Multi-hop question answering (MHQA) enables accurate answers to complex queries by retrieving and reasoning over evidence dispersed across multiple documents. Existing MHQA approaches mainly rely on iterative retrieval-augmented generation, which suffer from the following two major issues. 1) Existing methods prematurely commit to surface-level entities rather than underlying reasoning structures, making question decomposition highly vulnerable to lexical ambiguity. 2) Existing methods overlook the logical dependencies among reasoning steps, resulting in uncoordinated execution. To address these issues, we propose STRIDE, a framework that separates strategic planning, dynamic control, and grounded execution. At its core, a Meta-Planner first constructs an entity-agnostic reasoning skeleton to capture the abstract logic of the query, thereby deferring entity grounding until after the reasoning structure is established, which mitigates disambiguation errors caused by premature lexical commitment. A Supervisor then orchestrates sub-question execution in a dependency-aware manner, enabling efficient parallelization where possible and sequential coordination when necessary. By dynamically deciding whether to retrieve new evidence or infer from existing facts, it avoids redundant queries and error propagation, while fusing cross-branch information and reformulating failed queries to enhance robustness. Grounded fact extraction and logical inference are delegated to specialized execution modules, ensuring faithfulness through explicit separation of retrieval and reasoning. We further propose STRIDE-FT, a modular fine-tuning framework that uses self-generated execution trajectories from STRIDE, requiring neither human annotations nor stronger teacher models. Experiments show that STRIDE achieves robust and accurate reasoning, while STRIDE-FT effectively enhances open-source LLMs.

Recent years have witnessed the rapid development of Large Language Model-based Multi-Agent Systems (MAS), which excel at collaborative decision-making and complex problem-solving. Recently, researchers have further investigated Multi-Agent Debate (MAD) frameworks, which enhance the reasoning and collaboration capabilities of MAS through information exchange and debate among multiple agents. However, existing approaches often rely on unguided initialization, causing agents to adopt identical reasoning paths that lead to the same errors. As a result, effective debate among agents is hindered, and the final outcome frequently degenerates into simple majority voting. To solve the above problem, in this paper, we introduce Dynamic Multi-Agent Debate (DynaDebate), which enhances the effectiveness of multi-agent debate through three key mechanisms: (1) Dynamic Path Generation and Allocation, which employs a dedicated Path Generation Agent to generate diverse and logical solution paths with adaptive redundancy; (2) Process-Centric Debate, which shifts the focus from surface-level outcome voting to rigorous step-by-step logic critique to ensure process correctness; (3) A Trigger-Based Verification Agent, which is activated upon disagreement and uses external tools to objectively resolve deadlocks. Extensive experiments demonstrate that DynaDebate achieves superior performance across various benchmarks, surpassing existing state-of-the-art MAD methods.

Automated feature generation extracts informative features from raw tabular data without manual intervention and is crucial for accurate, generalizable machine learning. Traditional methods rely on predefined operator libraries and cannot leverage task semantics, limiting their ability to produce diverse, high-value features for complex tasks. Recent Large Language Model (LLM)-based approaches introduce richer semantic signals, but still suffer from a restricted feature space due to fixed generation patterns and from the absence of feedback from the learning objective. To address these challenges, we propose a Memory-Augmented LLM-based Multi-Agent System ( MALMAS ) for automated feature generation. MALMAS decomposes the generation process into agents with distinct responsibilities, and a Router Agent activates an appropriate subset of agents per iteration, further broadening exploration of the feature space. We further integrate a memory module comprising procedural memory, feedback memory, and conceptual memory, enabling iterative refinement that adaptively guides subsequent feature generation and improves feature quality and diversity. Extensive experiments on multiple public datasets against state-of-the-art baselines demonstrate the effectiveness of our approach.

Reward models (RMs) play a critical role in enhancing the reasoning performance of LLMs. For example, they can provide training signals to finetune LLMs during reinforcement learning (RL) and help select the best answer from multiple candidates during inference. In this paper, we provide a systematic introduction to RMs, along with a comprehensive survey of their applications in LLM reasoning. We first review fundamental concepts of RMs, including their architectures, training methodologies, and evaluation techniques. Then, we explore their key applications: (1) guiding generation and selecting optimal outputs during LLM inference, (2) facilitating data synthesis and iterative self-improvement for LLMs, and (3) providing training signals in RL-based finetuning. Finally, we discuss critical open questions regarding the selection, generalization, evaluation, and enhancement of RMs, based on existing research and our own empirical findings. Our analysis aims to provide actionable insights for the effective deployment and advancement of RMs for LLM reasoning.

Graph Foundation Models (GFMs), built upon the paradigm, have emerged as a research hotspot in graph learning. For GNN-based GFMs, graph prompt tuning has become the prevailing adaptation method for downstream tasks. Although recent methods explain why graph prompt tuning works, how to rigorously measure its adaptation capacity remains an open problem. Addressing this problem is critical for understanding the capability limits of graph prompt tuning and for developing more powerful adaptation methods. In this paper, we propose Prismatic Space Theory (PS-Theory), a novel mathematical framework to quantify the capacity of adaptation methods, while focusing on establishing the upper bound for the adaptation capacity of graph prompt tuning. Building upon the proposed PS-Theory, we further introduce essage uning for FMs (MTG), a lightweight approach that injects a small set of learnable message prototypes into each layer of the GNN backbone to adaptively guide message fusion without updating pre-trained weights. Through our PS-Theory, we prove that the adaptation capacity of MTG can exceed the theoretical upper bound of graph prompt tuning. Extensive experiments demonstrate that MTG consistently outperforms graph prompt baselines across diverse benchmark datasets, providing strong empirical support for our theoretical findings. Our code is available at https://anonymous.4open.science/r/MTG.

Knowledge-based Visual Question Answering (KB-VQA) aims to answer queries that necessitate reasoning over external knowledge sources beyond the visual content. Typically, current methods fuse multimodal features to retrieve external information, subsequently leveraging Multimodal Large Language Models (MLLMs) to derive answers from the retrieved evidence. However, these methods often struggle to capture structural associations within complex contexts to effectively filter noise. Furthermore, they frequently fail to ensure that the reasoning process remains strictly faithful to the retrieved evidence. To address these challenges, we propose SAFE-G, a Structure-Aware Faithful Evidence-guided Generation framework, which enables precise evidence localization and trustworthy reasoning. Specifically, we first employ a coarse-grained hybrid search fusing visual and textual modalities to recall candidate documents, and subsequently implement a structure-aware fine-grained graph retrieval that captures structural dependencies to filter noise and pinpoint precise evidence. Moreover, we introduce a reinforcement learning (RL) strategy with an evidence-grounded reward that assigns credit to correct answers only when the selected evidence is correct. This strict alignment constraint compels the model to anchor its response in the retrieved context, effectively enhancing its capability to locate evidence via multimodal features and perform faithful reasoning. Extensive experiments on the Encyclopedic-VQA and InfoSeek benchmarks demonstrate that SAFE-G outperforms prior methods by a margin of 8.9

Large language model (LLM) agents are increasingly deployed in user-facing services that require iterative tool use under dynamic business conditions. Reliable evaluation is essential for sustained improvement: it must reveal capability deficiencies, inform priorities, and assess interventions. Yet industrial agent service unfolds both through the iterative trajectory of a current request and through continued user interaction. Final-outcome assessment can therefore obscure where deficiencies arise and whether later service remains aligned with context from earlier exchanges. We propose ATLAS, a dual-horizon diagnostic evaluation framework for industrial tool-use agents. At the request horizon, trajectory-wise diagnostic signals relate deficiencies to execution locations and capability concerns. At the interaction horizon, user-wise signals assess whether service remains responsive across continued interaction. Together, these views provide structured diagnostic evidence for analyzing execution deficiencies and sustained service behavior. ATLAS instantiates them as executable signals with explicit evidence scopes and decision boundaries. LLM judge interfaces are calibrated against high-confidence references from real business logs; when needed, their decision behavior is distilled into efficient diagnostic models for lower-latency, lower-cost evaluation. The resulting feedback supports policy optimization. We evaluate ATLAS on Meituan Xiaotuan production traffic. Offline experiments assess diagnostic-signal fidelity and replay-based policy improvement, while online A/B experiments show concurrent gains in user engagement, downstream business outcomes, and sampled human-audit quality.

Lifelong Model Editing aims to continuously update evolving facts in Large Language Models while preserving unrelated knowledge and general capabilities, yet it remains plagued by catastrophic forgetting and model collapse. Empirically, we find that the few recent editors resilient over long horizons share the same core strategy: , which normalizes value gradients using running statistics. Removing LN causes immediate collapse, and we observe a counter-intuitive where early edits can facilitate later edits. This suggests that with LN, early edits can stabilize the model and promote the success of future edits. Yet the mechanism of LN remains a "black box", leaving its precise role in lifelong stability poorly understood. In this work, we provide the theoretical account of LN in the lifelong regime. Our analysis reveals a self-reinforcing stability loop and proves that, when combined with ridge-regularized regression, LN yields updates with and , directly mitigating forgetting and systemic collapse. Based on these insights, we derive , which strengthens this stability loop via an explicit warm-up stage and full whitening, improving long-horizon stability at minimal overhead. Extensive experiments validate our theory and demonstrate competitive performance.

Empirical Risk Minimization (ERM) models often rely on spurious correlations between features and labels during the learning process, leading to shortcut learning behavior that undermines robustness generalization performance. Current research mainly targets identifying or mitigating a single shortcut; however, in real-world scenarios, cues within the data are diverse and unknown. In empirical studies, we reveal that models rely more on strong shortcuts than weak ones, with their performance under multiple shortcuts typically falling between that of an individual shortcut. To address these challenges, we propose MiMu, a novel method integrated with Transformer-based ERMs designed to Mitigate Multiple shortcut learning behavior, which incorporates self-calibration strategy and self-improvement strategy. In the source model, we first propose the self-calibration strategy to prevent the model from relying on shortcuts and make overconfident predictions. Then, we design self-improvement strategy in target model to further reduce the reliance on multiple shortcuts. The random mask strategy involves randomly masking partial attention positions to diversify the focus of target model avoiding fixation on a fixed region. Meanwhile, the adaptive attention alignment module facilitates the alignment of attention weights to the calibrated source model, without the need for post-hoc attention maps or supervision. Finally, extensive experiments conducted on Natural Language Processing (NLP) and Computer Vision (CV) demonstrate the effectiveness of MiMu in improving the robustness generalization abilities.

Large vision-language models (VLMs) have demonstrated remarkable capabilities in open-world multimodal understanding, yet their high computational overheads pose great challenges for practical deployment. Some recent works have proposed methods to accelerate VLMs by pruning redundant visual tokens guided by the attention maps of VLM's early layers. Despite the success of these token pruning methods, they still suffer from two major shortcomings: (i) considerable accuracy drop due to insensitive attention signals in early layers, and (ii) limited speedup when generating long responses (e.g., 30 tokens). To address the limitations above, we present TwigVLM – a simple and general architecture by growing a lightweight module, named twig, upon an early layer of the base VLM. Compared with most existing VLM acceleration methods purely based on visual token pruning, our TwigVLM not only achieves better accuracy retention by employing a twig-guided token pruning (TTP) strategy, but also yields higher generation speed by utilizing a self-speculative decoding (SSD) strategy. Taking LLaVA-1.5-7B as the base VLM, experimental results show that TwigVLM preserves 96

Large Language Models fine-tuned with Reinforcement Learning from Human Feedback (RLHF-LLMs) can over-rely on aligned preferences without truly gaining self-knowledge, leading to hallucination and biases. If an LLM can better access its knowledge and know what it knows, it can avoid making false or unsupported claims. Therefore, it is crucial to evaluate whether LLMs have the ability to know what they know, as it can help to ensure accuracy and faithfulness in real-world applications. Inspired by research in Educational Psychology, surface learners who don’t really know are easily affected by teacher and peer guidance, we treat LLM as a student, incorporate role guidance in prompts to explore whether LLMs really know. Specifically, we propose a novel strategy called Role-Guided and Self-Reflection (RoSe) to fully assess whether LLM “knows it knows”. We introduce multiple combinations of different roles and strong reminder in prompts combined with self-reflection to explore what local information in prompt LLMs rely on and whether LLMs remain unaffected by external guidance with varying roles. Our findings reveal that LLMs are very sensitive to the strong reminder information. Role guidance can help LLMs reduce their reliance on strong reminder. Meanwhile, LLMs tend to trust the role of authority more when guided by different roles. Following these findings, we propose a double-calibrated strategy with verbalized confidence to extract well-calibrated data from closed-source LLM and fine-tune open-source LLMs. Extensive experiments conducted on fine-tuning open-source LLMs demonstrate the effectiveness of double-calibrated strategy in mitigating the reliance of LLMs on local information. For a thorough comparison, we not only employ public JEC-QA and openBookQA datasets, but also construct EG-QA which contains English Grammar multiple-choice question-answering and 14 key knowledge points for assessing self-knowledge and logical reasoning.

Lifelong model editing aims to dynamically adjust a model’s output with respect to specific facts, knowledge points, or behaviors, enabling the model to adapt to the ever-changing demands of the real world without requiring retraining. While some retrieval-based methods have demonstrated potential in lifelong editing scenarios by storing edited knowledge in external memory, they often suffer from limitations in usability, such as requiring additional training corpora or lacking support for reversible and detachable edits.To address these issues, we propose a plug-and-play method for knowledge retrieval and storage, i.e., Layer-Level Prompting (LLP), which enables seamless and efficient lifelong model editing. In our LLP framework, the reasoning process of LLMs is divided into two stages, respectively knowledge retrieval (Think) and knowledge injection(Recall). Specifically, the knowledge retrieval process is performed in the early layers of the model. Based on the retrieved information, the model is guided to access the updated knowledge stored in the subsequent layer to complete the knowledge editing process. Experimental results demonstrate that our method consistently outperforms existing techniques on lifelong model editing tasks, achieving superior performance on question answering and hallucination benchmarks across different LLMs.

Multi-hop question answering is a challenging task that requires capturing information from different positions in multiple documents. Recently, several methods propose to enhance Large Language Models (LLMs) by incorporating structured knowledge, aiming to grasp key information for solving this task. Despite certain achievements, they still face the following challenges: 1) The neglect of text-based reasoning capabilities. 2) Information redundancy between text and triples. 3) Information loss during structured knowledge extraction. To solve the above challenges, in this paper, we propose Dynamic Combination of Structured Knowledge (DCSK), a novel framework for integrating text-based and triple-based paradigms. Following Occam’s Razor, DCSK dynamically determine the necessity of structured knowledge by the designed multi-faceted evaluation, which systematically assess the correctness, clarity, and informativeness of text-based prediction. For questions that require structured knowledge, we develop an iterative fact refiner that screens for question-relevant triples, verifies their factual adequacy, and thereby effectively excludes irrelevant and redundant information. Furthermore, based on the verification, we construct an adaptive knowledge reasoner that dynamically adjusts the need for text supplementation, thus mitigating the information deficiency in selected triples. Extensive experiments on three MHQA datasets demonstrate the efficiency and effectiveness of DCSK.

Capturing human learning behavior based on deep learning methods has become a major research focus in both psychology and intelligent systems. Recent approaches rely on controlled experiments or rule-based models to explore cognitive processes. However, they struggle to capture learning dynamics, track progress over time, or provide explainability. To address these challenges, we introduce LearnerAgent, a novel multi-agent framework based on Large Language Models (LLMs) to simulate a realistic teaching environment. To explore human-like learning dynamics, we construct learners with psychologically grounded profiles-such as Deep, Surface, and Lazy-as well as a persona-free General Learner to inspect the base LLM's default behavior. Through weekly knowledge acquisition, monthly strategic choices, periodic tests, and peer interaction, we can track the dynamic learning progress of individual learners over a full-year journey. Our findings are fourfold: 1) Longitudinal analysis reveals that only Deep Learner achieves sustained cognitive growth. Our specially designed "trap questions" effectively diagnose Surface Learner's shallow knowledge. 2) The behavioral and cognitive patterns of distinct learners align closely with their psychological profiles. 3) Learners' self-concept scores evolve realistically, with the General Learner developing surprisingly high self-efficacy despite its cognitive limitations. 4) Critically, the default profile of base LLM is a "diligent but brittle Surface Learner"-an agent that mimics the behaviors of a good student but lacks true, generalizable understanding. Extensive simulation experiments demonstrate that LearnerAgent aligns well with real scenarios, yielding more insightful findings about LLMs' behavior.

Deep Neural Networks (DNNs), despite their notable progress across information retrieval tasks, encounter the issues of shortcut learning and struggle with poor generalization due to their reliance on spurious correlations between features and labels. Current research mainly mitigates shortcut learning behavior using augmentation and distillation techniques, but these methods could be laborious and introduce unwarranted biases. To tackle these, in this paper, we propose COMI, a novel method to COrrect and MItigate shortcut learning behavior. Inspired by the ways students solve shortcuts in educational scenarios, we aim to reduce model's reliance on shortcuts and enhance its ability to extract underlying information integrated with standard Empirical Risk Minimization (ERM). Specifically, we first design Correct Habit (CoHa) strategy to retrieve the top.. challenging samples for priority training, which encourages model to rely less on shortcuts in the early training. Then, to extract more meaningful underlying information, the information derived from ERM is separated into task-relevant and task-irrelevant information, the former serves as the primary basis for model predictions, while the latter is considered non-essential. However, within task-relevant information, certain potential shortcuts contribute to overconfident predictions. To mitigate this, we design Deep Mitigation (DeMi) network with shortcut margin loss to adaptively control the feature weights of shortcuts and eliminate their influence. Besides, to counteract unknown shortcut tokens issue in NLP, we adopt locally interpretable module-LIME to help recognize shortcut tokens. Finally, extensive experiments conducted on NLP and CV tasks demonstrate the effectiveness of COMI, which can perform well on both IID and OOD samples.