In real-world Tool-Integrated Reasoning (TIR) scenarios, a major source of inefficiency is that the toolcalls create pauses between LLM requests and cause KV-cache eviction. Also, the long, unfiltered response returned by external tools inflates the KV-cache, so each decode step spends more time loading the growing cache and thus becomes steadily slower as context length increases. However, existing efficiency metrics like token counts and toolcall counts fail to capture this real computational cost. To address this, we introduce PTE (Prefill Token Equivalents), a hardware-aware TIR-efficiency metric that unifies internal reasoning and external tool-use costs while explicitly accounting for non-reusable KV-Cache and long-tool-response scenarios, thus better reflects real-world scenarios. We conduct extensive experiments across five TIR benchmarks, quantify their PTE costs, and identify four inefficiency patterns that appear in TIR. In a simulated high-concurrency industrial setting, PTE explains wall-clock latency significantly better than token-count metric. We also discover that trajectories with higher PTE costs tend to have lower reasoning correctness, indicating that simply using more tools does not improve the quality of the answer. PTE offers a new perspective on the efficiency of Tool-Integrated Reasoning. The code is available.
论文检索
输入标题、作者或关键词，从 9,460 篇学术成果中精准定位
Multilingual Retrieval-Augmented Generation (mRAG) leverages cross-lingual evidence to ground Large Language Models (LLMs) in global knowledge. However, we show that current mRAG systems suffer from a language bias during reranking, systematically favoring English and the query’s native language. By introducing an estimated oracle evidence analysis, we quantify a substantial performance gap between existing rerankers and the achievable upper bound. Further analysis reveals a critical distributional mismatch: while optimal predictions require evidence scattered across multiple languages, current systems systematically suppress such “answer-critical” documents, thereby limiting downstream generation performance. To bridge this gap, we propose Language-Agnostic Utility-driven Reranker Alignment (LAURA), Experiments across diverse languages and generation models show that LAURA effectively mitigates language bias and consistently improves mRAG performance.
Large Language Models (LLMs) based autonomous agents demonstrate multifaceted capabilities to contribute substantially to economic production. However, existing benchmarks remain focused on single agentic capability, failing to capture long-horizon real-world scenarios. Moreover, the reliance on human-in-the-loop feedback for realistic tasks creates a scalability bottleneck, hindering automated rollout collection and evaluation. To bridge this gap, we introduce AgencyBench, a comprehensive benchmark derived from daily AI usage, evaluating 6 core agentic capabilities across 32 real-world scenarios, comprising 138 tasks with specific queries, deliverables, and rubrics. These scenarios require an average of 90 tool calls, 1 million tokens, and hours of execution time to resolve. To enable automated evaluation, we employ a user simulation agent to provide iterative feedback, and a Docker sandbox to conduct visual and functional rubric-based assessment. Experiments reveal that closed-source models significantly outperform open-source models (48.4% vs 32.1%). Further analysis reveals significant disparities across models in resource efficiency, feedback-driven self-correction, and specific tool-use preferences.
Arg-LLaDA: Argument Summarization via Large Language Diffusion Models and Sufficiency-Aware Refinement
PDF ↗Argument summarization aims to generate concise, structured representations of complex, multi-perspective debates. While recent work has advanced the identification and clustering of argumentative components, the generation stage remains underexplored. Existing approaches typically rely on single-pass generation, offering limited support for factual correction or structural refinement. To address this gap, we introduce Arg-LLaDA, a novel large language diffusion framework that iteratively improves summaries via sufficiency-guided remasking and regeneration. Our method combines a flexible masking controller with a sufficiency-checking module to identify and revise unsupported, redundant, or incomplete spans—yielding more faithful, concise, and coherent outputs. Empirical results on two benchmark datasets demonstrate that Arg-LLaDA surpasses state-of-the-art baselines in 7 out of 10 automatic evaluation metrics. In addition, human evaluations reveal substantial improvements across core dimensions, coverage, faithfulness, and conciseness,validating the effectiveness of our iterative, sufficiency-aware generation strategy.
While agent evaluation has shifted toward long-horizon tasks, most benchmarks still emphasize local, step-level reasoning rather than the global constrained optimization (e.g., time and financial budgets) that demands genuine planning ability. Meanwhile, existing LLM planning benchmarks underrepresent the active information gathering and fine-grained local constraints typical of real-world settings. To address this, we introduce DeepPlanning, a challenging benchmark for practical long-horizon agent planning. It features multi-day travel planning and multi-product shopping tasks that require proactive information acquisition, local constrained reasoning, and global constrained optimization. Evaluations on DeepPlanning show that even frontier agentic LLMs struggle with these problems, highlighting the importance of reliable explicit reasoning patterns and parallel tool use for achieving better effectiveness-efficiency trade-offs. Error analysis further points to promising directions for improving agentic LLMs over long planning horizons. We open-source the code and data to support future research.
Spectral Disentanglement: Rank-Aware Task Adaptation for Rehearsal-free Continual Learning in LLMs
PDF ↗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.
Large Language Models (LLMs) can extend their parameter knowledge limits by adopting the Tool-Integrated Reasoning (TIR) paradigm. However, existing LLM-based agent training framework often focuses on answers’ accuracy, overlooking specific alignment for behavior patterns. Consequently, agent often exhibits ineffective actions during TIR tasks, such as redundant and insufficient tool calls. How to calibrate erroneous behavioral patterns when executing TIR tasks, thereby exploring effective trajectories, remains an open-ended problem. In this paper, we propose ET-Agent, a training framework for calibrating agent’s tool-use behavior through two synergistic perspectives: Self-evolving Data Flywheel and Behavior Calibration Training. Specifically, we introduce a self-evolutionary data flywheel to generate enhanced data, used to fine-tune LLM to improve its exploration ability. Based on this, we implement an two-phases behavior-calibration training framework. It is designed to progressively calibrate erroneous behavioral patterns to optimal behaviors. Further in-depth experiments confirm the superiority of ET-Agent across multiple dimensions, including correctness, efficiency, reasoning conciseness, and tool execution accuracy. Our ET-Agent framework provides practical insights for research in the TIR field. Codes can be found in https://github.com/RUC-NLPIR/ET-Agent
Reasoning Over Space: Enabling Geographic Reasoning for LLM-Based Generative Next POI Recommendation
PDF ↗Generative recommendation with large language models (LLMs) reframes prediction as sequence generation, yet existing LLM-based recommenders remain limited in leveraging geographic signals that are crucial in mobility and local-services scenarios. Here, we present Reasoning Over Space (ROS), a framework that utilizes geography as a vital decision variable within the reasoning process. ROS introduces a Hierarchical Spatial Semantic ID (SID) that discretizes coarse-to-fine locality and POI semantics into compositional tokens, and endows LLM with a three-stage Mobility Chain-of-Thought (CoT) paradigm that models user personality, constructs an intent-aligned candidate space, and performs locality informed pruning. We further align the model with real world geography via spatial-guided Reinforcement Learning (RL). Experiments on three widely used location-based social network (LBSN) datasets show that ROS achieves over 10% relative gains in hit rate over strongest LLM-based baselines and improves cross-city transfer, despite using a smaller backbone model.
Beyond Timestamps: Bridging Forward and Backward Reasoning in Temporal Numerical and Relational Understanding
PDF ↗Temporal reasoning remains a critical challenge for large language models (LLMs), particularly when it requires encompassing relational dependencies and numerical constraints. Yet, existing benchmarks largely overlook the joint consideration of these two dimensions and primarily rely on single-task evaluation paradigms, making it difficult to assess whether correct answers reflect grounded reasoning or arise from superficial statistical recall. To address these gaps, we introduce TNR, a benchmark designed to evaluate both Temporal Numerical and Relational reasoning. We propose a bi-directional evaluation framework consisting of forward generation via Question Answering (QA) and backward verification via Fact Verification (FV). By measuring the alignment between QA and FV, we introduce a Consistency Rate to quantify the robustness of reasoning across these two directions. Experiments on a range of LLMs reveal notable discrepancies between QA and FV performance, particularly in numerical and interval-based tasks. Moreover, our bi-directional error analysis demonstrates that these inconsistencies often stem from heuristic shortcuts and statistical co-occurrences rather than grounded logical deduction, flaws that are frequently masked in standard single-task evaluations.
Evo-Attacker: Memory-Augmented Reinforcement Learning for Long-Horizon Tool Attacks on LLM-MAS
PDF ↗While Large Language Model-based Multi-Agent Systems (LLM-MAS) demonstrate remarkable capabilities in solving complex tasks by orchestrating specialized agents and external tools, the implicit trust in tool outputs creates a critical attack surface. Existing tool attacks are limited by domain specificity or fixed and static templates. To address these challenges, we propose Evo-Attacker, which formulates the tool attack as a self-evolving, memory-augmented reinforcement learning process. Evo-Attacker constructs a dynamic attack memory and employs deliberative reasoning to retrieve adversarial patterns and strategize modifying interventions at critical moments. Furthermore, we introduce Attack-Flow GRPO to optimize intermediate reasoning steps via terminal outcomes, addressing the long-horizon credit assignment challenge. Comprehensive experiments demonstrate that Evo-Attacker consistently outperforms baselines, highlighting its generalization and evolutionary capabilities and the urgent need for defensive tool safeguards.
STEM: Structure-Tracing Evidence Mining for Knowledge Graphs-Driven Retrieval-Augmented Generation
PDF ↗Knowledge Graph-based Question Answering (KGQA) plays a pivotal role in complex reasoning tasks but remains constrained by two persistent challenges: the structural heterogeneity of Knowledge Graphs(KGs) often leads to semantic mismatch during retrieval, while existing reasoning path retrieval methods lack a global structural perspective. To address these issues, we propose Structure-Tracing Evidence Mining (STEM), a novel framework that reframes multi-hop reasoning as a schema-guided graph search task. First, we design a Semantic-to-Structural Projection pipeline that leverages KG structural priors to decompose queries into atomic relational assertions and construct an adaptive query schema graph. Subsequently, we execute globally-aware node anchoring and subgraph retrieval to obtain the final evidence reasoning graph from KG. To more effectively integrate global structural information during the graph construction process, we design a Triple-Dependent GNN (Triple-GNN) to generate a Global Guidance Subgraph (Guidance Graph) that guides the construction. STEM significantly improves both the accuracy and evidence completeness of multi-hop reasoning graph retrieval, and achieves State-of-the-Art performance on multiple multi-hop benchmarks.
Large Vision-Language Models (LVLMs) excel at visual understanding but face severe computational bottlenecks when processing high-resolution images and long videos due to massive visual token counts. Token pruning mitigates this by selectively removing less informative tokens while maintaining performance. However, existing methods vary widely in pruning location (vision encoder vs. LLM decoder), importance criteria (attention vs. similarity vs. learned scores), and application strategy, lacking systematic comparison. This survey presents the first comprehensive review of token pruning for LVLMs. We propose a taxonomy categorizing methods into vision-side, LLM-side, and hybrid paradigms, systematically analyze token selection mechanisms and pruning strategy. We further discuss evaluation protocols and identify key challenges including prompt-adaptive pruning and hardware-aware design. Our survey provides a structured foundation for this rapidly growing research area.
When KV Cache Reuse Fails in Multi-Agent Systems: Cross-Candidate Interaction is Crucial for LLM Judges
PDF ↗Multi-agent LLM systems routinely generate multiple candidate responses that are aggregated by an LLM judge. To reduce the dominant prefill cost in such pipelines, recent work advocates KV cache reuse across partially shared contexts and reports substantial speedups for generation agents. In this work, we show that these efficiency gains do not transfer uniformly to judge-centric inference. Across GSM8K, MMLU, and HumanEval, we find that reuse strategies that are effective for execution agents can severely perturb judge behavior: end-task accuracy may appear stable, yet the judge’s selection becomes highly inconsistent with dense prefill. We quantify this risk using Judge Consistency Rate (JCR) and provide diagnostics showing that reuse systematically weakens cross-candidate attention, especially for later candidate blocks. Our ablation further demonstrates that explicit cross-candidate interaction is crucial for preserving dense-prefill decisions. Overall, our results identify a previously overlooked failure mode of KV cache reuse and highlight judge-centric inference as a distinct regime that demands dedicated, risk-aware system design.
Long Chain-of-Thought (CoT) reasoning has significantly advanced the capabilities of Large Language Models (LLMs), but this progress is accompanied by substantial memory and latency overhead from the extensive Key-Value (KV) cache. Although KV cache quantization is a promising compression technique, existing low-bit quantization methods often exhibit severe performance degradation on complex reasoning tasks. Fixed-precision quantization struggles to handle outlier channels in the key cache, while current mixed-precision strategies fail to accurately identify components requiring high-precision representation. We find that an effective low-bit KV cache quantization strategy must consider two factors: a key channel’s intrinsic quantization difficulty and its relevance to the query. Based on this insight, we propose MixKVQ, a novel plug-and-play method that introduces a lightweight, query-aware algorithm to identify and preserve critical key channels that need higher precision, while applying per-token quantization for value cache. Experiments on complex reasoning datasets demonstrate that our approach significantly outperforms existing low-bit methods, achieving performance comparable to a full-precision baseline at a substantially reduced memory footprint.
Equipping Large Language Models (LLMs) with pedagogical tutoring capabilities holds significant promise for education. Existing approaches simulate tutor behaviors or preferences and use them to prompt or fine-tune LLMs for dialogue tutoring. However, such methods often fail to sustain high-quality pedagogical conversations that provide explicit stepwise scaffolding and adapt to learners’ evolving cognitive states. To address this, we propose ScaffoldLM, a planning-guided tutoring framework with an assessment-driven memory for multi-turn math dialogue tutoring. ScaffoldLM first generates a stepwise pedagogical plan from solution steps, which serves as a stable backbone for explicit scaffolding. During tutoring, the tutoring memory is updated by an assessment-driven control loop that infers the learner’s cognitive state, evaluates whether the current step target is met, and adaptively selects tutoring actions. The plan, step-level progress, inferred learner states, and dialogue history are maintained in memory to support coherent multi-turn guidance. Experiments on multi-turn math tutoring benchmarks demonstrate that ScaffoldLM substantially improves pedagogical tutoring quality over strong baselines. Code is publicly available at https://github.com/BNU-ERC-ITEA/ScaffoldLM.
While Learned Data Compression (LDC) has achieved superior compression ratios, balancing precise probability modeling with system efficiency remains challenging. Crucially, uniform single-stream architectures struggle to simultaneously capture micro-syntactic and macro-semantic features, necessitating deep serial stacking that exacerbates latency. Compounding this, heterogeneous systems are constrained by device speed mismatches, where throughput is capped by Amdahl’s Law due to serial processing. To this end, we propose a Dual-Stream Multi-Scale Decoupler that disentangles local and global contexts to replace deep serial processing with shallow parallel streams, and incorporate a Hierarchical Gated Refiner for adaptive feature refinement and precise probability modeling. Furthermore, we design a Concurrent Stream-Parallel Pipeline, which overcomes systemic bottlenecks to achieve full-pipeline parallelism. Extensive experiments demonstrate that our method achieves state-of-the-art performance in both compression ratio and throughput, while maintaining the lowest latency and memory usage. The code is available at https://github.com/huidong-ma/FADE.
With the widespread deployment of public large language models (LLMs) such as ChatGPT, protecting user prompt privacy has become an increasingly critical issue. Existing privacy-preserving inference methods sacrifice either utility or efficiency, and often require model-specific modifications that limit their compatibility. In this paper, we propose SharedRequest, a model-agnostic framework for privacy-preserving LLM inference that reformulates privacy protection at the batch level rather than the individual-prompt level. The key idea is to obscure sensitive information by mixing original prompts with noisy variants, while grouping semantically equivalent instructions to amortize the inference cost over a large batch of queries with minimal impact on LLM response quality. This design is independent of the LLM architecture, requiring no access to model parameters or architectural modification. Empirical results demonstrate that SharedRequest achieves over 20% higher utility compared to prior differential privacy baselines, and its shared-prompt mechanism reduces query cost by up to 5\times compared to non-batched inference.
Large language models (LLMs) have shown promise in simulating human-like social behaviors. Social graphs provide high-quality supervision signals that encode both local interactions and global network structure, yet they remain underutilized for LLM training. To address this gap, we propose Graphia, the first general LLM-based social graph simulation framework that leverages graph data as supervision for LLM post-training via reinforcement learning. With GNN-based structural rewards, Graphia trains specialized agents to predict whom to interact with (destination selection) and how to interact (edge generation), followed by designed graph generation pipelines. We evaluate Graphia under two settings: Transductive Dynamic Graph Generation (TDGG), a micro-level task with our proposed node-wise interaction alignment metrics; and Inductive Dynamic Graph Generation (IDGG), a macro-level task with our proposed metrics for aligning emergent network properties. On three real-world networks, Graphia improves micro-level alignment by 6.1% in the composite destination selection score, 12% in edge classification accuracy, and 27.9% in edge content BERTScore over the strongest baseline. For macro-level alignment, it achieves 35.98% higher structural similarity and 28.71% better replication of social phenomena such as power laws and echo chambers. Our results show that social graphs can serve as high-quality supervision signals for LLM post-training, closing the gap between agent behaviors and network dynamics for LLM-based simulation. Code is available at https://github.com/Ji-Cather/Graphia.git.
Empathetic dialogue requires not only recognizing a user’s emotional state but also making strategy-aware, context-sensitive decisions throughout response generation.However, the lack of a comprehensive empathy strategy framework, explicit task-aligned multi-stage reasoning, and high-quality strategy-aware data fundamentally limits existing approaches, preventing them from effectively modeling empathetic dialogue as a complex, multi-stage cognitive and decision-making process.To address these challenges, we propose STRIDE-ED, a STRategy-grounded, Interpretable, and DEep reasoning framework that models Empathetic Dialogue through structured, strategy-conditioned reasoning.To support effective learning, we develop a strategy-aware data refinement pipeline integrating LLM-based annotation, multi-model consistency-weighted evaluation, and dynamic sampling to construct high-quality training data aligned with empathetic strategies.Furthermore, we adopt a two-stage training paradigm that combines supervised fine-tuning with multi-objective reinforcement learning to better align model behaviors with target emotions, empathetic strategies, and response formats.Extensive experiments demonstrate that STRIDE-ED generalizes across diverse open-source LLMs and consistently outperforms existing methods on both automatic metrics and human evaluations.
Chain-of-Thought (CoT) has become a cornerstone of reasoning in large language models, yet its effectiveness is constrained by the limited expressiveness of discrete token sampling. Recent latent reasoning approaches attempt to alleviate this limitation by replacing discrete tokens with soft embeddings (probability-weighted mixtures of token embeddings) or hidden states, but they commonly suffer from two issues: (1) global activation injects perturbations into high-confidence steps, impairing reasoning stability; and (2) soft embeddings quickly collapse toward the highest-probability token, limiting exploration of alternative trajectories. To address these challenges, we propose SeLaR (Selective Latent Reasoning), a lightweight and training-free framework. SeLaR introduces an entropy-gated mechanism that activates soft embeddings only at low-confidence steps, while preserving discrete decoding at high-confidence steps. Additionally, we propose an entropy-aware contrastive regularization that pushes soft embeddings away from the highest-probability token’s direction, encouraging sustained exploration of multiple latent reasoning paths. Experiments on five reasoning benchmarks demonstrate that SeLaR consistently outperforms standard CoT and state-of-the-art training-free methods.