Recursive self‑improvement (RSI) has emerged as an effective way to boost the capabilities of large language model (LLM) agents. By searching the harness space for a specific LLM on a given dataset, RSI can discover a harness that aligns with the model’s abilities and behavioral traits. However, this optimization often creates a strong coupling between the harness and its target LLM, limiting cross‑model generalization and requiring substantial compute to adapt to new LLMs.
Existing RSI methods typically optimize a harness using feedback from a single model and only later evaluate its generalization on other models, treating model differences merely as an evaluation criterion rather than as optimization evidence.
We introduce a new perspective that treats models as data and places the harness at the center of optimization. Building on this view, we propose the MaD‑RSI framework, which directly optimizes a shared harness across multiple heterogeneous LLMs by incorporating each model’s failure trajectories into the training process. Instead of assessing model differences after optimization, MaD‑RSI leverages these differences as training evidence to guide the harness’s self‑improvement. We also present a new evaluation suite with two complementary metrics to measure harness effectiveness across models. Experiments on several datasets and LLMs show that MaD‑RSI improves generalization while significantly reducing training costs.
Review