Trace2Skill: Verifier-Guided Skill Evolution for Long-Context EDA Agents

2026-05-22 · Source: cs.MA updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Software Development & Engineering · Depth: Expert, extended

Summary

Trace2Skill is a test-time scaling framework designed to improve hardware LLM agents tackling Complex Verilog Design Problems (CVDP) without requiring model fine-tuning. It evolves an agent's natural-language skill by mining repeated rollout traces for success and failure modes, converting them into diagnostics and oracle lessons. An oracle–mutator–selector loop then produces task-specific skills that guide subsequent search, editing, validation, and recovery. The framework also supports bounded runtime dense verifier feedback, providing sanitized functional observations to guide skill evolution and agent execution. On eight hard CVDP tasks that defeated the seed CVDP agent and frontier coding agents, Trace2Skill with dense verifier feedback achieved a 33.6% pass rate, solving 6/8 tasks, a significant improvement over the 0/8 baseline.

Key takeaway

For AI Scientists and Machine Learning Engineers developing agents for complex hardware design, Trace2Skill offers a compelling strategy to overcome hard failures. You should consider implementing a skill evolution framework with dense verifier feedback to adapt agents to specific tasks without costly model fine-tuning. This approach improves task pass rates and enables breakthroughs on previously unsolved problems by fostering skill-agent coadaptation, making failures actionable and guiding future agent behavior.

Key insights

Evolve natural-language agent skills with verifier feedback to solve complex hardware design problems without model fine-tuning.

Principles

• Treat agent's natural-language skill as an evolvable policy.
• Mine execution traces for both success and failure modes.
• Dense verifier feedback improves skill evolution and agent execution.

Method

Trace2Skill uses an oracle (GPT-5) to summarize lessons from traces, a mutator (Claude Sonnet 4.5) to propose child skills, and a selector to choose the next task-specific skill, guided by metrics like SelectQ.

In practice

• Use Claude Opus 4.6 as the tool agent for RTL inspection and editing.
• Employ GPT-5 for oracle summarization of actionable lessons.
• Apply dense verifier feedback for in-rollout corrections.

Topics

• LLM Agents
• Hardware Design Automation
• Verilog Design Problems
• Skill Evolution
• Verifier Feedback
• Test-Time Adaptation

Code references

• NVIDIA-NeMo/Gym

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills — Trace2Skill distills broad agent experience into transferable, declarative skills via parallel analysis and inductive consolidation.
• EvoTrainer: Co-Evolving LLM Policies and Training Harnesses for Autonomous Agentic Reinforcement Learning — EvoTrainer co-evolves LLM policies and training harnesses, moving beyond static recipe search for agentic RL.
• SkillAdaptor: Self-Adapting Skills for LLM Agents from Trajectories — SkillAdaptor enables stable, auditable, training-free skill adaptation for LLM agents via step-level failure attribution.
• Think Twice, Act Once: Verifier-Guided Action Selection For Embodied Agents — Explicit test-time verification with a specialized, generatively trained verifier significantly boosts embodied agent robustness.
• Microsoft’s open-source SkillOpt automatically upgrades AI agent skills without touching model weights — SkillOpt applies deep-learning optimization to text-based agent skills, enabling autonomous, verified performance improvements without altering model weights.
• SkillHone: A Harness for Continual Agent Skill Evolution Through Persistent Decision History — SkillHone enables continuous agent skill evolution by persistently recording and utilizing decision history and evaluation feedback.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.MA updates on arXiv.org.