Skip to main content
Back

Sakana AI's QD-Rec Builds Competing Rationales to Outsmart Single-Score Recommendations

Sakana AI's QD-Rec evolves a diverse archive of item plus reason pairs instead of a single ranked list, aimed at high-stakes decisions.

  • Sakana AI released QD-Rec, a quality-diversity method for open-ended recommendation over a fixed candidate pool.
  • Each archive entry pairs an item with a natural-language hypothesis explaining why it fits the decision maker.
  • A contextual bandit picks items, an LLM evolves hypotheses from archive stepping stones, an LLM judge admits novel or better pairs.
  • Tested on about 6,300 companies, 10,000 synthetic customers, 17,500 papers, plus a Goodreads reader setup.
  • Beats from-scratch and no-bandit ablations on mean quality and archive size in most scenarios after 1,000 rounds.
  • Code is Apache 2.0 on GitHub; results are LLM-judged, not real-world decision outcomes.

Sakana AI has introduced QD-Rec, a recommendation method that uses quality-diversity search to build an archive of candidate items paired with natural-language rationales. The project page describes a system for fixed candidate pools and decision-maker profiles, such as a list of acquisition targets and a buyer’s strategic goals. Its output lets an analyst inspect several strong, distinct arguments across the pool.

The research paper, QD-Rec: Open-Ended Evolution of Recommendations via Quality-Diversity, calls this task open-ended recommendation. Each unit of search is an item-hypothesis pair: a candidate plus an explanation connecting it to the profile’s goals, interests, and background. Quality-diversity algorithms retain strong solutions across different behavioral niches, allowing QD-Rec to preserve competing rationales while its language-model proposer mutates and recombines them over repeated rounds.

Why one score loses useful context

A single acquisition target may appeal to an M&A team because of its technology, geographic reach, customer base, or engineering talent. Compressing those arguments into one score removes information about how each candidate could serve the buyer’s strategy. QD-Rec keeps the arguments attached to their candidates, giving decision makers a structured set of rationales to compare.

Three components drive each search round

Each QD-Rec search round updates the archive through item selection, hypothesis generation, and local evaluation:

  1. Select an item. A contextual bandit uses item embeddings and previous evaluation scores to balance promising candidates with less certain ones. A diversity bonus favors items unlike those already represented in the archive.