MaD-RSI: Model-as-Data for Recursive Self-Improvement of Agent Harnesses

Key points

• However, such optimization often creates a strong coupling between the learned harness and its target LLM, limiting generalization across models and requiring substantial computational resources to adapt to new LLMs. Moreover, existing RSI methods typically optimize a harness using feedback from a single model and evaluate its generalization to other models only after training.1
• In this paper, we introduce a new perspective that treats models as data and places the harness at the center of optimization.1
• Building on this perspective, we propose MaD-RSI, a framework that directly optimizes a shared harness across multiple heterogeneous LLMs by incorporating their diverse failure trajectories into the training process.1
• Furthermore, we introduce a new evaluation framework with two complementary metrics designed to assess the effectiveness of harness optimization across models.1

Sources (1)

• [1]MaD-RSI: Model-as-Data for Recursive Self-Improvement of Agent Harnesses
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 01:41 PM

  Recursive self-improvement (RSI) of agent harnesses has emerged as an effective approach to enhancing the capabilities of large language model (LLM) agents.

  However, such optimization often creates a strong coupling between the learned harness and its target LLM, limiting generalization across models and requiring substantial computational resources to adapt to new LLMs. Moreover, existing RSI methods typically optimize a harness using feedback from a single model and evaluate its generalization to other models only after training.

Extractive summary: sentences quoted from the sources.