The paper presents MaD-RSI, a method for recursive self-improvement (RSI) of agent harnesses—the outer logic that directs how a large language model (LLM) acts on a task. The abstract notes that such RSI has become an effective approach for enhancing LLM agent capabilities.
As the title suggests, the method adopts a 'model-as-data' perspective. According to the available abstract, MaD-RSI works by searching the harness space for a given LLM and dataset. However, the abstract is cut off, so the exact mechanics of how the model is treated as data remain unspecified in the source.
Because this is a single preprint announcement, the claims have not been peer-reviewed. The brief, truncated abstract limits what can be independently assessed, but the work appears to build on existing RSI ideas by making the model itself part of the searchable configuration space.