Large language models (LLMs) reach state-of-the-art performance across many NLP tasks, but their large parameter counts introduce heavy computational and memory overhead, which complicates deployment in resource-constrained settings. Pruning is a standard compression strategy that induces sparsity to lower these costs. However, most pruning methods for LLMs depend on calibration data and expensive weight updates, which limits practical scalability. To address these limitations, we introduce Haar Wavelet Subband Pruning (), a post-training framework that requires no calibration data and no weight updates. applies a two-dimensional Haar wavelet transform to each weight matrix and decomposes it into four frequency subbands. It then assigns a uniform sparsity ratio to all subbands so that both low- and high-frequency components are retained in a balanced manner. Our theoretical analysis shows that the subband design of provides a deterministic per-subband retention guarantee, which helps mitigate the potential bias of global magnitude pruning toward dominant frequency components. Experiments on the LLaMA, OPT and Qwen model families show that achieves competitive accuracy relative to strong pruning baselines while substantially reducing pruning time. Compared with magnitude pruning, which serves as a simple calibration-free baseline, generally achieves better downstream performance across a wide range of sparsity levels and model scales.

Large language models (LLMs) rely on tool use to act as autonomous agents, yet often fail in multi-step execution due to insufficient tool-related knowledge and ineffective knowledge activation. Therefore, we present a systematic study on how knowledge influences tool-use performance, covering the stages of knowledge acquisition, activation, and internalization. In the knowledge acquisition stage, we acquire and evaluate various forms of experiential knowledge, and our analysis shows that simple instance-level knowledge can already provide strong and reliable gains, while abstract intent-level knowledge offers limited benefits. At inference time, to activate knowledge, we find that prompting LLM to expand the depth of reasoning yields diminishing returns, whereas expanding the width of reasoning by parallel sampling with aggregation more effectively activates latent experiential knowledge. At training time, for knowledge internalization, post-training with knowledge-augmented data further improves performance, with reinforcement learning outperforming supervised fine-tuning. Based on these insights, we propose the Knowledge-Augmented Tool Execution (KATE), a knowledge-augmented tool execution framework that integrates experiential knowledge with reasoning-width-expanded inference and knowledge-aware training. Experiments on BFCL-V3 and AppWorld demonstrate consistent and substantial improvements over strong baselines across model scales. Our Code is available at https://github.com/hypasd-art/KATE.

Temporal knowledge graph forecasting(TKGF) asks a model to rank the mostplausible future entity for a query such as(s, r, ?, t) from historical events. Recenttraining-free methods use large languagemodels (LLMs) for this task, but their accuracydepends heavily on which past events areshown in the prompt under a tight contextbudget. We present LANTERN, a training-freeprompting framework that addresses thisbottleneck by combining two complementaryviews of history: a long-window strengthscore for stable interaction patterns anda short-window novelty score for suddenchanges. LANTERN first filters unhelpfulevents, then selects a compact evidence setwith Pareto-greedy selection, and finally addsone structure-aware analogical demonstration.Across ICEWS14, ICEWS05-15, ICEWS18,and GDELT, LANTERN consistently outperforms the state-of-the-art training-free baselineAnRe under the same backbone and 2-hopcandidate protocol, improving Hits@1 by upto 2.5 points and MRR by up to 1.2 points.

As the parameter size of language models continues to grow, effective model compression is required to reduce their computational and memory overhead. Existing compression methods suffer from bottleneck issues: when the compression ratio is increased, performance degrades significantly. Low-rank decomposition and quantization are two prominent compression methods that have been proven to significantly reduce the computational and memory requirements of Large Language Models (LLMs) while maintaining model accuracy. Evidently, combining these two methods will break through the existing compression bottleneck. However, how these two methods interact when combined remains a critical question for developers, as many assume they are orthogonal, meaning their combination would not introduce additional errors beyond those independently introduced by each method. This paper provides the first mathematical proof that low-rank decomposition and quantization are non-orthogonal. We validate these findings through a series of experiments on large language models. Our results demonstrate that these methods are non-orthogonal, and their combination leads to significant performance degradation. Importantly, we propose a novel approach Diagonal Adhesive Method (DAM), which can effectively combine the two methods and mitigate the performance loss. Our research provides deep insights into model compression and lays a solid theoretical and experimental foundation for future related studies.

Large language model (LLM) agents increasingly rely on accumulated memory to solve long-horizon decision-making tasks. However, most existing approaches store memory in fixed representations and reuse it at a single or implicit level of abstraction, which limits generalization and often leads to negative transfer when distribution shift. This paper proposes the Meta-Cognitive Memory Abstraction method (MCMA), which treats memory abstraction as a learnable cognitive skill rather than a fixed design choice. MCMA decouples task execution from memory management by combining a frozen task model with a learned memory copilot. The memory copilot is trained using direct preference optimization; it determines how experience should be structured, abstracted, and reused. Memories are further organized into a hierarchy of abstraction levels, enabling selective reuse based on task similarity. When no memory is transferable, MCMA transfers the ability to abstract and manage memory by transferring the memory copilot. Experiments on ALFWorld, ScienceWorld, and BabyAI demonstrate substantial improvements in performance, out-of-distribution generalization, and cross-task transfer over several baselines.

Post-Training Quantization (PTQ) is a critical strategy for efficient large language models (LLMs) deployment. However, existing scaling laws primarily focus on general performance, overlooking crucial fine-grained factors and how quantization differentially impacts diverse knowledge capabilities. To address this, we establish Task-Stratified Knowledge Scaling Laws. By stratifying capabilities into memorization, application, and reasoning, we develop a framework that unifies model size, bit-width, and fine-grained factors: group size and calibration set size. Validated on 293 diverse PTQ configurations, our framework demonstrates strong fit and cross-architecture consistency. It reveals distinct sensitivities across knowledge capabilities: reasoning is precision-critical, application is scale-responsive, and memorization is calibration-sensitive. We highlight that in low-bit scenarios, optimizing these fine-grained factors is essential for preventing performance collapse. These findings provide an empirically-backed foundation for designing knowledge-aware quantization strategies.

Post-Training Quantization (PTQ) is critical for the efficient deployment of Large Language Models (LLMs). While 4-bit quantization is widely regarded as an optimal trade-off, reducing the precision to 2-bit usually triggers a catastrophic “performance cliff.” It remains unclear whether the underlying mechanisms differ fundamentally. Consequently, we conduct a systematic mechanistic analysis, revealing two qualitatively distinct failure modes: Signal Degradation, where the computational patterns remain intact but information precision is impaired by cumulative error; and Computation Collapse, where key components fail to function, preventing correct information processing and destroying the signal in the early layers. Guided by this diagnosis, we conduct mechanism-aware interventions, demonstrating that targeted, training-free repair can mitigate Signal Degradation, but remains ineffective for Computation Collapse. Our findings provide a systematic diagnostic framework for PTQ failures and suggest that addressing Computation Collapse requires structural reconstruction rather than mere compensation.

Large Language Models (LLMs) have shown great promise in tool-making, yet existing frameworks often struggle to efficiently construct reliable toolsets and are limited to single-task settings. To address these challenges, we propose GATE (Graph-based Adaptive Tool Evolution), an adaptive framework that dynamically constructs and evolves a hierarchical graph of reusable tools across multiple scenarios. We evaluate GATE on open-ended tasks (Minecraft), agent-based tasks (TextCraft, DABench), and code generation tasks (MATH, Date, TabMWP). Our results show that GATE achieves up to 4.3× faster milestone completion in Minecraft compared to the previous state-of-the-art method, and provides an average improvement of 9.23% over existing tool-making methods in code generation tasks and 10.03% in agent tasks. Further analysis shows that GATE exhibits strong adaptive evolution capabilities, effectively balancing tool quantity, complexity, and functionality while maintaining high efficiency. Code and data are available at https://github.com/ayanami2003/GATE.

Reinforcement Learning with Verifiable Rewards (RLVR) is a promising approach for enhancing agentic search. However, its performance is often hindered by reward sparsity, whereby agents receive very limited positive feedback despite incurring significant exploration costs. In this paper, we formalize this challenge as a new research problem termed Reward Density Optimization, which aims to improve the reward obtained per unit of exploration cost. To address this problem, we introduce InfoFlow, a systematic framework that operates along three complementary dimensions: 1) Sub-goal Scaffolding: which decomposes long-horizon tasks into intermediate objectives and assigns process-level rewards to provide denser learning signals; 2) Pathfinding Hints: which injects corrective guidance into stalled trajectories to increase the ratio of successful trials; and 3) Dual-agent Refinement: which employs a dual-agent architecture to offload the cognitive burden of deep exploration. We evaluate InfoFlow on several popular agentic search benchmarks, where it significantly outperforms strong baselines and enables lightweight LLMs to achieve performance comparable to that of advanced proprietary models.

Chain-of-Thought (CoT) has become a standard method for improving reasoning capabilities in large language models (LLMs) by eliciting step-by-step thinking, but its effectiveness in multimodal tasks remains unclear. In this paper, we aim to systematically investigate the key question: What can multimodal Chain-of-Thought reasoning do, and where and why does it fall short? To this end, we evaluate 12 multimodal tasks across perception and reasoning categories using both 14 non-reasoning models and 8 reasoning models. Our analysis reveals several important findings: (1) CoT is not a free lunch and should be used selectively depending on the specific requirements of each task. For perception tasks, CoT can lead to undesirable side effects, such as reduced performance in visual grounding and object counting. In contrast, it proves effective for reasoning tasks involving mathematical, scientific, and multi-image reasoning; (2) Compared to original models, existing open-source multimodal reasoning models often yield only marginal overall improvements, possibly due to an overemphasis on mathematical reasoning at the expense of broader capabilities; (3) Visual reasoning remains a key bottleneck for current multimodal CoT, as models exhibit a Look Light, Think Heavy” pattern where verbal reflection rises and falls during reasoning, whereas visual reflection consistently diminishes. These findings suggest that while multimodal CoT handles verbal reflection relatively well, it lacks the ability to maintain deep visual introspection throughout the reasoning process.

Continual Learning (CL) for Large Language Models (LLMs) faces a fundamental Stability-Plasticity Dilemma: balancing the plasticity to acquire new capabilities with the stability to preserve prior knowledge. While Parameter-Efficient Fine-Tuning methods, such as LoRA, enable efficient adaptation, we identify a critical flaw in current approaches termed Rank-Blindness: the enforcement of a single rank constraint across diverse tasks, which entangles task-shared and task-specific knowledge, leading to catastrophic forgetting of earlier tasks and underfitting on complex new ones. To address this, we propose SpaRTA, a novel rehearsal-free framework guided by a rank-spectrum perspective that explicitly disentangles knowledge into two orthogonal subspaces. Specifically, SpaRTA employs a low-rank branch to capture task-shared representations and a high-rank branch to model task-specific features. To integrate these complementary representations, we introduce a context-aware dynamic router that adaptively fuses the two branches based on input semantics, while an explicit orthogonality constraint minimizes interference between shared and specific parameter subspaces. This design effectively isolates task-specific updates from shared knowledge, preventing the overwriting of prior capabilities while preserving strong adaptation capacity. Extensive experiments demonstrate that SpaRTA achieves a superior stability-plasticity balance compared to single-rank baselines. Notably, the proposed spectral disentanglement strategy substantially reduces inter-task interference and yields strong zero-shot generalization on unseen tasks. Our code will be available at https://github.com/Xnhyacinth/SpaRTA.

Post-training quantization has emerged as a widely adopted technique for compressing and accelerating the inference of Large Language Models (LLMs). The primary challenges in LLMs quantization stem from activation outliers, which significantly degrade model performance especially at lower bit precision. While recent approaches attempt to mitigate outliers through linear transformations across feature dimensions, our analysis reveals that the transformed weights and activations still exhibit persistent outlier patterns with concentrated magnitude distributions. In this paper, we first model the mathematical relationship between quantization error and outliers, and then introduce a new metric Flatness to quantify the distribution of outliers. Based on this, we derive the theoretical optimal solution with respect to Flatness. Building on these insights, we propose Bidirectional Diagonal Quantization (BDQ), a novel post-training quantization framework that effectively disperses outlier patterns through optimized matrix transformations. BDQ strategically distributes outlier magnitudes across matrix dimensions via learned diagonal operations. Extensive experiments demonstrate that BDQ establishes a new quantization benchmark. It achieves less than 1% accuracy drop in W4A4 quantization on the LLaMA-3-8B model. In the more challenging W2A4KV16 experiment, compared to state-of-the-art approaches, BDQ reduces the performance gap by 39.1% on the DeepSeek-R1-Distill-LLaMA-70B model.

Explainable diagnosis requires that authoritative medical knowledge provide the rationales linking a patient’s clinical manifestations to the diagnostic conclusion. Although large language models (LLMs) hold great potential to facilitate explainable diagnosis, their effectiveness is often constrained by insufficient diagnostic expertise. To address this limitation, we propose Self-learned Explainable Knowledge Augmented Diagnosis (SEKAD), a unified LLM-based framework for faithful and explainable diagnosis. Our approach builds a high-quality diagnostic knowledge base through a record-driven explanation learning paradigm, as well as applies this knowledge via an explanation-based diagnostic process that ensures faithful inference. Experiments on the DiReCT and JAMA benchmarks show that SEKAD consistently outperforms strong baselines across the metrics. In particular, on the DiReCT benchmark, SEKAD improves the explanation completeness metric from 64.5% to 76.9% over the best existing methods, highlighting its effectiveness in enhancing diagnostic explainability and showing that our text mining approach produces knowledge that is both reliable in quality and large in quantity.

Multimodal web agents can assist humans in operating repetitive GUI tasks, where effective task planning is essential for decomposing complex tasks into executable actions. While small open-source MLLMs are cost-efficient and privacy-preserving compared with commercial large models, they suffer from weak planning and limited cross-website generalization. To address these limitations, we introduce the planning experience exploration and utilization (PEEU) method, which autonomously explores environments to discover experiences and utilizes hindsight experience to synthesize strictly aligned, high-level training data. To quantitatively analyze the generalization behaviors driving this performance, we propose the task decomposition hierarchical analysis framework (TDHAF) to systematically study compositional generalization across three task granularities: low, middle and high levels. Our analysis reveals that mastering low-level atomic skills does not guarantee high-level planning competence, while high-level task training yields stronger OOD generalization. Experiments on real-world benchmarks demonstrate PEEU’s superior effectiveness: our 7B model achieves 30.6% accuracy, outperforming the much larger Qwen2.5-VL-32B model. These demonstrate constructing hindsight high-level tasks and leveraging experiences is crucial for OOD planning abilities of small MLLMs.

Large language models (LLMs) can carry out human-like dialogue, but unlike humans, they are stateless due to the superposition property. However, during multi-turn, multi-agent interactions, LLMs begin to exhibit consistent, character-like behaviors—hinting at a form of emergent lifelong learning. Despite this, existing benchmarks often fail to capture these dynamics, primarily focusing on static, open-ended evaluations. To address this gap, we introduce LifeState-BENCH, a benchmark designed to assess lifelong learning in LLMs. It features two episodic datasets—Hamlet and a synthetic script collection—rich in narrative structure and character interactions. Our fact-checking evaluation probes models’ self-awareness, episodic memory retrieval, and relationship tracking, across both parametric and non-parametric approaches. Experiments on models like Llama3.1-8B, GPT-4-turbo, and DeepSeek R1, we demonstrate that non-parametric methods significantly outperform parametric ones in managing stateful learning. However, all models exhibit challenges with catastrophic forgetting as interactions extend, highlighting the need for further advancements in lifelong learning.

LLM-based Multi-agent systems (MAS) have shown strong capabilities across a wide range of domains. Their success largely hinges on the collaboration topology design, which has emerged as a central research focus in the automated MAS design.However, existing approaches are fundamentally constrained by their reliance on homogeneous LLMs, which significantly limits overall system intelligence.In response to this limitation, we for the first time propose the concept of Automated Design of Heterogeneous-LLMs-based MAS (ADHM).ADHM sheds light on a promising avenue for advancing collective intelligence, which focuses on the automated design of cost-effective MAS composed of diverse LLMsand roles to suit various queries.Toward this challenging goal, we propose Hetero-Designer, a novel pipeline that efficiently encodes intricate dependencies among queries, LLMs and roles through a novel Binary-Star Transformer and constructs Hetero-MAS in an autoregressive graph generation process. Extensive experiments demonstrate that Hetero-Designer is: (1) high-performing on various benchmarks, (2) economical in reducing overhead, (3) extensible to unseen LLMs and roles.

This paper notices that while symbolic instruction and neural parameters play different roles on steering LLMs’ behavior, both instructions and parameters are the compression of task data, they are supposed be strongly correlated and can be learned to predict one from the other. Therefore, This paper proposes a novel neural network framework, SHIP (Shuttle between the Instructions and the Parameters), to model and learn the bi-directional mappings between the instructions and the parameters of LLMs. We verify that SHIP can effectively map one of the instructions/parameters to the other by evaluating it on the tasks of instruction deduction and induction. The results show that SHIP performs better than existing baseline methods in terms of deductive capabilities while significantly surpassing them in inductive capabilities. Moreover, SHIP can effectively combine the two mapping processes to perform excellent inductive reasoning. We further discuss how the latent fusing methods and latent dimensions affect SHIP’s performance, and show SHIP can effectively generalize with pre-training. The code and data for this paper are released at https://anonymous.4open.science/r/Shuttle-Between-Instructions-Parameters

Enabling Large Language Models (LLMs) to evolve sustainably requires simultaneously preserving previously acquired knowledge (Past), effectively acquiring new task-specific skills (Present), and reserving sufficient parameter capacity for subsequent adaptation (Future). However, existing continual learning (CL) paradigms often prioritize immediate performance through dense updates, leading to catastrophic forgetting and rapid exhaustion of model capacity. To harmonize these conflicting demands, we draw inspiration from the brain’s functional partitioning and propose the Null-Space Constrained Parameter Region Specificity Method (PaRSP). PaRSP establishes a dynamic “Task-Region Mapping” that distinguishes between specialized neurons and generalist neurons. By precisely localizing a sparse “functional core” for each task, PaRSP restricts updates to specific regions via null-space orthogonality, preserving the vast majority of the network as an immutable “long-term memory bank.” This induced sparsity not only enhances plasticity via targeted adaptation and minimizes interference to ensure stability, but also strategically reserves substantial capacity, securing sustainability for future evolution. Extensive experiments validate PaRSP’s state-of-the-art performance, particularly on Standard CL and Long Sequence benchmarks, effectively harmonizing the stability-plasticity-sustainability trade-off. Code is available at https://github.com/JinhuiBot/PaRSP

Retrieval-Augmented Generation (RAG) has become a standard paradigm for grounding Large Language Models (LLMs) with external knowledge. However, RAG performance often degrades substantially when faced with noisy, outdated, or conflicting retrieved information. In this work, we empirically demonstrate that Prior-Guided Reasoning—a strategy that explicitly elicits the model’s parametric knowledge as prior information to guide reasoning on retrieved documents—effectively mitigates the impact of external conflicts. Building on this, we propose BrPr (Bernoulli-gated reinforcement learning for Prior-Guided reasoning), a framework that achieves robust performance across varying degrees of external inconsistency. Furthermore, by employing a Bernoulli-gated dropout mechanism during training, BrPr distills the prior-driven reasoning capability into the model parameters, enabling efficient latent reasoning without explicit prior generation. The experimental results demonstrate that BrPr consistently exhibits superior robustness to external conflicts and noise.

Retrieval-augmented generation (RAG) plays a critical role in user-generated content (UGC) platforms, but its effectiveness critically depends on accurate query–document relevance assessment. Despite recent advances in applying large language models (LLMs) to relevance modeling, UGC platforms present unique challenges: 1) ambiguous user intent due to sparse user feedback in RAG scenarios, and 2) asymmetric relevance, where relevance is driven by localized answer-bearing content rather than global query–document similarity. To address these issues, we propose the Reinforced Reasoning model for Relevance Assessment (R³A), which decomposes relevance assessment into intent inference and evidence grounding. R³A leverages auxiliary high-clicked documents to infer latent query intent, and extracts verbatim evidence fragments to ground relevance decisions, reducing noise sensitivity and improving asymmetric relevance modeling. Experimental results demonstrate that R³A substantially outperforms strong baselines on offline benchmarks, while the distilled R³A-1.5B model achieves significant gains in large-scale online A/B testing, effectively balancing performance and practical deployability.

Large language models (LLMs) have achieved remarkable success but still tend to generate factually erroneous responses, a phenomenon known as hallucination. A recent trend is to use preference learning to fine-tune models to align with factuality. However, existing work primarily evaluates fine-tuned models on in-domain (ID) datasets and the factuality on out-of-domain (OOD) datasets remains underexplored. In this paper, we conduct a comprehensive evaluation of the factuality of different models tuned by various preference learning algorithms and demonstrate that their performance on OOD datasets either increases minimally or decreases. Subsequently, we reveal that the main cause of model’s failure to uphold factuality under a distribution shift is under-alignment, rather than over-alignment, by analyzing the token distribution shift of the models before and after tuning. Finally, we propose APEFT (Atomic Preference Enhanced Factuality Tuning), a framework that enhances model’s awareness of factuality at the granularity of individual facts. Extensive experiments demonstrate that APEFT improves model performance by an average of on both ID and OOD datasets, which is highly effective.

Large language models (LLMs) based Multilingual Knowledge Graph Completion (MKGC) aim to predict missing facts by leveraging LLMs’ multilingual understanding capabilities, improving the completeness of multilingual knowledge graphs (KGs).However, existing MKGC research underutilizes the multilingual capabilities of LLMs and ignores the shareability of cross-lingual knowledge.In this paper, we propose a novel MKGC framework that leverages multilingual shared knowledge to significantly enhance performance through two components: Knowledge-level Grouped Mixture of Experts (KL-GMoE) and Iterative Entity Reranking (IER).KL-GMoE efficiently models shared knowledge, while IER significantly enhances its utilization.To evaluate our framework, we constructed a mKG dataset containing 5 languages and conducted comprehensive comparative experiments with existing state-of-the-art (SOTA) MKGC method.The experimental results demonstrate that our framework achieves improvements of 5.47%, 3.27%, and 1.01% in the Hits@1, Hits@3, and Hits@10 metrics, respectively, compared with SOTA MKGC method.Further experimental analysis revealed the properties of knowledge sharing in settings of unseen and unbalanced languages.We have released the dataset and code for our work on https://github.com/gaoxiaofei07/KL-GMoE.

Large Language Models (LLMs) hold significant promise for accelerating academic ideation but face critical challenges in grounding ideas and mitigating confirmation bias during refinement. To address these limitations, we propose MotivGraph-SoIQ, a novel framework that enhances LLM ideation by integrating a Motivational Knowledge Graph (MotivGraph), which provides essential grounding from research literature, with a Q-Driven Socratic Ideator. The Ideator, a dual-agent system utilizing Socratic questioning, facilitates a rigorous refinement process that mitigates confirmation bias and significantly improves idea quality across dimensions of novelty, experimental feasibility, and motivation. Our experimental results demonstrate MotivGraph-SoIQ’s effectiveness. Comparative studies show significant quantitative improvements over SOTA methods across LLM-based scoring, ELO ranking, and human evaluation. Ablation studies further validate the crucial contributions of both the MotivGraph for enhancing idea novelty and practicality, and the Socratic dialogue with the teacher agent for substantial quality improvement. This work underscores the potential of combining structured knowledge with interactive, critique-based refinement for robust LLM ideation.

While hallucinations of large language models could be alleviated through retrieval-augmented generation and citation generation, how the model utilizes internal knowledge is still opaque, and the trustworthiness of its generated answers remains questionable. In this work, we introduce Context-Prior Augmented Citation Generation task, requiring models to generate citations considering both external and internal knowledge while providing trustworthy references, with 5 evaluation metrics focusing on 3 aspects: answer helpfulness, citation faithfulness, and trustworthiness. We introduce RAEL, the paradigm for our task, and also design INTRALIGN, an integrated method containing customary data generation and an alignment algorithm. Our experimental results show that our method achieves a better cross-scenario performance with regard to other baselines. Our extended experiments further reveal that retrieval quality, question types, and model knowledge have considerable influence on the trustworthiness in citation generation.

Despite the significant progress made by existing retrieval augmented language models (RALMs) in providing trustworthy responses and grounding in reliable sources, they often overlook effective alignment with human preferences. In the alignment process, reward models (RMs) act as a crucial proxy for human values to guide optimization. However, it remains unclear how to evaluate and select a reliable RM for preference alignment in RALMs. To this end, we propose RAG-RewardBench, the first benchmark for evaluating RMs in RAG settings. First, we design four crucial and challenging RAG-specific scenarios to assess RMs, including multi-hop reasoning, fine-grained citation, appropriate abstain, and conflict robustness. Then, we incorporate 18 RAG subsets, six retrievers, and 24 RALMs to increase the diversity of data sources. Finally, we adopt an LLM-as-a-judge approach to improve preference annotation efficiency and effectiveness, exhibiting a strong correlation with human annotations. Based on the RAG-RewardBench, we conduct a comprehensive evaluation of 45 RMs and uncover their limitations in RAG scenarios. Additionally, we also reveal that existing trained RALMs show almost no improvement in preference alignment, highlighting the need for a shift towards preference-aligned training.

Chain-of-thought (CoT) prompting demonstrates varying performance under different reasoning tasks.Previous work attempts to evaluate it but falls short in providing an in-depth analysis of patterns that influence the CoT. In this paper, we study the CoT performance from the perspective of effectiveness and faithfulness. For the former, we identify key factors that influence CoT effectiveness on performance improvement, including problem difficulty, information gain, and information flow. For the latter, we interpret the unfaithful CoT issue by conducting a joint analysis of the information interaction among the question, CoT, and answer. The result demonstrates that, when the LLM predicts answers, it can recall correct information missing in the CoT from the question, leading to the problem. Finally, we propose a novel algorithm to mitigate this issue, in which we recall extra information from the question to enhance the CoT generation and evaluate CoTs based on their information gain. Extensive experiments demonstrate that our approach enhances both the faithfulness and effectiveness of CoT.

This paper systematically addresses the challenge of rule retrieval, a crucial yet underexplored area. Vanilla retrieval methods using sparse or dense retrievers to directly search for relevant rules to support downstream reasoning, often suffer from low accuracy. This is primarily due to a significant semantic gap between the instantiated facts in the queries and the abstract representations of the rules. Such misalignment results in suboptimal retrieval quality, which in turn negatively impacts reasoning performance. To overcome these challenges, we propose Self-Induction Augmented Retrieval (SIAR), a novel approach that utilizes Large Language Models (LLMs) to induce potential inferential rules that might offer benefits for reasoning by abstracting the underlying knowledge and logical structure in queries. These induced rules are then used for query augmentation to improve retrieval effectiveness. Additionally, we introduce Rule Relevance ReEstimate (R³), a method that re-estimates the relevance of retrieved rules by assessing whether the abstract knowledge they contain can be instantiated to align with the facts in the queries and the helpfulness for reasoning. Extensive experiments across various settings demonstrate the effectiveness and versatility of our proposed methods.

Recent advancements in large language models (LLMs) have demonstrated remarkable capabilities in complex reasoning tasks, such as math problem-solving and code generation. However, multi-hop question answering (MHQA) over long contexts, which demands both robust knowledge-intensive reasoning and efficient processing of lengthy documents, remains a significant challenge. Existing approaches often struggle to balance these requirements, either neglecting explicit reasoning or incurring expensive computational costs due to full-attention mechanisms over long contexts. To address this, we propose Search-in-Context (SIC), a novel framework that integrates Monte Carlo Tree Search (MCTS) with dynamic key-value (KV) retrieval to enable iterative, context-aware reasoning. SIC dynamically retrieves critical KV pairs (e.g., 4K tokens) at each step, prioritizing relevant evidence while mitigating the “lost in the middle” problem. Furthermore, the paper introduces a Process-Reward Model (PRM) trained on auto-labeled data to guide the MCTS process with stepwise rewards, promoting high-quality reasoning trajectories without manual annotation. Experiments on three long-context MHQA benchmarks (HotpotQA, 2WikiMultihopQA, MuSiQue) and a counterfactual multi-hop dataset demonstrate SIC’s superiority, achieving state-of-the-art performance while significantly reducing computational overhead.

Large Language Models (LLMs) can improve commonsense reasoning through generating intermediate knowledge. However, the effectiveness of this knowledge introspection is not always guaranteed. This paper first systematically investigates and reveals an introspection paradox: while simple introspection tends to benefit weaker models, it often degrades the performance of stronger ones, particularly on simpler tasks. Our deep analysis indicates that this paradox arises from a complex interplay among model capability, task difficulty and the quality of generated knowledge. Further interpretability analysis reveals the origins of low-quality knowledge generation. To better employ introspected knowledge in LLM, this paper proposes a training-free Adaptive Introspection Strategy that operates in two stages using only the model’s internal states: Knowledge Detection, which dynamically identifies and discards potentially low-quality knowledge, and Knowledge Regeneration, which employs attention smoothing to guide the model away from harmful failure modes during knowledge generation. Extensive experiments on five Llama models with different sizes and eight commonsense reasoning benchmarks demonstrate that our approach effectively mitigates the limitations of standard introspection and has consistent performance gains across almost all settings.

Multimodal knowledge editing is an important method for modifying outdated or incorrect knowledge in Multimodal Large Language Models (MLLMs). However, existing datasets for multimodal knowledge editing lack multi-granularity knowledge. In this paper, we present a more realistic dataset called M2Edit, which includes three distinct types of knowledge: entity, relation, and action. Additionally, existing knowledge editing methods for MLLMs lack the ability to handle multi-granularity knowledge and generalize to multimodal data. To address these limitations, we propose the multimodal knowledge editing method MLE. This approach identifies key knowledge layers within different components and collaboratively edits the various components of MLLMs. As a result, we observe significant improvements in visual generality performance, ranging from 4.8 to 10.8, and achieve the best overall performance on knowledge data of different granularities.

The paper presents KMatrix-2, an open-source toolkit that supports comprehensive heterogeneous knowledge collaborative enhancement for Large Language Models (LLMs). As the successor of KMatrix, our toolkit offers powerful modular components and typical enhancement patterns for convenient construction of mainstream knowledge-enhanced LLMs systems. Besides, it provides unified knowledge integration and joint knowledge retrieval methods to achieve more comprehensive heterogeneous knowledge collaborative enhancement. Compared with KMatrix which mainly focuses on descriptive knowledge, this work additionally considers procedural knowledge. Moreover, systematic inter-context and context-memory knowledge conflict resolution methods are offered for better knowledge integration. Some key research questions in heterogeneous knowledge-enhanced Large Language Models systems are analyzed, and our toolkit’s capability in building such systems is validated.

“Although Large Language Models (LLMs) have demonstrated strong instruction-following abil-ity, they are further supposed to be controlled and guided by inferential rules in real-world scenarios to be safe, accurate, and intelligent. This demands the possession of inferential rule-following capability of LLMs. However, no prior work has made a clear evaluation of the inferential rule-following capability of LLMs. Previous studies that try to evaluate the inferential rule-following capability of LLMs fail to distinguish the inferential rule-following scenarios from the instruction-following scenarios. Therefore, this paper first clarifies the concept of inferential rule-following and proposes a comprehensive benchmark, RuleBench, to evaluate a diversified range of inferential rule-following abilities. Our experimental results on a variety of LLMs show that they are still limited in following rules. Our analysis based on the evaluation results provides insights into the improvements for LLMs toward a better inferential rule-following intelligent agent. We further propose Inferential Rule-Following Tuning (IRFT). The experimental results show that through IRFT, LLMs can learn abstract inferential rule-following abilities from purely synthetic data and then generalize to RuleBench. The data and code can be found at:https://gitee.com/forangel2014/llm-rule-following-code”

Knowledge editing, which aims to update the knowledge encoded in language models, can be deceptive. Despite the fact that many existing knowledge editing algorithms achieve near-perfect performance on conventional metrics, the models edited by them are still prone to generating original knowledge. This paper introduces the concept of "superficial editing” to describe this phenomenon. Our comprehensive evaluation reveals that this issue presents a significant challenge to existing algorithms. Through systematic investigation, we identify and validate two key factors contributing to this issue: (1) the residual stream at the last subject position in earlier layers and (2) specific attention modules in later layers. Notably, certain attention heads in later layers, along with specific left singular vectors in their output matrices, encapsulate the original knowledge and exhibit a causal relationship with superficial editing. Furthermore, we extend our analysis to the task of superficial unlearning, where we observe consistent patterns in the behavior of specific attention heads and their corresponding left singular vectors, thereby demonstrating the robustness and broader applicability of our methodology and conclusions. Our code is available at https://github.com/jiakuan929/superficial-editing.

With the development of large language models, they are widely used as agents in various fields. A key component of agents is memory, which stores vital information but is susceptible to jailbreak attacks. Existing research mainly focuses on single-agent attacks and shared memory attacks. However, real-world scenarios often involve independent memory. In this paper, we propose the Troublemaker Makes Chaos in Honest Town (TMCHT) task, a large-scale, multi-agent, multi-topology text-based attack evaluation framework. TMCHT involves one attacker agent attempting to mislead an entire society of agents. We identify two major challenges in multi-agent attacks: (1) Non-complete graph structure, (2) Large-scale systems. We attribute these challenges to a phenomenon we term toxicity disappearing. To address these issues, we propose an Adversarial Replication Contagious Jailbreak (ARCJ) method, which optimizes the retrieval suffix to make poisoned samples more easily retrieved and optimizes the replication suffix to make poisoned samples have contagious ability. We demonstrate the superiority of our approach in TMCHT, with 23.51%, 18.95%, and 52.93% improvements in line, star topologies, and 100-agent settings. It reveals potential contagion risks in widely used multi-agent architectures.

As Multimodal Large Language Models (MLLMs) advance, multimodal agents show promise in real-world tasks like web navigation and embodied intelligence. However, due to limitations in a lack of external feedback, these agents struggle with self-correction and generalization. A promising approach is to use reward models as external feedback, but there is no clear on how to select reward models for agents. Thus, there is an urgent need to build a reward bench targeted at agents. To address these challenges, we propose Agent-RewardBench, a benchmark designed to evaluate reward modeling ability in MLLMs. The benchmark is characterized by three key features: (1) Multiple dimensions and real-world agent scenarios evaluation. It covers perception, planning, and safety with 7 scenarios; (2) Step-level reward evaluation. It allows for the assessment of agent capabilities at the individual steps of a task, providing a more granular view of performance during the planning process; and (3) Appropriately difficulty and high-quality. We carefully sample from 10 diverse models, difficulty control to maintain task challenges, and manual verification to ensure the integrity of the data. Experiments demonstrate that even state-of-the-art multimodal models show limited performance, highlighting the need for specialized training in agent reward modeling. Code is available at github.

We demonstrate that features, rather than neurons, serve as superior analytical units for understanding the mechanisms of factual knowledge in Language Models (LMs). Previous studies primarily utilize MLP neurons as units of analysis; however, neurons suffer from polysemanticity, leading to limited knowledge expression and poor interpretability. We first conduct preliminary experiments to validate that SAE can effectively decompose neurons into features. With this established, our core findings reveal three key advantages of features over neurons: (1) Features exhibit stronger influence on knowledge expression and superior interpretability. (2) Features demonstrate enhanced monosemanticity, showing distinct activation patterns between related and unrelated facts. (3) Feature-based method demonstrates superior performance over neuron-based approaches in erasing privacy-sensitive information from LMs. Additionally, we propose FeatureEdit, the first feature-based editing method. Code and dataset will be available.

Knowledge neuron theory provides a key approach to understanding the mechanisms of factual knowledge in Large Language Models (LLMs), which suggests that facts are stored within multi-layer perceptron neurons. This paper further explores Degenerate Knowledge Neurons (DKNs), where distinct sets of neurons can store identical facts, but unlike simple redundancy, they also participate in storing other different facts. Despite the novelty and unique properties of this concept, it has not been rigorously defined and systematically studied. Our contributions are: (1) We pioneer the study of structures in knowledge neurons by analyzing weight connection patterns, providing a comprehensive definition of DKNs from both functional and structural aspects. (2) Based on this definition, we develop the Neuronal Topology Clustering method, leading to a more accurate DKN identification. (3) We demonstrate the practical applications of DKNs in two aspects: guiding LLMs to learn new knowledge and relating to LLMs’ robustness against input errors.

Large Language Models (LLMs) have demonstrated remarkable capabilities in various tasks, yet they often struggle with context-faithfulness generations that properly reflect contextual knowledge. While existing approaches focus on enhancing the decoding strategies, they ignore the fundamental mechanism of how contextual information is processed within LLMs’ internal states. As a result, LLMs remain limited in their ability to fully leverage contextual knowledge. In this paper, we propose Context-aware Layer Enhancement (CaLE), a novel intervention method that enhances the utilization of contextual knowledge within LLMs’ internal representations. By employing 𝒱-usable information analysis, CaLE strategically amplifies the growth of contextual information at an optimal layer, thereby enriching representations in the final layer. Our experiments demonstrate that CaLE effectively improves context-faithful generation in Question-Answering tasks, particularly in scenarios involving unknown or conflicting contextual knowledge.

Personalized tool utilization is essential for aligning large language models (LLMs) with user preference in interaction scenarios with various tools. However, most of the current benchmarks primarily focus on either personalization of text generation or direct tool-utilizing, without considering both. In this work, we introduce a novel benchmark ETAPP for evaluating personalized tool invocation, establishing a sandbox environment, and a comprehensive dataset of 800 testing cases covering diverse user profiles. To improve the accuracy of our evaluation, we propose a key-point-based LLM evaluation method, mitigating biases in the LLM-as-a-judge system by manually annotating key points for each test case and providing them to LLM as the reference. Additionally, we evaluate the excellent LLMs and provide an in-depth analysis. Furthermore, we investigate the impact of different tool-invoking strategies on LLMs’ personalization performance and the effects of fine-tuning in our task. The effectiveness of our preference-setting and key-point-based evaluation method is also validated. Our findings offer insights into improving personalized LLM agents. Our code is available at https://github.com/hypasd-art/ETAPP.

Large Language Models (LLMs) offer a transparent brain with accessible parameters that encode extensive knowledge, which can be analyzed, located and transferred. Consequently, a key research challenge is to transcend traditional knowledge transfer paradigms rooted in symbolic language and achieve genuine Parametric Knowledge Transfer (PKT). Significantly, exploring effective methods for transferring knowledge across LLMs of different scales through parameters presents an intriguing and valuable research direction. In this paper, we first demonstrate Alignment in parametric space is the fundamental prerequisite to achieve successful cross-scale PKT. We redefine the previously explored knowledge transfer as Post-Align PKT (PostPKT), which utilizes extracted parameters for LoRA initialization and requires subsequent fine-tune for alignment. Hence, to reduce cost for further fine-tuning, we introduce a novel Pre-Align PKT (PrePKT) paradigm and propose a solution called LaTen (Locate-Then-Align) that aligns the parametric spaces of LLMs across scales only using several training steps without following training. Comprehensive experiments on four benchmarks demonstrate that both PostPKT and PrePKT face challenges in achieving consistently stable transfer. Through in-depth analysis, we identify Neural Incompatibility as the ethological and parametric structural differences between LLMs of varying scales, presenting fundamental challenges to achieving effective PKT. These findings provide fresh insights into the parametric architectures of LLMs and highlight promising directions for future research on efficient PKT. Our code is available at https://github.com/Trae1ounG/Neural_Incompatibility.

The emerging paradigm of enabling Large Language Models (LLMs) to generate citations in Question-Answering (QA) tasks is lacking in a unified framework to standardize and fairly compare different citation generation methods, leading to difficulties in reproduction and innovation. Therefore, we introduce Citeflow, an open-source and modular framework fostering reproduction and the implementation of new designs. Citeflow is highly extensible, allowing users to utilize four main modules and 14 components to construct a pipeline, evaluate an existing method, and understand the attributing LLM-generated contents. The framework is also paired with a visual interface, Citefix, facilitating case study and modification of existing citation generation methods. Users can use this interface to conduct LLM-powered case studies according to different scenarios. Citeflow and Citefix are highly integrated into the toolkit CiteLab, and we use an authentic process of multiple rounds of improvement through the Human-LLM interaction interface to demonstrate the efficiency of our toolkit on implementing and modifying citation generation pipelines. Citelab is released at https://github.com/SjJ1017/Citelab

As large language models (LLMs) continue to advance, there is a growing urgency to enhance the interpretability of their internal knowledge mechanisms. Consequently, many interpretation methods have emerged, aiming to unravel the knowledge mechanisms of LLMs from various perspectives. However, current interpretation methods differ in input data formats and interpreting outputs. The tools integrating these methods are only capable of supporting tasks with specific inputs, significantly constraining their practical applications. To address these challenges, we present an open-source Knowledge Mechanisms Revealer&Interpreter (Know-MRI) designed to analyze the knowledge mechanisms within LLMs systematically. Specifically, we have developed an extensible core module that can automatically match different input data with interpretation methods and consolidate the interpreting outputs. It enables users to freely choose appropriate interpretation methods based on the inputs, making it easier to comprehensively diagnose the model’s internal knowledge mechanisms from multiple perspectives. Our code is available at https://github.com/nlpkeg/Know-MRI. We also provide a demonstration video on https://youtu.be/NVWZABJ43Bs.