---
id: "lib-2026-10-20-qd-rec-quality-diversity-search-for-decision"
type: "library"
title: "QD-Rec: quality-diversity search for decision-support recommendations"
title_original: "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」"
lang: "JA"
org: "Sakana AI"
source_type: "Other"
date: "2026-10-20"
week: "2026-W43"
url: "https://nosov.ai/library/lib-2026-10-20-qd-rec-quality-diversity-search-for-decision/"
original_url: "https://sakana.ai/qd-rec/"
tier: "A"
tier_name: "Strong evidence"
tier_reason: "Sakana AI research post backed by an arXiv paper and public code, with a clear method tested on three domains against two baselines; results are scored by an LLM judge and the post gives no aggregate figures."
summary_basis: "full-text"
original_minutes: "10"
tags: ["decision support", "quality diversity search", "recommendations", "llm methods"]
---

# QD-Rec: quality-diversity search for decision-support recommendations

Original title: 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」

Tier A (Strong evidence): Sakana AI research post backed by an arXiv paper and public code, with a clear method tested on three domains against two baselines; results are scored by an LLM judge and the post gives no aggregate figures.

Summary based on the full text.

## In short

Sakana AI's Japanese blog post (20 October 2026) introduces QD-Rec, a recommendation method that applies quality-diversity search from open-ended evolution research to decision support. The unit of recommendation is a pair: a candidate (a company, customer or paper) plus a written hypothesis for why to pick it. An LLM writes new hypotheses by mutating or crossing earlier ones kept in an archive, and new entries only compete with their nearest neighbour. On M&A, customer targeting and paper recommendation, QD-Rec built archives that were larger and had higher average scores than baselines in most settings. Paper: arXiv 2610.23470; code on GitHub.

## Key numbers

- ~6,300: US companies (August 2024 data) searched for acquisition candidates in the M&A task
- ~17,500: papers from recent AI conferences searched in the paper-recommendation task
- 1,000: generate-and-evaluate rounds after which QD-Rec and baseline archives were compared
- 1 to 5: judge score (out of 10) rising when a failed airline-acquisition idea was transplanted to Liberty TripAdvisor for JPMorgan

## What it reports

- Three steps repeat: a bandit picks the next candidate, balancing promising and unexplored ones; a proposer LLM writes a child hypothesis from archived parents; a judge (LLM or human) scores it.
- A hypothesis with a new angle enters the archive unconditionally; otherwise it faces a one-to-one contest against the single most similar archived hypothesis and replaces it if it wins.
- Test domains: about 6,300 US companies with ~123-word business descriptions and seven financial fields; 10,000 synthetic customer profiles (~685 words, California); about 17,500 conference papers (titles and abstracts).
- Baselines: Novelty (writes each hypothesis from scratch) and Novelty+Evo (evolves hypotheses but picks candidates at random). After 1,000 rounds QD-Rec archives had higher average quality and larger size in many settings; quality and size kept rising with more rounds.
- Worked example: a from-scratch case for JPMorgan buying Liberty TripAdvisor scored 1/10; an airline-acquisition parent also scored 1/10; transplanting the parent's "embed lending, payments and rewards in the travel purchase flow" idea gave a child scoring 5/10.
- Second example: a nurse rejected as a customer for an online-learning service (2/10) became a 7/10 recommendation after borrowing the logic of a kindergarten-teacher hypothesis (7/10).
- Proposer was claude-sonnet-5 and judge gpt-5.6-terra in the examples.

## How they got it

Fixed-budget experiments on three public or synthetic datasets, comparing archive quality (judge scores) and archive size across methods, with five seeds and standard errors in the charts. Archive diversity is visualised by projecting entry embeddings to 2D with PCA on a 25x25 grid. The proposer was a single LLM call on short inputs.

## Limits

- Quality is scored by an LLM judge; the post itself names the risk of the proposer gaming the judge without improving substance.
- The blog shows charts and examples but no aggregate effect sizes; checking them needs the paper.
- The JPMorgan example contains a fact the model invented (a member base "over 100 million"); Sakana flags it and notes the target was acquired in April 2025, after the data date.
- Inputs were deliberately thin (short descriptions, a few financial fields), so real-world usefulness is untested.

## Why it matters for companies

Strategy, sales and research teams often need a shortlist with distinct rationales rather than a single top-ranked answer. QD-Rec offers a pattern for that: keep a structured archive of reasons, let weak ideas serve as stepping stones, and cap cost through bandit selection. The judge is the control point. Sakana suggests putting internal experts in that role or splitting it between people and models, which is where governance effort belongs before using this for M&A or targeting.

## Key passages in translation

> QD-Recは、企業や顧客、論文などの「候補」と、それを選ぶ理由を述べた「仮説」の組を推薦の単位とし、過去の優れた仮説をもとに生成AIが新しい仮説を書くサイクルを繰り返します。

QD-Rec takes as its unit of recommendation a pair of a "candidate", such as a company, customer or paper, and a "hypothesis" stating why to choose it, and repeats a cycle in which generative AI writes new hypotheses based on strong past ones.

> LLMの採点には、内容の質を高めずに採点者の弱点を突いて高得点を得るハックのリスクもあります。

LLM scoring also carries the risk of hacks that win high scores by exploiting the judge's weaknesses without improving the quality of the content.

> 本記事で示す買収案は、推薦手法を評価するためにモデルが生成した仮説です。

The acquisition proposals shown in this article are hypotheses generated by the model to evaluate the recommendation method.

Original: https://sakana.ai/qd-rec/