Trace2Skill: Verifier-Guided Skill Evolution for Long-Context EDA Agents
Trace2Skill is a framework aimed at scaling tests to improve hardware LLM agents in addressing Complex Verilog Design Problems (CVDP). These challenges involve identifying verifier-relevant RTL, testbenches, include paths, and build dependencies within extensive repository snapshots, executing precise modifications, and recovering from infrequent hidden-verifier failures. Instead of refining an RTL-focused model or simply generating additional candidate solutions, Trace2Skill views the agent's natural-language capabilities as an adaptable policy. It analyzes repeated rollout traces to identify success and failure patterns, transforming them into detailed diagnostics and oracle insights. An oracle, mutator, and selector loop is employed to develop task-specific skills that assist in subsequent search, editing, validation, and recovery. Additionally, due to the often vague nature of final pass/fail labels regarding critical failures, Trace2Skill offers support for bounded runtime den.
Key facts
- Trace2Skill is a test-time scaling framework for hardware LLM agents.
- It addresses Complex Verilog Design Problems (CVDP).
- No RTL-specialized model fine-tuning is required.
- The framework treats natural-language skill as an evolvable policy.
- It mines rollout traces for success and failure modes.
- Uses an oracle, mutator, and selector loop to generate task-specific skills.
- Supports bounded runtime den for handling hard failures.
- Published on arXiv under ID 2605.21810.
Entities
Institutions
- arXiv