arXiv subjects

Huazheng Wang

Publications and source records attributed to Huazheng Wang.

At least 19 recordsLinked to original sources

arXiv2026

When Does Multi-Agent RL Improve LLM Workflows? Workflow, Scale, and Policy-Sharing Tradeoffs

Outcome-reward RL can train every role of a multi-agent LLM workflow. Its benefit is hard to predict because the reward assigns no credit to individual roles and a shared policy trades pooled data against specialization. We ask when this training improves a workflow over its base model and single-agent RL, and what each role learns. We train three workflows (Eval-Opt, Voting, Orch-Workers) with GRPO on math and code at three Qwen3 scales, with one policy for all roles (Shared-Policy, SP) or one per role type (Isolated-Policy, IP). Every workflow gains by adapting to its structure; this does not make a role better than training it alone. Single-agent RL in the same structure matches or beats trained Voting and Orch-Workers, and within the grid only Eval-Opt under IP exceeds it. Eval-Opt's evaluator learns when to retry and its generator learns to complete the retried attempts, a joint strategy that a single policy trained alone does not acquire. Under SP, per-output token averaging sets each role's gradient share, a hidden hyperparameter absent from single-agent RL, and the solving roles learn slowly; reweighting it gives SP the speed and peak of IP, and its decline. Under IP, a parallel role, whose instances share one adapter, takes one optimizer step per instance per training step and can drift after the peak as its entropy rises and accuracy declines. Merging those steps delays the decline. Our code is at: https://github.com/XHMY/marl-llm-workflows

cs.AI↗

arXiv2026

Computer-Use Agents Gain Little Benefit from a World Model When Conditioned on Tutorials

Recent work has improved computer-use agents (CUAs) by combining retrieved task-relevant tutorials and look-ahead planning using world models with MLLM judges. However, the contribution of each component -- tutorials, the judge, and the world model -- to overall improvement has not been evaluated in isolation. To fill this gap, we first construct OSWorld-Tutorial and ScienceBoard-Tutorial: two datasets for CUAs with curated multimodal tutorials, covering everyday and scientific domains. We then propose RAG-PolicyJudge (RAG-PJ), a policy-judge framework for CUAs conditioned on task-specific tutorials: the policy proposes candidate actions, and the judge selects among them. To isolate the effect of world models given tutorials, we compare RAG-PJ with and without common world models: using either zero-shot text-based descriptions of next states or an oracle that returns ground-truth visual observations. We find that RAG-PJ gains little benefit from these rollouts when tutorials are present, even with perfect visual futures. This suggests prior world-model gains in tutorial-conditioned methods come mainly from tutorial understanding and the judge's ability to select the proposed actions, rather than from predicted rollouts. Error analysis reveals that the bottleneck is not rollout fidelity, but the judge's ability to reason over future states and recognize task completion. This highlights the need to improve not just next-state prediction in world models, but also reasoning over their predicted rollouts. Datasets and code are available at https://github.com/realjoenguyen/RAGPJ-official.

cs.LG↗

arXiv2026

Lifelong Context Adaptation through Bidirectional Memory-Skill Learning

Lifelong LLM agents must learn from an expanding history of interactions while making each decision within a bounded context window. Memory makes past experience available, but the agent may still need to infer how to solve each new task from the retrieved records. Skills provide reusable procedures, yet may omit the task-specific facts needed to apply them. Because retrieved memory and skills shape the agent's behavior, they also influence the experience available for future learning. We introduce MemCryst, a framework for lifelong context adaptation through bidirectional learning of memory and skills with a frozen language model. Recurring episodes stored in memory support skill creation and revision, and the interactions guided by those skills supply new memory for subsequent learning. Facts retrieved from memory and learned skills jointly form the agent's context, while task outcomes adjust the future retrieval influence of the records actually served without rewriting their content. Across ten settings spanning embodied control, long-document reading, conversation, mathematics, and code, MemCryst leads seven reported comparisons. On ALFWorld, joint serving achieves 63.17% success, compared with 59.83% for memory-only and 50.50% for skill-only serving. Code is available at https://github.com/tianyi0216/Memcryst.git

cs.AI↗

arXiv2026

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↗

arXiv2026

Verifier-Harness Co-Evolution for Training-Free Self-Improvement on Open-Ended Tasks

Training-free self-improvement lets an agent use feedback to revise the response it is about to return and to update the harness state that shapes how it handles later tasks. On open-ended tasks that feedback rarely comes from reference answers or executable tests, so the agent has to rely on a model-based verifier. When the verifier is wrong, the current response can get worse, and misleading experience can persist across the task stream. A verifier that never changes also grows stale as the agent's responses improve and task demands shift. We introduce Unified Co-Evolution (UCoE), a training-free method that evolves an executable verifier program together with the agent's harness state while all model parameters stay fixed. UCoE synthesizes candidate verifier programs and tests them on response contrasts, pairs of responses that differ in one declared property and have a known preferred side. It selects the candidate whose discrimination on its weakest contrast family is highest, a quantity we call driving capacity. The selected verifier diagnoses one weakness in each draft, guides a single repair, decides which response is served, and determines whether the accepted repair may enter the harness state. As tasks complete, their diagnostics seed new candidate verifiers, which must beat the current verifier on the same contrasts before adoption. Across six open-ended benchmarks, UCoE averages 59.00% preference against matched direct generation with Qwen3-8B and 57.19% with Qwen3.8-27B, margins of 9.00 and 7.19 percentage points. On three verifiable benchmarks, the same method raises the mean from 36.06% for direct generation to 38.48%. Code is available at https://github.com/tianyi0216/Unified-Co-Evolution.

cs.AI↗

arXiv2026

Who&When Pro: Can LLMs Really Attribute Failures in AI Agents?

Automated failure attribution uses LLMs to identify where and why agentic systems fail. As agents become more capable, their failures become subtler, making automated attribution increasingly important. We introduce Who&When Pro, a large-scale benchmark for automated failure attribution in agentic systems. Using a strictly controlled pipeline that injects a failure only after exactly replaying a successful prefix, we construct 12,326 failed trajectories with golden labels across 3 modalities and 26 benchmarks covering various scenarios. Beyond benchmarking, we conduct extensive experiments and analyses, revealing systematic patterns in how models attribute failures across modalities, protocols, and model families, and providing empirical guidance for future automated failure attribution systems.

cs.AI↗

arXiv2026

AsynCodeBench: Benchmarking Collaboration of Asynchronous Multi-Agent Systems in Software Engineering

Multi-agent coding has emerged as an increasingly active direction in software engineering, where complex development tasks are decomposed across multiple specialized agents working on different parts of the problem. Despite the shift from individual problem solving to distributed collaboration, multi-agent systems still lack a direct measure of collaboration and are largely evaluated through task-level outcomes inherited from single-agent coding, conflating individual coding capability with cross-agent coordination. We introduce AsynCodeBench, a dependency-centric benchmark for asynchronous multi-agent software engineering that represents each task with an explicit dependency graph and executable Dependency Checkers. Through this dependency-tracking process, we propose two complementary measures: Asynchronous Dependency Pass Rate (ADPR), which measures how many cross-agent dependencies are ultimately satisfied, and Dependency Resolution Step (DRS), which measures when each dependency first becomes satisfied during execution. AsynCodeBench comprises 19 tasks from real-world repositories, exposing 52 directed dependencies as explicit units for evaluating cross-agent collaboration. Experiments across model families, scales, and generations reveal a clear gap between coding and collaboration capability: improvements in coding performance do not necessarily translate into stronger collaboration, and task-level metrics can diverge substantially from dependency-level collaboration measures. Dependency-trajectory analysis further reveals that successful coordination often emerges not gradually, but through concentrated bursts in which many dependencies become resolved over a short portion of the execution trajectory, a pattern we term a hopping window.

cs.SE↗

arXiv2026

When Greedy Sampling Explores: KL-Regularized Contextual Bandits without Eluder-Dimension Dependence

We study KL-regularized contextual bandits under both reward and preference feedback. While existing regret guarantees typically depend on the eluder dimension, we show that simple greedy sampling can achieve polylogarithmic regret without explicit dependence on this complexity measure. For reward feedback, we analyze a greedy algorithm that samples directly from the Gibbs policy induced by the estimated reward. We extend the result to preference feedback under both general preference and Bradley--Terry models, while also sharpening existing dimension-dependent guarantees. Our analysis reveals a trade-off between greedy sampling and upper confidence bound-style exploration: greedy sampling enjoys stronger regret guarantees when KL regularization is sufficiently strong, whereas additional exploration yields sharper bounds as the regularization weakens.

cs.LG↗

arXiv2026

Towards Affordable Energy: A Gymnasium Environment for Electric Utility Demand-Response Programs

Extreme weather and volatile wholesale electricity markets expose residential consumers to catastrophic financial risks, yet demand response at the distribution level remains an underutilized tool for grid flexibility and energy affordability. While a demand-response program can shield consumers by issuing financial credits during high-price periods, optimizing this sequential decision-making process presents a unique challenge for reinforcement learning despite the plentiful offline historical smart meter and wholesale pricing data available publicly. Offline historical data fails to capture the dynamic, interactive feedback loop between an electric utility's pricing signals and customer acceptance and adaptation to a demand-response program. To address this, we introduce DR-Gym, an open-source, online Gymnasium-compatible environment designed to train and evaluate demand-response from the electric utility's perspective. Unlike existing device-level energy simulators, our environment focuses on the market-level electric utility setting and provides a rich observational space relevant to the electric utility. The simulator additionally features a regime-switching wholesale price model calibrated to real-world extreme events, alongside physics-based building demand profiles. For our learning signal, we use a configurable, multi-objective reward function for specifying diverse learning objectives. We demonstrate through baseline strategies and data snapshots the capability of our simulator to create realistic and learnable environments.

cs.AI↗

arXiv2026

Pin Once, Swap Light: Subspace-Aligned Centroid-Residual Training for Efficient Ultra-LoRA Serving

Modern multi-tenant Low-Rank Adapters (LoRAs) serving systems concurrently host tens to hundreds of LoRA adapters. Though powerful, this introduces a critical system dilemma between serving efficiency and task performance: higher-rank adapters generally achieve better downstream task performance, but their GPU VRAM footprint and Host-to-Device PCIe swapping overhead severely constrain scalability. Conversely, ultra-low-rank adapters ($r \le 2$) minimize both VRAM footprint and PCIe transfer overhead, but suffer from downstream task performance degradation. To solve this problem, we propose Subspace-Aligned LoRA Training (SALT), a serving efficiency-aware hierarchical fine-tuning framework. Our solution operates in three phases. First, a provider jointly trains high-capacity domain centroids on public data within the domain using a novel alignment regularizer that coheres in-domain task subspaces into a unified basis. Next, users fine-tune ultra-low-rank task residual adapters on private data atop those frozen centroids. Finally, during inference, the provider pins the centroid in GPU VRAM and dynamically swaps in each user's task residual on demand. Across LLMs of varying scales, SALT recovers high-rank accuracy using $r \le 2$ residuals, achieving up to 18.5% absolute accuracy gains over state-of-the-art compression baselines and reducing per-adapter memory by up to 16x. When integrated into vLLM, SALT improves serving throughput by up to 51% under PCIe bandwidth pressure and 28% under GPU VRAM constraints for Llama-3.2-3B.

cs.LG↗

arXiv2026

From RLVR to RLSVR: Task Transformation Induces Self-Verifiable Rewards for Open-Ended LLM Self-Improvement

Reinforcement Learning with Verifiable Rewards (RLVR) has driven recent progress in reasoning-oriented large language models (LLMs) by enabling large-scale optimization. However, its applicability remains largely limited to domains such as mathematics and coding, where correctness can be deterministically verifiable. Open-ended tasks instead often rely on human preferences, reward models, or LLM-based judges, introducing evaluation bias, judge capability bottlenecks, and additional inference costs. Drawing on the principle of self-supervised learning, which constructs pretext tasks to derive supervision from the data itself, we propose Reinforcement Learning with Self-Verifiable Rewards (RLSVR), a task-transformation-based training paradigm for extending RLVR to open-ended tasks. RLSVR transforms open-ended tasks into verifiable proxy environments whose internal rules and interaction outcomes automatically generate reward signals. We instantiate RLSVR with SpyRL, a Self-PlaY Reinforcement Learning method inspired by social deduction game Who Is the Spy?. Agents receive asymmetric information, complete the same target task, and vote to identify a designated spy. Because the spy identity is predetermined, voting outcomes provide fully verifiable rewards, while successful identification remains closely related to output quality. Experiments on text summarization, creative writing, and mathematical reasoning show that SpyRL outperforms existing self-improvement methods on non-verifiable tasks and yields consistent gains on verifiable reasoning tasks. These results demonstrate that task transformation can extend scalable RLVR-based self-improvement beyond inherently verifiable domains. Models and code have been released at https://github.com/wangqinsi1/RLSVR/tree/SpyRL.

cs.AI↗

arXiv2026

Speculative Pipeline Decoding: Higher-Accuracy Drafting with Hidden Latency via Pipeline Parallelism

Speculative Decoding (SD) accelerates low-concurrency LLM inference with a draft-then-verify paradigm. Mainstream methods, however, rely on multi-token prediction, which incurs compounding prediction difficulty and exposed draft latency. We propose Speculative Pipeline Decoding (SPD), which partitions the target LLM into $n$ pipeline stages so that $n$ tokens of a single sequence advance in parallel. To keep the pipeline saturated, a Pipeline Draft Module (PDM) aggregates multi-depth target features to predict the next token and runs concurrently with each pipeline step, yielding bounded prediction difficulty, higher acceptance, and hidden draft latency. Experiments show that SPD achieves higher theoretical and wall-clock speedup than EAGLE-3 at moderate pipeline width, while more aggressive widths still leave room for further gains. Our code is available at https://github.com/yuyijiong/speculative_pipeline_decoding

cs.CL↗

arXiv2026

Online KL-Regularized Reinforcement Learning with Function Approximation under Misspecification

We study KL-regularized contextual bandits and episodic reinforcement learning (RL) under general function approximation with model misspecification. Existing guarantees rely on realizability and therefore do not extend to misspecified models, where classical regret bounds may fail. This work introduces KL misspecification formulations for contextual bandits and episodic RL and analyzes regression-based algorithms with Gibbs policy updates. High-probability KL-regret guarantees with explicit misspecification terms are established, recovering the standard realizable KL-regularized setting as a special case.

cs.LG↗

arXiv2026

When Can You Poison Rewards? A Tight Characterization of Reward Poisoning in Linear MDPs

We study reward poisoning attacks in reinforcement learning (RL), where an adversary manipulates rewards within constrained budgets to force the target RL agent to adopt a policy that aligns with the attacker's objectives. Prior works on reward poisoning mainly focused on sufficient conditions to design a successful attacker, while only a few studies discussed the infeasibility of targeted attacks. This paper provides the first precise necessity and sufficiency characterization of the attackability of a linear MDP under reward poisoning attacks. Our characterization draws a bright line between the vulnerable RL instances, and the intrinsically robust ones which cannot be attacked without large costs even running vanilla non-robust RL algorithms. Our theory extends beyond linear MDPs -- by approximating deep RL environments as linear MDPs, we show that our theoretical framework effectively distinguishes the attackability and efficiently attacks the vulnerable ones, demonstrating both the theoretical and practical significance of our characterization.

cs.LG↗

arXiv2026

EvoPool: Evolutionary Programmatic Annotation for Label-Efficient Specialized Supervision

Large language models excel at general tasks but underperform smaller supervised models in specialized, high-stakes domains where training labels are costly. We address this regime with EvoPool, an evolutionary multi-agent framework inspired by Darwinian evolution. Three specialized agents iteratively propose executable annotator code, a small validation set provides a fitness signal, and a deterministic gate keeps only annotators that pass viability, diversity, and marginal-contribution checks across generations. Pool votes are mapped to soft training labels by EvoAgg, a text-aware aggregator combining semantic features with annotator-vote features. The authored pool runs at near-zero per-example cost and is 4500 to 31000x faster than LLM annotation on 100K examples. Across 7 of 8 LLM-weak specialized and complex tasks spanning biomedical relation extraction, legal-clause classification, complex reasoning, and dense multi-label biomedical classification, EvoPool beats the strongest LLM annotation baseline by an average +0.141 macro-F1, peaking at +0.301 on ChemProt and +0.265 on PubMed. Code is available at: https://github.com/tianyi0216/EvoPool

cs.CL↗

arXiv2026

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↗

arXiv2026

EVOCHAMBER: Test-Time Co-evolution of Multi-Agent System at Individual, Team, and Population Scales

We argue that multi-agent test-time evolution is not single-agent evolution replicated N times. A single-agent learner can only evolve its own context and memory. A multi-agent system additionally evolves who collaborates, how they collaborate, and how knowledge flows across the population. These components have no single-agent counterpart and can produce phenomena such as emergent specialization. Yet prior test-time methods either confine experiences to individual agents, forfeiting cross-agent learning, or broadcast symmetrically to all agents, erasing the specialization that makes collaboration valuable. We present EVOCHAMBER, a training-free framework that instantiates test-time evolution at three levels over a coevolving agent pool. At its core is CODREAM (Collaborative Dreaming), a post-task protocol triggered on team failure or disagreement, in which agents collaboratively reflect, distill insights, and route them asymmetrically from strong to weak agents on the failed niche, preserving specialization while filling knowledge gaps. Team-level operators assemble niche-conditioned teams and select collaboration structures online. Population-level lifecycle operators fork, merge, prune, and seed agents under performance pressure. On three heterogeneous task streams with Qwen3-8B, EVOCHAMBER reaches 63.9% on competition math, 75.7% on code, and 87.1% on multi-domain reasoning, outperforming the best baseline by 32% relative on math and confirming asymmetric cross-agent transfer as the primary driver in ablation. Starting from several identically initialized agents, four to five stable niche specialists spontaneously emerge, a structural signature of multi-agent evolution that no single-agent learner can express. See our code at: https://github.com/Mercury7353/EvoChamber

cs.AI↗

arXiv2026

Beyond Static Bias: Adaptive Multi-Fidelity Bandits with Improving Proxies

As an extension of the classical multi-armed bandit problem, multi-fidelity multi-armed bandits (MF-MAB) enable individual arms to be evaluated using diverse feedback sources that vary in both cost and accuracy. Prior stochastic models typically assume fixed low-to-high fidelity discrepancies, whereas modern proxy sources, such as learning-based simulators and Large Language Models (LLMs), can be improved using additional calibration. We investigate adaptive MF-MAB with improving proxy sources, and focus on the canonical two-fidelity case in which the low-fidelity source becomes more informative with repeated use. To capture this dynamic, we introduce a selected-average mismatch bound that converts dynamic low-fidelity observations into improvement-aware confidence bounds for the high-fidelity target. We propose the Threshold-Based Adaptive Continuation Companion (TACC), an optimistic algorithm that uses a bounded continuation rule to decide when low-fidelity sampling remains cost-effective and when to escalate. We prove an instance-dependent regret bound showing that, for detected intermediate arms, adaptive continuation replaces logarithmic high-fidelity confirmation with bounded low-fidelity continuation. Experiments on synthetic bandits and an LLM-as-a-judge policy-evaluation task examine when continuation improves cost-weighted regret.

cs.LG↗