WEFT: Whole-System Evolution for Tool-Use Post-Training Scales Agentic Interaction with MegaMCP State Isolation

核心背景与行业痛点

Scaling tool-use capabilities through post-training is foundational for modern autonomous coding and enterprise agents. However, recent paradigms primarily emphasize synthesizing executable sandbox environments (such as Agent-World). In practice, an environment is only one constituent of an agentic interaction system that equally depends on nuanced task definitions, execution harnesses, and robust evaluators. Scaling environments in isolation produces severe cross-component incoherence, generating noisy reward signals that destabilize policy optimization and trigger overfitting to synthetic quirks.

架构亮点与底层机制

Researchers from Fudan University present WEFT (Whole-system Evolution For Tool-use Post-training, arXiv:2609.36887):

1. Whole-System Co-Construction: Jointly scales the entire agentic topology across environmental diversity (spanning diverse enterprise APIs), task complexity, and multi-turn interaction modalities.
2. Execution-Driven Self-Evolution: Inspects full execution traces and runtime state invariants to pinpoint root causes behind failures across environments, harnesses, or rubrics, applying targeted updates verified by follow-up rollouts.
3. Stable Post-Training Optimizations: Introduces prefix-preserving sampling to safeguard confirmed progress during RL exploration, coupled with atomic-turn credit assignment to concentrate policy gradients on critical tool-invocation decisions.
4. MegaMCP Infrastructure: Integrates with the Model Context Protocol (MCP) standard to orchestrate isolated, snapshot-recoverable state management across massive concurrent agent rollouts over shared backend services.

权威 Benchmark 与实测跑分对比

Empirically evaluated against state-of-the-art environment-scaling baselines:

1. Dominates Across BFCL V4, tau^2-Bench, and Claw-Eval: Both WEFT-8B and WEFT-14B establish new performance highs against identically sized foundation models.
2. Up to +12.27 Percentage Point Gains Over Agent-World-14B: WEFT-14B outperforms Agent-World-14B by 6.41 percentage points on BFCL V4, 2.23 points on tau^2-Bench, and an extraordinary 12.27 percentage points on Claw-Eval.
3. Robust Long-Horizon Scaling: The scaled WEFT-35B-A3B variant maintains superior accuracy on arduous long-horizon task benchmarks, including Toolathlon-Verified and AutomationBench.

开发者实战落地与开箱指南

WEFT provides a definitive blueprint for machine learning engineers building tool-using agents. By replacing isolated environment generation with whole-system evolution and deploying MegaMCP for resilient state management, teams can construct stable, sample-efficient reinforcement learning pipelines that transfer reliably to production APIs and complex multi-agent workflows.