Searcharxiv⌕ Search

arXiv subjects

Jishen Zhao

Publications and source records attributed to Jishen Zhao.

At least 19 recordsLinked to original sources

Multi-Agent Memory Inconsistency: Where Factual Conflicts Disappear in LLM Agents

Multi-agent delegation lets LLM systems distribute evidence collection across workers with separate contexts. This creates a consistency challenge: facts retrieved by workers may be missing from the evidence available to the lead for a joint conclusion. We introduce a benchmark and evidence-flow profiler for studying multi-agent memory inconsistency through controlled factual conflicts in Harvey LAB document-review tasks. Pooled across three models with comparable generation spending, requiring delegation increases retrieval of both conflicting claims from 50.2% to 68.7% relative to single-agent execution, yet conflict detection falls from 16.2% to 13.2%. Within this setting, workers report both claims for only 35.2% of retrieved pairs, and pooled matched comparisons show higher detection than single-agent execution when one worker reads both claims but lower detection when the claims are split across workers. Worker-to-lead handoffs are a central target for improving conflict detection in multi-agent LLMs.

cs.AI↗

The Physical Cost of Context: Quality-Constrained Resource Accounting for Coding Agents

Coding agents resend a growing context on every model call, and context policies are usually judged by the prompt tokens they save. Executed work also depends on call count and prefix reuse: a caching runtime prefills only the tokens after the longest cached prefix. A shorter prompt can therefore cost more if it rewrites cached history. Our quality-gated protocol splits executed prefill into logical prompt minus reused prefix and measures KV residency, time, and energy. On 450 localization cases derived from SWE-rebench V2, preloading ten retrieved blocks (Eager) and adding anchored history compaction (Compact) both pass a five-point non-inferiority gate. In a separate 900-cell study, Eager cuts model calls by 35.7% and the GPU-energy proxy by 33.2% against on-demand search (Grep); Compact saves nothing further. Enabling prefix caching leaves Eager's prompt volume similar (+8.3%; its 95% CI spans zero) but cuts its uncached prefill by 67.7%. Deleting history can turn a token saving into extra work. With mini-SWE-agent on 100 SWE-bench Verified instances, keeping only the last five observations (Mask-5) cuts Eager's prompt volume by 33.3% but loses 135.9k reusable tokens per task. Uncached prefill rises by 608.5%, the energy proxy by 125.6%, and the policy misses the quality gate. Policies thus save physical work by shortening the trajectory and can multiply it by editing reusable history.

cs.SE↗

TileWeave: Heterogeneous Tiling for Sparse Prefill on Mobile SoCs

Smartphone applications increasingly require large language models to process long contexts, whereas high prefill latency and memory demands constrain local inference. Block-sparse attention reduces computation and memory traffic, but its irregular selections conflict with mobile NPUs' preference for large, regular tiles. Efficient execution must exploit cross-query reuse while avoiding redundant computation and preserving selections needed for accuracy. In this paper, we present TileWeave, which partitions sparse attention by this cross-query reuse: highly shared key blocks run in coarse NPU tiles, while low-reuse blocks run on the GPU only for the query blocks that selected them. In addition, we design an NPU-native pooled-key selector that keeps selection inexpensive, and a runtime that publishes GPU partials in the order the NPU consumes them and merges them in its existing output store. On RULER, HeteroSparse matches XAttention's accuracy with Qwen3-1.7B and improves it with Llama-3.2-3B-Instruct; its mean LongBench scores match XAttention as well. On Snapdragon SM8750 and SM8850 phones, it achieves 1.35--3.23$\times$ faster full-model prefill than XAttention over 1k--16k tokens and up to 1.58$\times$ faster than dense attention.

cs.LG↗

RankLane: Efficient MoE Prefill on Accelerators with Ranked Expert Capacities

Mixture-of-experts (MoE) models expand model capacity by activating only a subset of experts for each token. On accelerators with static execution plans, expert capacities and placement are fixed before routing reveals each expert's load. Uniform worst-case capacities waste work on padding; smaller capacities tied to expert identities can overflow. We present RankLane, which ranks each device's experts by current load and assigns them to fixed-capacity lanes. Runtime indices select experts and token rows while tier populations, shapes and device placement remain fixed. An aggregation graph preserves device ownership and reduces each token's local contributions on one core. Routing calibration and measured execution costs guide tier selection; overflowing chunks rerun at full capacity. We evaluate full-model prefill of Qwen3-30B-A3B and Gemma 4 26B-A4B with MXFP6 weights on four Cloud AI 100 Ultra SoCs, using 128-2048-token chunks and including overflow detection and in-program recovery support. Against a dense baseline, RankLane achieves end-to-end prefill speedups of 1.27-4.18x on Qwen3 and 1.93-3.46x on Gemma.

cs.LG↗

EvalResearchBench: Can AI Agents Build Their Own Feedback for Recursive Self-Improvement?

Recursive self-improvement (RSI) relies on evaluation feedback to assess progress and guide further research, yet repeatedly running complex benchmarks is costly and slows iteration. Human experts reduce this cost by selecting benchmark subsets or designing compact suites. We ask whether AI agents can automate this design process and introduce EvalResearchBench (ERB), a benchmark for autonomous evaluation research. Given target materials, development references, candidate APIs, and fixed time and API budgets, an agent called the researcher selects or synthesizes tasks, implements graders, and revises them in pilot tests before freezing an executable evaluator for coding, co-work, and reasoning. We study 9 researchers and 13 candidate models and compare each frozen evaluator with 14 target benchmarks on score concordance and pairwise agreement. The best evaluators order about 75\% of candidate pairs as the targets do, below the 91\% ceiling set by disagreements among the targets. No researcher leads on every metric, and the best evaluator on development targets is not the best on sealed targets hidden from the researcher. A human-designed sample of public tasks remains a strong baseline, and the evaluator with the lowest recorded execution cost attains the highest pairwise agreement. Agents repair tasks and graders through pilot feedback, yet their evaluators can still truncate answers, exhaust the evaluation budget, or let a few questions dominate a domain score.

cs.AI↗

CodeNib: A Multi-View Data System for Serving Repository Context to Coding Agents

An incremental code index can complete every update without error and still return the wrong graph half the time. Coding agents depend on such reused views for search and navigation, yet the systems that supply them never measure where a precomputed answer may stand in for the trusted route: localization agents rebuild a graph per task, tool servers query a live language server per request, and code-intelligence databases serve developers, not agents. CodeNib is a multi-view data system that makes that measurement possible. It is the first to compile lexical, dense, and structural views of one commit behind a single manifest and source-address contract, so that a lexical hit, a graph occurrence, and a dense block can be checked against each other and against a rebuild or a live server, and it serves search, navigation, and bounded context to agents through one cost-visible runtime. The comparisons overturn three assumptions. Execution status is not correctness: symbol-level repair completes on every held-out change but reproduces the rebuild on 13 of 24, so the runtime admits it per repository, while vector reuse is exact on 27 of 30 held-out changes at 40$\times$ and static definitions match the live server on 86\% of requests at a 1.2\,s ready point. Shorter answers finish more tasks: behind LocAgent, CodeNib's views let the agent finish 78 of 90 runs against 65 at lower cost, and selected context policies use 50--87\% fewer tokens than grep/read across five models. And context pays most when it is taken back: on 200 SWE-bench Verified tasks, handing the agent CodeNib's candidates and retiring them after its first source read resolves 79 issues, the most of five context policies, against 66 for on-demand exploration and 56 for Aider's repository map.

cs.SE↗

Dynamic HBM Repartitioning for Multi-Turn MoE Serving

Long-running multi-turn requests accumulate reusable key-value (KV) state. Once this state exceeds a fixed GPU KV-cache allocation, serving systems evict reusable prefixes, repeat prefill work, and may preempt requests. This pressure is particularly acute for Mixture-of-Experts (MoE) models: their expert weights occupy most GPU high-bandwidth memory (HBM), even though each token activates only a sparse subset of experts. A static boundary between weights and the KV cache prevents serving systems from using expert memory to preserve reusable state as conversations grow. We present VAMP, an MoE serving framework that changes this HBM boundary at runtime. When a KV allocation cannot be satisfied, VAMP compares the estimated future work of three alternatives: staging expert weights from host memory, evicting cached prefixes that may require re-prefill, or preempting and rescheduling requests. It then converts a bounded expert-weight region into KV-cache capacity when that action has the lowest estimated penalty. CUDA Virtual Memory Management page remapping performs this conversion without copying resident KV data. We implement VAMP in the vLLM serving engine and evaluate Qwen3-Next-80B on recorded and controlled multi-turn workloads. Across five replays of a recorded 2,103-turn SWE-bench agent workload, VAMP with a 15% maximum expert-offloading ratio reduces time-to-first-token (TTFT) p90 from 26.1 s to 1.10 s (23.6 times) and increases request throughput by 20.7% relative to unmodified vLLM, while increasing time per output token (TPOT) by 31.1%.

cs.DC↗

CHORUS: Complementary Experts for High-Coverage Testbench Stimulus Generation

Large language models (LLMs) have advanced code generation, where executable feedback provides a more reliable learning signal than textual imitation alone. Hardware verification is an important application of code generation and accounts for a substantial fraction of modern chip design effort, with high-coverage testbench stimulus generation as a key task. We present CHORUS, a post-training framework that pushes performance beyond what a conventional supervised fine-tuning (SFT)-to-reinforcement learning (RL) pipeline achieves. CHORUS builds on two observations. First, staged SFT produces behaviorally diverse checkpoints, and dense-reward RL turns them into strong experts with comparable aggregate performance but distinct task-level strengths. Second, these complementary strengths can be exploited through either training-free model merging or further post-training to outperform the best individual expert. By consolidating the resulting specialists into a single 4B model, CHORUS achieves 88.0% Pass@1 on CVDP-ECov, outperforming DeepSeek-R1 (671B) by 13.5 percentage points.

cs.AI↗

LLM4Cov: Execution-Aware Agentic Learning for High-coverage Testbench Generation

Execution-aware LLM agents offer a promising paradigm for learning from tool feedback, but such feedback can be expensive and slow to obtain, making online reinforcement learning (RL) less practical in certain scenarios. High-coverage hardware verification exemplifies this challenge due to its reliance on industrial simulators and non-differentiable execution signals. We propose LLM4Cov, an offline agent-learning framework that models verification as single-step state transitions guided by deterministic evaluators. Building on this formulation, we introduce execution-validated data curation, policy-aware agentic data synthesis, and worst-state-prioritized sampling to enable scalable learning under execution constraints. We further curate a reality-aligned benchmark adapted from an existing verification suite through a revised evaluation protocol. Using the proposed pipeline, a compact 4B-parameter model achieves 69.2% pass rate and 90.4% average coverage in CVDP-ECov under agentic evaluation, outperforming its teacher by 5.3% and 10.5%, demonstrating competitive performance against models an order of magnitude larger.

cs.AI↗

AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications

Large Language Models (LLMs) are increasingly used as autonomous agents in complex, long-horizon applications, where effective memory is critical for sustained performance. Yet existing memory benchmarks are largely dialogue-centric, while real agent memory consists of continuous agent-environment interaction trajectories composed of states, actions, observations, and tool outputs. To address this gap, we introduce **AMA-Bench** (**A**gent **M**emory with **A**ny length), a benchmark for evaluating long-horizon memory in realistic agentic settings. AMA-Bench combines real-world agent trajectories from representative applications with expert-curated QA, as well as synthetic trajectories that scale to arbitrary horizons with rule-based QA. Our study shows that existing memory systems underperform because they fail to capture causal and objective information and rely heavily on lossy similarity-based retrieval. We further propose **AMA-Agent**, a memory system based on causality-graph construction and tool-augmented retrieval. AMA-Agent achieves **57.22%** accuracy on AMA-Bench, outperforming the strongest baseline by **11.16%**. Resources are available at: [https://ama-bench.github.io/](https://ama-bench.github.io/).

cs.AI↗

MetaAgent-X : Breaking the Ceiling of Automatic Multi-Agent Systems via End-to-End Reinforcement Learning

Automatic multi-agent systems aim to instantiate agent workflows without relying on manually designed or fixed orchestration. However, existing automatic MAS approaches remain only partially adaptive: they either perform training-free test-time search or optimize the meta-level designer while keeping downstream execution agents frozen, which creating a frozen-executor ceiling and leaving the end-to-end training of self-designing and self-executing agentic models unexplored. To address this, we introduce MetaAgent-X, an end-to-end reinforcement learning framework that jointly optimizes automatic MAS design and execution. MetaAgent-X enables script-based MAS generation, execution rollout collection, and credit assignment for both designer and executor trajectories. To support stable and scalable optimization, we propose Executor Designer Hierarchical Rollout and Stagewise Co-evolution to improve training stability and expose the dynamics of designer-executor co-evolution. MetaAgent-X consistently outperforms existing automatic MAS baselines, achieving up to 21.7% gains. Comprehensive ablations show that both designer and executor improve throughout training, and that effective automatic MAS learning follows a stagewise co-evolution process. These results establish end-to-end trainable automatic MAS as a practical paradigm for building self-designing and self-executing agentic models.

cs.AI↗

ChipMATE: Multi-Agent Training via Reinforcement Learning for Enhanced RTL Generation

Existing API-based agentic systems for RTL code generation are fundamentally misaligned with industrial practice: they assume a golden testbench is available at generation time, rely on closed-source APIs incompatible with chip vendors' air-gapped security requirements, and cannot be trained on vendors' proprietary RTL codebases, leaving valuable internal data unused. Recent self-trained models address the deployment constraint but remain single-turn generators that overlook the critical role of verification in real industrial flows. To bridge these gaps, we present ChipMATE, the first self-trained multi-agent framework for RTL generation. Inspired by industrial practice where correctness emerges from cross-comparison between independently written RTL modules and reference models, ChipMATE pairs a Verilog agent with a Python reference-model agent that mutually verify each other's outputs without any golden oracle. We design a backtrack-based inference workflow to prevent error propagation across turns, and a two-stage training pipeline that first trains each agent individually to saturate its code-generation capability, then trains the team jointly to collaborate effectively. To support the training, we further build a hybrid data-generation framework that produces 64.4K high-quality reference model training samples. ChipMATE achieves 75.0\% and 80.1\% pass@1 on VerilogEval V2 with 4B and 9B base models, outperforming all existing self-trained models and even DeepSeek V4 with 1600B parameters. Our code and model weights are publicly available in https://github.com/zhongkaiyu/ChipMATE.

cs.MA↗

Multi-Agent Memory from a Computer Architecture Perspective: Visions and Challenges Ahead

As LLM agents evolve into collaborative multi-agent systems, their memory requirements grow rapidly in complexity. This position paper frames multi-agent memory as a computer architecture problem. We distinguish shared and distributed memory paradigms, propose a three-layer memory hierarchy (I/O, cache, and memory), and identify two critical protocol gaps: cache sharing across agents and structured memory access control. We argue that the most pressing open challenge is multi-agent memory consistency. Our architectural framing provides a foundation for building reliable, scalable multi-agent systems.

cs.AR↗

Double-P: Hierarchical Top-P Sparse Attention for Long-Context LLMs

As long-context inference becomes central to large language models (LLMs), attention over growing key-value caches emerges as a dominant decoding bottleneck, motivating sparse attention for scalable inference. Fixed-budget top-k sparse attention cannot adapt to heterogeneous attention distributions across heads and layers, whereas top-p sparse attention directly preserves attention mass and provides stronger accuracy guarantees. Existing top-p methods, however, fail to jointly optimize top-p accuracy, selection overhead, and sparse attention cost, which limits their overall efficiency. We present Double-P, a hierarchical sparse attention framework that optimizes all three stages. Double-P first performs coarse-grained top-p estimation at the cluster level using size-weighted centroids, then adaptively refines computation through a second top-p stage that allocates token-level attention only when needed. Across long-context benchmarks, Double-P consistently achieves near-zero accuracy drop, reducing attention computation overhead by up to 1.8x and delivers up to 1.3x end-to-end decoding speedup over state-of-the-art fixed-budget sparse attention methods.

cs.LG↗

ChipBench: A Next-Step Benchmark for Evaluating LLM Performance in AI-Aided Chip Design

While Large Language Models (LLMs) show significant potential in hardware engineering, current benchmarks suffer from saturation and limited task diversity, failing to reflect LLMs' performance in real industrial workflows. To address this gap, we propose a comprehensive benchmark for AI-aided chip design that rigorously evaluates LLMs across three critical tasks: Verilog generation, debugging, and reference model generation. Our benchmark features 44 realistic modules with complex hierarchical structures, 89 systematic debugging cases, and 132 reference model samples across Python, SystemC, and CXXRTL. Evaluation results reveal substantial performance gaps, with state-of-the-art Claude-4.5-opus achieving only 30.74\% on Verilog generation and 13.33\% on Python reference model generation, demonstrating significant challenges compared to existing saturated benchmarks where SOTA models achieve over 95\% pass rates. Additionally, to help enhance LLM reference model generation, we provide an automated toolbox for high-quality training data generation, facilitating future research in this underexplored domain. Our code is available at https://github.com/zhongkaiyu/ChipBench.git.

cs.AI↗

Stronger-MAS: Multi-Agent Reinforcement Learning for Collaborative LLMs

Multi-agent systems (MAS) and reinforcement learning (RL) are widely used to enhance the agentic capabilities of large language models (LLMs). MAS improves task performance through role-based orchestration, while RL uses environmental rewards to learn stronger policies, such as GRPO-style optimization. However, applying on-policy RL to MAS remains underexplored and presents unique challenges. Algorithmically, standard GRPO grouping assumptions break down because prompts vary by role and by turn. System-wise, the training stack must support MAS-workflow rollouts and on-policy updates for both single-policy and multi-policy models. We propose AT-GRPO, which includes (i) an agent- and turn-wise grouped RL algorithm tailored to MAS and (ii) a training system that supports both single- and multi-policy regimes. Across game, planning, coding, and math tasks, AT-GRPO delivers substantial gains. On long-horizon planning, it increases accuracy from a 14.0 to 47.0 percent single-agent RL baseline to 96.0 to 99.5 percent. It also improves reasoning performance, with average gains of 3.87 to 7.62 percent on coding tasks and 9.0 to 17.93 percent on math. Code and environments are available at: https://github.com/pettingllms-ai/PettingLLMs.

cs.LG↗

PRO-V-R1: Reasoning Enhanced Programming Agent for RTL Verification

Register-Transfer Level (RTL) verification is a primary bottleneck, consuming 60-70% of development time. While Large Language Models (LLMs) show promise for RTL automation, their performance and research focus have overwhelmingly centered on RTL generation rather than verification. Current methods for RTL verification rely on large scale proprietary models (e.g., GPT-4o) to generate Python-based functional references, incurring a high cost and raising data-privacy risks. To date, an end-to-end open-source solution for autonomous verification remains absent. We introduce PRO-V-R1, the first trainable open-source agentic framework for autonomous RTL verification. Our contributions are threefold: (1) we design PRO-V sys, a modular agentic system that couples LLM-based reasoning with programmatic tool use for RTL verification; (2) we establish a data construction pipeline that leverages existing RTL datasets to build simulation-validated, expert-level trajectories tailored for supervised fine-tuning (SFT) RTL verification agents; and (3) we implement an efficient reinforcement learning (RL) algorithm that uses verification-specific rewards derived from program-tool feedback to optimize the end-to-end verification workflow. Our empirical evaluation demonstrates PRO-V-R1 achieves a 57.7% functional correctness rate and 34.0% in robust fault detection, significantly outperforming the base model's 25.7% and 21.8% (respectively) from the state-of-the-art (SOTA) automatic verification system. This configuration also outperforms large-scale proprietary LLMs in functional correctness and shows comparable robustness for fault detection.

cs.AI↗

OrcaLoca: An LLM Agent Framework for Software Issue Localization

Recent developments in Large Language Model (LLM) agents are revolutionizing Autonomous Software Engineering (ASE), enabling automated coding, problem fixes, and feature improvements. However, localization -- precisely identifying software problems by navigating to relevant code sections -- remains a significant challenge. Current approaches often yield suboptimal results due to a lack of effective integration between LLM agents and precise code search mechanisms. This paper introduces OrcaLoca, an LLM agent framework that improves accuracy for software issue localization by integrating priority-based scheduling for LLM-guided action, action decomposition with relevance scoring, and distance-aware context pruning. Experimental results demonstrate that OrcaLoca becomes the new open-source state-of-the-art (SOTA) in function match rate (65.33%) on SWE-bench Lite. It also improves the final resolved rate of an open-source framework by 6.33 percentage points through its patch generation integration.

cs.SE↗