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

Key points

• 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.1
• 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.1
• A verifier that never changes also grows stale as the agent's responses improve and task demands shift.1
• 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.1

Sources (1)

• [1]Verifier-Harness Co-Evolution for Training-Free Self-Improvement on Open-Ended Tasks
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 02:29 AM

  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.

  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.

Extractive summary: sentences quoted from the sources.