HarnessRL: Co-Evolving Policies with a Just-in-Time Training Harness for Open-Ended RL

Key points

• Both reflect a training configuration whose task sampling, exploration, and evaluation stay fixed as the policy evolves.1
• We introduce HarnessRL, a training-time harness that co-evolves with the policy through three interfaces for task sampling, rollout guidance, and rubric criteria.1
• HarnessRL maintains a training-time memory from which a harness evolver retrieves within-task history and cross-task evidence.1
• Using this evidence, the evolver attributes each reward failure to the interface whose update can address it.1

Sources (2)

• [1]HarnessRL: Co-Evolving Policies with a Just-in-Time Training Harness for Open-Ended RL
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 01:53 PM

  Rubric-based reinforcement learning (RL) trains policies for open-ended generation with natural-language criteria as rewards, but these rewards can become ineffective as training progresses.

  Both reflect a training configuration whose task sampling, exploration, and evaluation stay fixed as the policy evolves.

• [2]HarnessRL: Co-Evolving Policies with a Just-in-Time Training Harness for Open-Ended RL
  Hugging Face Daily Papers · Oct 19, 12:00 AM · same content

Extractive summary: sentences quoted from the sources.