Sakana AI releases QD-Rec to evolve recommendations and their reasoning

Sakana AI has released QD-Rec, a recommendation method that develops reasons for choosing an option, then reuses those ideas to find others. Published with research and public code, it aims to give people varied proposals to compare—not just a ranked list of companies, customers or papers.

How recommendations become parents

The basic unit is a candidate paired with a hypothesis: a written explanation of why that candidate might be useful. For an acquisition, that means a business scenario, not simply a company name and score. One company can support several different proposals.

Rather than search only for a single winner, QD-Rec keeps recommendations built on different lines of reasoning. This approach is called Quality-Diversity search: preserve useful differences while improving the options within each category. An unusual proposal can survive even when its score is low, providing material for a later idea.

Each cycle starts by choosing a candidate. A selection technique called a bandit balances options with promising past scores against those barely explored. A language model then draws on saved hypotheses as parents. It adapts one parent to the new candidate, called mutation, or combines ideas from multiple parents, called crossover.

A scorer evaluates the result. A proposal with a new angle gets its own place in the archive, the system’s saved collection. Otherwise, it competes only against the most similar existing proposal and replaces it if it scores higher. That local contest lets several strategies coexist instead of forcing every recommendation into one overall ranking.

Larger archives across three tasks

Sakana tested the method on acquisition research using about 6,300 US companies, customer targeting using 10,000 fictional profiles, and recommendations drawn from about 17,500 AI-conference papers. The same algorithm searched for acquisition rationales, reasons a product might suit a customer, or connections between papers and a researcher’s interests.

After 1,000 generation-and-evaluation rounds, Sakana says QD-Rec produced larger archives with higher average evaluation scores in many settings than Novelty, a baseline that writes each hypothesis from scratch. The experiments used language models as scorers. Another comparison, Novelty+Evo, reused hypotheses but selected candidates randomly, testing the contribution of the bandit selection step.

A customer-targeting example shows what borrowing an idea can change. A fresh proposal dismissed an online learning service as a poor fit for a fictional nurse. QD-Rec adapted a rationale written for a teacher moving into leadership: professional training may leave gaps in project management, data analysis and presentation skills.

Better reasoning is not verified evidence

The acquisition tests expose a practical limit. Companies were represented by short business descriptions and seven financial fields, using August 2024 data. Sakana notes that Liberty TripAdvisor Holdings, a target in one generated example, was acquired by Tripadvisor in April 2025. Real decisions would require more detailed, current information.

Some generated acquisition arguments also included claims absent from the inputs. One asserted an airline loyalty program had more than 100 million members, a figure Sakana could not confirm. The company explicitly labels the acquisition proposals experimental hypotheses, not actual corporate plans or investment advice.

Sakana also warns that proposals could exploit an AI scorer’s weaknesses without improving their substance. It suggests human scoring, shared human-and-model evaluation, and checks against expert judgments. Richer source data or a research agent could strengthen proposal generation. The company has published code and implementation examples for multiple recommendation domains.