool for NLP applications. Yet, it is challenging to encode large knowledge bases as compact offline structures while simultaneously achieving accurate, low-latency online retrieval. We propose **ZoomRAG**, a coarse-to-fine, hierarchical graph inference method to tackle the challenges. ZoomRAG formulates the retrieval task as random walks across multi-scale relational graphs. *At the coarse level*, it constructs a global relational graph and performs a query-initiated random walk to quickly locate a few relevant documents over the entire corpus. *At the finer level*, it “zooms into” the selected documents to capture fine-grained semantic and temporal relations, and conducts a second random walk to pinpoint salient evidence chunks for generation. This coarse-to-fine strategy substantially reduces offline indexing costs and accelerates online retrieval. Moreover, random-walk based topological reasoning over rich, multi-scale relational structures enables ZoomRAG to effectively aggregate multi-hop evidence while suppressing noise. Finally, we address the difficulty of handling concurrent RAG queries by **algorithm-parallel ZoomRAG**. Overall, ZoomRAG avoids building expensive knowledge graphs while achieving 2.2% – 4.9% absolute gains in accuracy over SOTA RAG models, with an average online retrieval latency per-query as low as 0.019 secs by processing hundreds of queries concurrently.

Raman Saparkhan, Majd Hawasly, Md Rizwan Parvez, Mohammad Raza

Self-consistency (SC) is a popular technique for improving the reasoning accuracy of large language models by aggregating multiple sampled outputs, but it comes at a high computational cost due to extensive sampling. We introduce a hybrid ensembling approach that leverages the complementary strengths of two distinct modes of reasoning: Chain-of-Thought (CoT) and Program-of-Thought (PoT). We describe a general framework for combining these two forms of reasoning in self-consistency, as well as particular strategies for both full sampling and early-stopping. We show that CoT-PoT ensembling not only improves overall accuracy, but also drastically reduces the number of samples required for SC by a factor of 9.3x. In particular, the majority of tasks (78.6%) can be addressed with only two samples, which has not been possible with any prior SC methods.

Kaleen Shrestha, Abhinav Gupta, Harish Dukkipati, Zhonghao Shi, Maja Mataric

Humans are able to predict each other’s actions by reasoning about the others’ underlying goals, preferences, and motives, such as greed and risk-aversion. Game theory provides a framework for studying human behaviors through incentivized games that simulate social situations. We utilized two validated games from the cognitive science literature—the Social Prediction Game (SPG) and the Inspection Game (IG)—to systematically study how well several recent open- and closed-source LLMs predict player actions and whether they can leverage and generalize the players’ motives learned from the iterated games. Our results indicate that state-of-the-art LLMs can achieve accuracy close to human levels in predicting players’ actions with underlying human motives in SPGs. However, unlike humans, who rely on reasoning about players’ motives to inform their predictions, LLMs failed to recognize statistical patterns in players’ actions. As a result, LLM prediction accuracy did not improve over multiple rounds. Our results in the IG further demonstrate that, unlike humans, LLMs were unable to recognize a player’s underlying motives and to generalize their understanding of the same player to a new context. This suggests that LLMs may lack reasoning capabilities. Our findings offer insights into differences in human and LLM reasoning mechanisms, suggesting that further research into human-AI alignment is needed before utilizing LLMs for human behavior modeling and simulation in this and related contexts.

Zhichao Sheng, Shilin Zhou, Chen Gong, Zhenghua Li

Large Audio Language Models (LALMs) employing the Chain-of-Thought paradigm have demonstrated remarkable reasoning capabilities. Though different problems naturally require varying depths of reasoning, existing methods often determine whether to perform reasoning, lacking fine-grained mechanisms to adapt reasoning length to problem complexity. As a result, LALMs often adopt a one-size-fits-all reasoning strategy, leading to redundant overthinking for simple tasks and insufficient reasoning for complex ones. In this paper, we conduct an in-depth analysis of LALM reasoning behavior and argue that effective and efficient reasoning should be adaptively aligned with task difficulty. To this end, we propose a difficulty-adaptive reasoning method for LALMs. Specifically, we introduce a reward function that dynamically links reasoning length to the model’s perceived problem difficulty, encouraging shorter reasoning for easy tasks and longer reasoning for more complex ones. Extensive experiments on three datasets demonstrate that our method consistently improves performance while reducing average reasoning length by at least 50%, achieving higher efficiency without sacrificing accuracy.

Sarmistha Das, Vaibhav Vishal, Syed Ibrahim Ahmad, Sriparna Saha, Manish Gupta

Financial decision-making in multilingual settings demands accurate numerical reasoning grounded in diverse modalities, yet existing benchmarks largely overlook this high-stakes, real-world challenge, especially for Indic languages. We introduce FinVQA, a benchmark for evaluating financial numerical and multimodal reasoning in multilingual Indic contexts. FinVQA spans English, Hindi, Bengali, Marathi, Gujarati, and Tamil, and comprises 18,900 samples across 14 financial domains. The dataset captures diverse reasoning paradigms under realistic constraints, and is structured across three difficulty levels (easy, moderate, hard) and four question formats: multiple choice, fill-in-the-blank, table matching, and true/false. To address these challenges, we propose FIND, a framework that combines supervised fine-tuning with constraint-aware decoding to promote faithful numerical reasoning, robust multimodal grounding, and structured decision-making. Together, FinVQA and FIND establish a rigorous evaluation and modeling paradigm for high-stakes multilingual multimodal financial reasoning.

Yifei Cao, Changhao Jiang, Jiabao Zhuang, Jiajun Sun, Ming Zhang, Zhiheng Xi, Hui Li, Shihan Dou, Yuran Wang, Yunke Zhang 等

Speech quality assessment (SQA) is typically formulated as a score regression task based on subjective ratings, such as the Mean Opinion Score (MOS), which inherently suffer from inconsistent standards and limit cross-dataset training and evaluation. To address these limitations, we reformulate SQA as a preference-based comparison paradigm and construct MOS-Pref, a large-scale MOS-derived preference dataset. Building on MOS-Pref, we systematically implement and evaluate three reward modeling paradigms: scalar, semi-scalar, and generative reward models, alongside existing SQA approaches. Our experiments reveal three key findings: (1) scalar models achieve the strongest overall performance, consistently exceeding 74% accuracy; (2) score regression-based approaches generally underperform preference-based methods in both overall performance and generalization; and (3) all reward models struggle on pairs with very small MOS gap. Motivated by these observations, we propose a MOS-aware GRM design that incorporates MOS gap into the reward function during reinforcement learning. Experimental results show that the MOS-aware GRM significantly improves fine-grained speech quality discrimination. We hope this work fosters more rigorous and scalable research in SQA.