arXiv subjects

Rongzheng Wang

Publications and source records attributed to Rongzheng Wang.

12 recordsLinked to original sources

arXiv2026

DeferKV: Rethinking Eviction Timing for One-Shot KV Cache Compression

Long-context large language models (LLMs) have demonstrated strong capabilities across a wide range of tasks, but the growing KV cache introduces substantial memory and inference overhead. Existing one-shot KV cache compression methods typically commit to irreversible eviction immediately after prefill, before any signal from actual generation becomes available. Our quantitative analysis shows that early queries from the actual generation stage provide attention signals that are more consistent with subsequent decode attention, with the largest single-step gain occurring at the prefill-decode boundary. Based on this observation, we propose DeferKV, which moves the eviction decision from the end of prefill to the first real decoding step and temporally combines prompt-side and decode-side observations, thereby better aligning KV importance estimation with subsequent generation requirements. DeferKV requires no additional training, draft model, or future-query prediction module, making it simple and easy to deploy. Experiments on LongBench, RULER, and Needle-in-a-Haystack demonstrate that DeferKV consistently improves model performance under KV cache compression while maintaining low inference latency.

cs.CL↗

arXiv2026

Visual Graph Reasoning via Knowledge Compilation

Visual graph reasoning requires answering graph-theoretic questions directly from graph images, where graph topology and state are conveyed visually rather than given in symbolic form. Despite recent progress of vision-language models (VLMs), current approaches to visual graph reasoning still fail on simple visual graph problems. This reveals a fundamental limitation of existing approaches: they prioritize final-answer supervision over the intermediate recovery of an explicit graph representation that preserves graph topology and state from visual input. To address this limitation, we propose VGCompiler, a compilation-centric paradigm for visual graph reasoning via knowledge compilation. VGCompiler organizes reasoning around two compilers: a representation compiler that recovers a structure-preserving intermediate graph representation from visual input, and an operation compiler that compiles query intent under the recovered graph state into an executable graph operation. Specifically, we build VGCompiler on Qwen3-VL-8B and train it with reinforcement learning guided by a layered reward over executability, compiled graph validity, representation quality, and operation quality. VGCompiler uses a frozen observer to summarize graph and question conditions into lightweight signatures, enabling archive retrieval and code reuse across similar regimes. Experiments on three benchmarks GVLQA, VisionGraph, and VGCURE, show that Qwen-VGCompiler, built on an 8B backbone, surpasses the strongest closed-source VLM baseline by 28.9% and the strongest code-based baseline by 23.7%, while maintaining high efficiency. We further evaluate VGCompiler on three real-world domains, including metro routing, logistics delivery, and network fault assessment, where it generalizes across heterogeneous visual graphs and domain-grounded tasks.

cs.MM↗

arXiv2026

When Safety Becomes a Vulnerability: Exploiting LLM Alignment Homogeneity for Transferable Blocking in RAG

Retrieval-Augmented Generation (RAG) systems are vulnerable to blocking attacks, in which poisoned documents cause large language models (LLMs) to refuse benign queries. Existing attacks rely on adversarial suffixes or explicit instructions, which are increasingly ineffective against modern LLMs, susceptible to prompt injection filtering, or require feedback from the target system. We observe overlapping risk categories and refusal criteria across safety-aligned LLMs, a phenomenon we term alignment homogeneity. This shared attack surface makes refusal-inducing context transferable across models. Accordingly, we propose TabooRAG, which optimizes one document per query for retrieval and refusal induction in a surrogate RAG environment, then transfers it to an unknown target system. Rather than injecting instructions, TabooRAG constructs query-relevant risk context to trigger alignment-driven refusal. To reduce optimization cost, it reuses validated strategies through a query-aware strategy library. Across nine LLMs and three datasets, TabooRAG achieves state-of-the-art ASR after filtering, with a 67.3% relative gain over the average per-setting best baseline. Further experiments show that TabooRAG remains effective with diverse surrogate models, against unseen target models, and under stronger RAG pipelines and existing defenses.

cs.CR↗

arXiv2026

Stop When Further Reasoning Won't Help: Attention-State Adaptive Generation in Reasoning Models

By incorporating test-time compute scaling, large reasoning models (LRMs) can solve complex problems through explicit chain-of-thought (CoT) reasoning processes. However, they often suffer from overthinking, resulting in redundant token outputs and degraded accuracy. Current methods to mitigate this issue remain limited: training-based approaches require substantial computational resources, while training-free methods rely on well-crafted prompts or unreliable confidence signals. In this work, we investigate early stopping from the perspective of attention distributions and propose a simple method, ASAG, which infers the model's reasoning state and adaptively adjusts the generation strategy. The proposed framework is training-free and plug-and-play, enabling seamless integration into existing LRMs. Extensive experiments on nine benchmarks demonstrate consistent improvements across mainstream LRMs with varying parameter scales, including the DeepSeek-R1-Distill and Qwen3 series. Specifically, ASAG improves average accuracy by 3.2% while reducing the number of generated tokens by nearly 40% across all reasoning tasks on Qwen3-8B.

cs.CL↗

arXiv2026

Toward Robust GraphRAG: Mitigating Retrieval Drift and Hallucination from Imperfect Knowledge Graphs

Graph Retrieval-Augmented Generation (GraphRAG) has become a common approach for multi-hop reasoning by using knowledge graphs (KGs) as structured retrieval indexes. However, most existing GraphRAG methods implicitly assume that LLM-constructed KGs provide structural support for evidence chaining. In this paper, we show that this assumption does not always hold in practice through an empirical analysis, and identify two recurring KG issue modes often overlooked by current retrievers: spurious noise and incomplete information. Spurious noise induces retrieval drift toward plausible but unsupported triples, whereas incomplete information leads to retrieval hallucination by forcing continuation through under-supported graph structure. To address these challenges, we propose CS-RAG, a robust GraphRAG framework that mitigates the impact of imperfect KGs during retrieval rather than relying on KG repair. CS-RAG first plans each query as an ordered sequence of executable atomic constraints and performs fine-grained anchor- and relation-aware retrieval to constrain evidence acquisition around the intended hop semantics. It then applies a sufficiency check to decide whether the retrieved evidence can safely induce variable bindings for subsequent propagation and activates textual recovery when structural support is insufficient, thereby reducing hallucinated structural continuation. Experiments on three multi-hop QA benchmarks show that CS-RAG is less sensitive to builder choice and remains stable under controlled KG issue injection. Code is available at: https://github.com/myz12138/CS-RAG/

cs.IR↗

arXiv2026

Rethinking LLM-Driven Heuristic Design: Generating Efficient and Specialized Solvers via Dynamics-Aware Optimization

Large Language Models (LLMs) have advanced the field of Combinatorial Optimization through automated heuristic generation. Instead of relying on manual design, this LLM-Driven Heuristic Design (LHD) process leverages LLMs to iteratively generate and refine solvers to achieve high performance. However, existing LHD frameworks face two critical limitations: (1) Endpoint-only evaluation, which ranks solvers solely by final gap to a reference solution, ignoring the convergence process and runtime efficiency; (2) High adaptation costs, where distribution shifts necessitate re-adaptation to generate specialized solvers for heterogeneous instance groups. To address these issues, we propose Dynamics-Aware Solver Heuristics (DASH), a framework that co-optimizes solver search mechanisms and runtime schedules guided by a convergence-aware metric, thereby identifying efficient and high-performance solvers. Furthermore, to mitigate expensive re-adaptation, DASH incorporates Profiled Library Retrieval (PLR), which maintains group-specialized solvers for profile-aware warm starts. These solvers are archived concurrently during evolution, allowing DASH to reuse matched specialists across heterogeneous distributions without restarting adaptation. Experiments on four combinatorial optimization problems demonstrate that DASH improves runtime efficiency by over 4 times while outperforming prior LHD baselines in the overall balance between gap and runtime across diverse problem scales. Furthermore, by enabling profile-aware warm starts, DASH maintains lower gap under distribution shift while reducing LLM adaptation costs by about 90%.

cs.LG↗

arXiv2026

Fixed Anchors Are Not Enough: Dynamic Retrieval and Persistent Homology for Dataset Distillation

Decoupled dataset distillation (DD) compresses large corpora into a few synthetic images by matching a frozen teacher's statistics. However, current residual-matching pipelines rely on static real patches, creating a fit-complexity gap and a pull-to-anchor effect that reduce intra-class diversity and hurt generalization. To address these issues, we introduce RETA -- a Retrieval and Topology Alignment framework for decoupled DD. First, Dynamic Retrieval Connection (DRC) selects a real patch from a prebuilt pool by minimizing a fit-complexity score in teacher feature space; the chosen patch is injected via a residual connection to tighten feature fit while controlling injected complexity. Second, Persistent Topology Alignment (PTA) regularizes synthesis with persistent homology: we build a mutual k-NN feature graph, compute persistence images of components and loops, and penalize topology discrepancies between real and synthetic sets, mitigating pull-to-anchor effect. Across CIFAR-100, Tiny-ImageNet, ImageNet-1K, and multiple ImageNet subsets, RETA consistently outperforms various baselines under comparable time and memory, especially reaching 64.3% top-1 accuracy on ImageNet-1K with ResNet-18 at 50 images per class, +3.1% over the best prior.

cs.CV↗

arXiv2026

KEPo: Knowledge Evolution Poison on Graph-based Retrieval-Augmented Generation

Graph-based Retrieval-Augmented Generation (GraphRAG) constructs the Knowledge Graph (KG) from external databases to enhance the timeliness and accuracy of Large Language Model (LLM) generations. However, this reliance on external data introduces new attack surfaces. Attackers can inject poisoned texts into databases to manipulate LLMs into producing harmful target responses for attacker-chosen queries. Existing research primarily focuses on attacking conventional RAG systems. However, such methods are ineffective against GraphRAG. This robustness derives from the KG abstraction of GraphRAG, which reorganizes injected text into a graph before retrieval, thereby enabling the LLM to reason based on the restructured context instead of raw poisoned passages. To expose latent security vulnerabilities in GraphRAG, we propose Knowledge Evolution Poison (KEPo), a novel poisoning attack method specifically designed for GraphRAG. For each target query, KEPo first generates a toxic event containing poisoned knowledge based on the target answer. By fabricating event backgrounds and forging knowledge evolution paths from original facts to the toxic event, it then poisons the KG and misleads the LLM into treating the poisoned knowledge as the final result. In multi-target attack scenarios, KEPo further connects multiple attack corpora, enabling their poisoned knowledge to mutually reinforce while expanding the scale of poisoned communities, thereby amplifying attack effectiveness. Experimental results across multiple datasets demonstrate that KEPo achieves state-of-the-art attack success rates for both single-target and multi-target attacks, significantly outperforming previous methods.

cs.LG↗

arXiv2025

NeuroPath: Neurobiology-Inspired Path Tracking and Reflection for Semantically Coherent Retrieval

Retrieval-augmented generation (RAG) greatly enhances large language models (LLMs) performance in knowledge-intensive tasks. However, naive RAG methods struggle with multi-hop question answering due to their limited capacity to capture complex dependencies across documents. Recent studies employ graph-based RAG to capture document connections. However, these approaches often result in a loss of semantic coherence and introduce irrelevant noise during node matching and subgraph construction. To address these limitations, we propose NeuroPath, an LLM-driven semantic path tracking RAG framework inspired by the path navigational planning of place cells in neurobiology. It consists of two steps: Dynamic Path Tracking and Post-retrieval Completion. Dynamic Path Tracking performs goal-directed semantic path tracking and pruning over the constructed knowledge graph (KG), improving noise reduction and semantic coherence. Post-retrieval Completion further reinforces these benefits by conducting second-stage retrieval using intermediate reasoning and the original query to refine the query goal and complete missing information in the reasoning path. NeuroPath surpasses current state-of-the-art baselines on three multi-hop QA datasets, achieving average improvements of 16.3% on recall@2 and 13.5% on recall@5 over advanced graph-based RAG methods. Moreover, compared to existing iter-based RAG methods, NeuroPath achieves higher accuracy and reduces token consumption by 22.8%. Finally, we demonstrate the robustness of NeuroPath across four smaller LLMs (Llama3.1, GLM4, Mistral0.3, and Gemma3), and further validate its scalability across tasks of varying complexity. Code is available at https://github.com/KennyCaty/NeuroPath.

cs.IR↗

arXiv2025

DSAS: A Universal Plug-and-Play Framework for Attention Optimization in Multi-Document Question Answering

While large language models (LLMs) show considerable promise across various fields, they have notable limitations in handling multi-document question answering (Multi-doc QA) tasks. The first challenge is long-range dependency modeling, where LLMs struggle to focus on key information in long texts, which weakens important semantic connections. Second, most LLMs suffer from the ''lost-in-the-middle'' issue, where they have difficulty processing information in the middle of long inputs. Current solutions either truncate global dependencies or demand costly finetuning, ultimately lacking a universal and simple solution for these challenges. To resolve these limitations, we propose Dual-Stage Adaptive Sharpening (DSAS) containing two modules. (i) The Contextual Gate Weighting (CGW) module alleviates ''lost-in-the-middle'' by assessing paragraph relevance through layer-wise attention tracking and position-aware weighting. (ii) The Reciprocal Attention Suppression (RAS) module enhances focus on critical paragraphs by suppressing information exchange between key and irrelevant texts, thus mitigating the limitations in long-range dependency modeling. Notably, DSAS functions as a plug-and-play solution requiring no architectural modifications or extra training parameters. Extensive experiments on four benchmarks demonstrate DSAS's efficacy across mainstream LLMs (Llama, Qwen, Mistral, and Deepseek), with an average F1-score improvement of 4.2% in Multi-doc QA tasks on Llama-3.1-8B-Instruct and Qwen2.5-14B-Instruct. Ablation studies confirm the essential contributions of both the CGW and RAS modules. In addition, detailed discussions in the Appendix further validate the robustness and scalability of DSAS.

cs.CL↗

arXiv2025

GraphCogent: Mitigating LLMs' Working Memory Constraints via Multi-Agent Collaboration in Complex Graph Understanding

Large language models (LLMs) show promising performance on small-scale graph reasoning tasks but fail when handling real-world graphs with complex queries. This phenomenon arises from LLMs' working memory constraints, which result in their inability to retain long-range graph topology over extended contexts while sustaining coherent multi-step reasoning. However, real-world graphs are often structurally complex, such as Web, Transportation, Social, and Citation networks. To address these limitations, we propose GraphCogent, a collaborative agent framework inspired by human Working Memory Model that decomposes graph reasoning into specialized cognitive processes: sense, buffer, and execute. The framework consists of three modules: Sensory Module standardizes diverse graph text representations via subgraph sampling, Buffer Module integrates and indexes graph data across multiple formats, and Execution Module combines tool calling and tool creation for efficient reasoning. We also introduce Graph4real, a comprehensive benchmark that contains four domains of real-world graphs (Web, Transportation, Social, and Citation) to evaluate LLMs' graph reasoning capabilities. Our Graph4real covers 21 different graph reasoning tasks, categorized into three types (Structural Querying, Algorithmic Reasoning, and Predictive Modeling tasks), with graph scales up to 10 times larger than existing benchmarks. Experiments show that Llama3.1-8B based GraphCogent achieves a 50% improvement over massive-scale LLMs like DeepSeek-R1 (671B). Compared to state-of-the-art agent-based baseline, our framework outperforms by 20% in accuracy while reducing token usage by 80% for in-toolset tasks and 30% for out-toolset tasks. Code will be available after review.

cs.AI↗

arXiv2024

GraphTool-Instruction: Revolutionizing Graph Reasoning in LLMs through Decomposed Subtask Instruction

Large language models (LLMs) have been demonstrated to possess the capabilities to understand fundamental graph properties and address various graph reasoning tasks. Existing methods fine-tune LLMs to understand and execute graph reasoning tasks by specially designed task instructions. However, these Text-Instruction methods generally exhibit poor performance. Inspired by tool learning, researchers propose Tool-Instruction methods to solve various graph problems by special tool calling (e.g., function, API and model), achieving significant improvements in graph reasoning tasks. Nevertheless, current Tool-Instruction approaches focus on the tool information and ignore the graph structure information, which leads to significantly inferior performance on small-scale LLMs (less than 13B). To tackle this issue, we propose GraphTool-Instruction, an innovative Instruction-tuning approach that decomposes the graph reasoning task into three distinct subtasks (i.e., graph extraction, tool name identification and tool parameter extraction), and design specialized instructions for each subtask. Our GraphTool-Instruction can be used as a plug-and-play prompt for different LLMs without fine-tuning. Moreover, building on GraphTool-Instruction, we develop GTools, a dataset that includes twenty graph reasoning tasks, and create a graph reasoning LLM called GraphForge based on Llama3-8B. We conduct extensive experiments on twenty graph reasoning tasks with different graph types (e.g., graph size or graph direction), and we find that GraphTool-Instruction achieves SOTA compared to Text-Instruction and Tool-Instruction methods. Fine-tuned on GTools, GraphForge gets further improvement of over 30% compared to the Tool-Instruction enhanced GPT-3.5-turbo, and it performs comparably to the high-cost GPT-4o. Our codes and data are available at https://anonymous.4open.science/r/GraphTool-Instruction.

cs.LG↗