オープンエンドな探索を意思決定支援へ：推薦の質と多様性を高める「QD-Rec」を提案

Sakana AI's QD-Rec treats recommendation as building a varied set of proposals, each pairing a candidate (a company to acquire, a customer to target, a paper to read) with a written hypothesis explaining why it is worth choosing. It borrows from quality-diversity search, which looks for many good solutions that differ from one another rather than a single best one. In each cycle a contextual bandit picks the next candidate, balancing promising ones against unexplored ones; a proposer LLM writes a hypothesis for it by adapting an earlier hypothesis from the archive or combining several; and an evaluator scores it. A new entry with a novel angle gets its own niche, while one resembling an existing entry only has to beat its nearest neighbour, so different rationales coexist instead of all competing for first place. The method is described in an arXiv paper and a technical article.

Reusing earlier hypotheses is the point. In one example, a rejected idea that JPMorgan buy Southwest Airlines, scored 1 out of 10, contributed the idea of building lending, payments and rewards into travel purchases; moved to Liberty TripAdvisor, it scored 5, where a hypothesis written from scratch for the same company scored 1. Across M&A over about 6,300 US companies, customer targeting over 10,000 fictional profiles, and paper recommendation over about 17,500 conference papers, with claude-sonnet-5 and claude-haiku-4-5 as proposers and gpt-5.6-terra as evaluator, the authors report that QD-Rec achieved the highest mean quality in most settings and usually built the largest archive within 1,000 evaluations, compared with baselines that write every hypothesis from scratch or pick candidates at random.

Those quality numbers are LLM-judge scores, and the same scores that steer the search are used to report quality, which leaves room for proposals that exploit the judge rather than improve; the authors name this reward-hacking risk and suggest human experts take on or share the grading in real use. The examples also contain claims the model added beyond its input, and the company data date from August 2024, so the acquisition cases are evaluation outputs, not advice.