ARTFEED — Contemporary Art Intelligence

DemoEvolve: Demonstration-Bootstrapped Harness Evolution for Language Agents

ai-technology · 2026-05-26

A new AI research paper introduces DemoEvolve, a method for improving frozen language-model agents by evolving their external harness structures using human demonstrations. The approach addresses the problem of sparse rewards in long-horizon stochastic environments, where self-generated rollouts become fragile. DemoEvolve uses competent human trajectories as expert reference experience to guide harness-level diagnosis, overcoming the limitations of reward-only search.

Key facts

  • DemoEvolve is a demonstration-bootstrapped approach to harness evolution.
  • It improves frozen language-model agents by modifying executable structures around them.
  • Prior work shows self-generated rollouts can support harness search.
  • In long-horizon stochastic environments, self-practice becomes fragile due to sparse rewards and high variance.
  • Human trajectories serve as expert reference experience for the coding proposer.
  • The paper is published on arXiv with ID 2605.24539.
  • The method is described as sample-efficient fast adaptation.
  • Base model's general capabilities remain intact while task-specific competence is acquired.

Entities

Institutions

  • arXiv

Sources