WEAVE: Weaving Real-World Code Evolution into Dependency-Structured Tasks for Benchmarking and Training Code Agents
We introduce WEAVE, a framework that transforms verified PR histories into executable, dependency-structured tasks.
ProofPaper ↗
Key points
- Real-world software engineering rarely consists of isolated changes.
- This mismatch obscures dependencies among changes and limits progress measurement and learning signals for long-horizon code agents.
- We further introduce Graph Credit, a deterministic, test-grounded metric that measures dependency-consistent partial progress and supports dense, verifiable rewards for reinforcement learning.
- Experiments on task construction, agent evaluation, supervised fine-tuning, and reinforcement learning show that WEAVE yields coherent executable tasks, enables more faithful progress measurement than flat scoring, and provides effective supervision and verifiable rewards for training code agents.
Sources (1)
- [1]WEAVE: Weaving Real-World Code Evolution into Dependency-Structured Tasks for Benchmarking and Training Code AgentsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 03:12 PM
We introduce WEAVE, a framework that transforms verified PR histories into executable, dependency-structured tasks.
Real-world software engineering rarely consists of isolated changes.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 15, 2026Selection-Based Structured Reasoning: Toward Efficient Multimodal Search Agents
- Oct 14, 2026Explorer-Zero: Improving Self-Play for Long-Horizon Tasks with Decomposable Verifiable Rewards
- Oct 14, 2026Learning Perception to Action Rules through Reinforcement Learning for Theory-of-Mind Reasoning
- Oct 14, 2026Player-1: Learning to Play from Synthetic Experience
- Oct 14, 2026The Design Space of Tri-Modal Masked Diffusion Models
- Oct 14, 2026MAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments