The system uses a Quality-Diversity archive: an LLM generates new hypotheses by mutating or combining prior ones, while an AI or human evaluator scores them. Novel directions can occupy new niches; otherwise, a proposal competes with its most similar archived counterpart.
The source reports tests in three settings: M&A candidate discovery, customer targeting, and research-paper recommendation. After 1,000 generation-and-evaluation cycles, QD-Rec built larger archives with higher average evaluations in many settings than the described baselines. Sakana AI frames the approach as supporting comparison across distinct recommendations, not merely selecting a single top-scoring item.
These are reported experimental results, not evidence of deployment outcomes. The article also notes that generated recommendations can contain unsupported claims, so the outputs should not be treated as decision-ready without validation. Source: https://sakana.ai/qd-rec