arXiv subjects

Menghao Zhang

Publications and source records attributed to Menghao Zhang.

6 recordsLinked to original sources

arXiv2026

TRACE: A Self-Evolving Skill Bank for Consistent, Limit-Aware LLM Agents

Reliable deployment of LLM agents in user-facing products depends not on raw task-solving ability but on consistency and limit-awareness: behaving the same way across repeated trials, and recognizing when a request cannot, or cannot yet, be safely fulfilled. CAR-bench exposes this reliability gap in the domain of in-car assistants: an LLM-simulated user issues incomplete or ambiguous requests, requiring the agent to resolve uncertainty through multi-turn dialogue and tool use while strictly adhering to domain policies. Even frontier models show a substantial gap between what they can solve at least once (Pass@3) and what they solve consistently across trials (Pass^k). We bridge this gap with TRACE (TRAjectory-Contrastive Evolution), which iteratively improves a skill-based agent's behavioral knowledge without modifying model weights. This knowledge is organized as a Skill Bank of modular, retrievable skills, each encoding a self-contained set of tool-use rules and behavioral guidelines. TRACE evolves this bank through an agentic self-evolution loop: after each evaluation round, it groups trajectories by the skills invoked and refines each skill by contrasting successful and failed behaviors. The updated bank then guides subsequent rounds, while during deployment the Actor performs state-conditioned skill orchestration at every turn. On GPT-5.5, TRACE improves consistency (Pass^3) by 34.6 points, from 59.9% to 94.5%, while shrinking the gap between potential and reliable performance to just 4.0 points. On the official hidden set, TRACE achieved first place using GPT-5.6-Sol, attaining a Pass^3 score of 70%-a 40% relative improvement over the baseline. These results show that TRACE converts high model potential into stable, consistent performance gain. Project homepage: https://darwin-agent.github.io/Car-bench-TRACE.

cs.CL↗

arXiv2026

Collective Communication for Distributed LLM Systems: Planning, Runtime Adaptation, and Computation Coordination

Distributed large language model (LLM) systems increasingly rely on collective communication primitives such as AllReduce (AR), ReduceScatter (RS), AllGather (AG), and AlltoAll (A2A). In modern LLM training and serving clusters, heterogeneous GPU interconnects, multi-NIC networking, mixed parallelism strategies, low-latency inference requests, and high-throughput training pipelines have motivated increasingly diverse ways to plan, execute, and overlap collective communication. This paper presents a tutorial-style, collective-centric taxonomy for collective communication. We organize recent advances into three layers: communication planning, which generates topology-aware collective schedules; communication execution and adaptation, which maps these schedules onto GPU runtimes and hardware in real clusters; and computation-communication coordination, which turns collective optimization into end-to-end training and inference benefits. We further discuss open challenges and future opportunities for collective communication in distributed LLM systems.

cs.DC↗

arXiv2026

SpecBox: Speculative Sandbox Scheduling for Efficient LLM Agent Serving

As LLM agents increasingly rely on the Model Context Protocol (MCP) to invoke isolated external sandboxes, disaggregated sandbox deployment introduces a fundamental tension between resource utilization and interactive tail latency. Persistent long-lived sandbox reservations incur excessive memory overhead at scale, while lazy on-demand instantiation generates severe cold-start penalties that degrade response performance under multi-tenant, multi-turn agent workloads. To resolve this dilemma, we present SpecBox, a runtime built around speculative sandbox preallocation tailored for dynamic LLM agent execution pipelines. At its core, SpecBox implements keyword matching and streaming semantic embedding to enable intent-driven sandbox prewarming, which identifies pending tool execution demands mid-LLM token generation and fully overlaps sandbox bootstrapping with model inference. To extend prewarming windows across sequential agent steps, the framework leverages context-aware stochastic prefetching atop a sandbox dependency graph to probabilistically forecast future sandbox switches ahead of execution. We complement these speculative mechanisms with two orthogonal optimizations: a semantic result cache that prunes redundant repeated sandbox invocations, and a dedicated out-of-band shared-memory transport plane that bypasses conventional network serialization to deliver zero-copy artifact transfers. Evaluated on high-concurrency multi-turn agent traces, our prototype demonstrates that SpecBox cuts P99 end-to-end latency by up to $2.9\times$ relative to the on-demand sandbox baseline, while slashing peak memory consumption by $45.9\%$ compared to permanently reserved sandbox deployments.

cs.DC↗

arXiv2026

An Efficient, Reliable and Observable Collective Communication Library in Large-scale GPU Training Clusters

Large-scale LLM training requires collective communication libraries to exchange data among distributed GPUs. As a company dedicated to building and operating large-scale GPU training clusters, we encounter several practical limitations of NCCL in production, including 1) SM competition between computation and communication, 2) expensive restart costs under link failures, and 3) insufficient observability of transient collective communication anomalies. To address these challenges, we propose VCCL, an efficient, reliable, and observable collective communication library in large-scale GPU training clusters. VCCL removes SM-consuming P2P kernels by moving intra-node data movement and stream dependency enforcement to CPU threads and GPU copy engines. VCCL also introduces a primary-backup QP mechanism to tolerate frequent NIC port failures, and designs a window-based monitor to observe network anomalies at O(μs) level. We opensource VCCL and deploy it in production training clusters for several months. Compared with NCCL, VCCL improves training throughput by up to 5.28% and reduces massive GPU resource wastage through runtime fault tolerance and finegrained monitor. We also share experience and lessons we learned during the deployment of VCCL in large-scale clusters.

cs.DC↗

arXiv2026

PlexRL: Cluster-Level Orchestration of Serviceized LLM Execution for RLVR

Reinforcement learning with verifiable rewards (RLVR) has recently unlocked strong reasoning capabilities in large language models (LLMs), triggering rapid exploration of new algorithms and data. However, RLVR training is notoriously inefficient: long-tailed rollouts, tool-induced stalls, and asymmetric resource requirements between rollout and training introduce substantial idle time that cannot be eliminated by job-local optimizations such as synchronous pipelining, asynchronous rollout, or colocated execution. We argue that this inefficiency is structural. While idle gaps are unavoidable within individual RLVR jobs, they are largely anti-correlated across jobs and therefore exploitable at the cluster level. Leveraging this observation, we present PlexRL, a cluster-level runtime for multiplexing unified LLM services across RLVR jobs. By centrally managing model placement, state transitions, and function-level scheduling under strict affinity constraints, PlexRL time-slices LLM execution across jobs to fill otherwise idle periods without expensive model migration. Our implementation and evaluations demonstrate that PlexRL significantly improves effective cluster capacity and reduces user GPU hour cost by maximum 37.58% while preserving algorithmic flexibility and introducing minimal per-job overhead.

cs.DC↗

arXiv2016

Expenditure Aware Rating Prediction for Recommendation

The rating score prediction is widely studied in recommender system, which predicts the rating scores of users on items through making use of the user-item interaction information. Besides the rating information between users and items, lots of additional information have been employed to promote recommendations, such as social relation and geographic location. Expenditure information on each transaction between users and items is widely available on e-commerce websites, often appearing next to the rating information, while there is seldom study on the correlation between expenditures and rating scores. In this paper, we first study their correlations in real data sets and propose the expenditure aware rating prediction problem. From the data sets crawled from a well-known social media platform Dianping in China, we find some insightful correlations between expenditures and rating scores: 1) transactions or experiences with higher expenditures usually lead to higher rating scores; 2) when the real expenditures are higher than users' normal spending behavior, the users usually give higher scores; and 3) there are multiple grades of expenditure behaviors. Based on these three observations, we propose an Expenditure ware RatingPrediction method (EARP), based on low-rank matrix factorization, to effectively incorporate the expenditure information. Extensive experiments on five real data sets show that EARP not only always outperforms other state-of-the-art baselines but also discovers the latent characteristics of users and businesses.

cs.SI↗