QD-Rec: Open-Ended Evolution of Recommendations via Quality-Diversity

• A fixed candidate pool still leaves an open-ended search over the many possible reasons for choosing each item.The system returns diverse, high-quality item–hypothesis pairs rather than a single ranked list.Source: §1 · p. 1
• QD-Rec selects items with a contextual bandit, generates hypotheses from archive exemplars, and retains novel pairs or non-novel pairs that beat a nearby recommendation.The archive exemplars serve as stepping stones for generating new hypotheses.Source: §1 · p. 2, Figure 1 · p. 2
• The paper formalizes open-ended recommendation over fixed-pool item–hypothesis pairs and evaluates the method across four domains and two proposers.Source: §1 · p. 2
• QD-Rec gives the highest mean quality in most LLM-judged settings, usually builds the largest archive, and outperforms diversity-aware rerankers on human-rated quality.Source: §1 · p. 3

• Quality–diversity methods maintain diverse high-performing solutions in an archive, while novelty search rewards behavioral novelty as a stepping stone to later progress.MAP-Elites retains elites across behavioral niches.Source: §2 · p. 3
• Foundation-model evolutionary loops condition proposals on archived artifacts, but unlike systems that create new artifacts, QD-Rec selects items from a fixed pool and generates hypotheses for them.QD-Rec treats item–hypothesis pairs as the artifacts in the loop.Source: §2 · p. 3
• Prior recommender systems use bandits for sequential recommendation and diversify displayed lists, whereas QD-Rec uses a contextual bandit to select existing items and enforces diversity at archive admission.The final archive, rather than a displayed list, is the object evaluated in QD-Rec.Source: §2 · pp. 3–4

• Given a decision-maker profile and fixed candidate pool, open-ended recommendation returns a high-quality, diverse set of item–hypothesis pairs using at most T costly evaluations.A hypothesis is a natural-language explanation of why an item fits the profile; the same item may have different hypotheses.Source: §3 · p. 4

4. QD-Rec

• QD-Rec stores each retained recommendation with its embedding and quality score, and compares new entries with their nearest archive neighbors.The entry embedding is computed from a structured item profile followed by the hypothesis text.Source: §4 · p. 4, §4.1 · p. 4

4.1. Local novelty, ranking, and archive update

• Novel entries enter the archive immediately; otherwise, a blind duel against the single nearest neighbor determines whether the new entry replaces it.The judge receives no pointwise scores and is not told which candidate is new.Source: §4.1 · p. 4

4.2. Item selection and hypothesis generation

• A diversity-aware linear Thompson-sampling bandit balances sampled item quality against distance from the nearest item already covered by the archive.The novelty bonus affects item selection, not the normalized reward used to update the posterior; the reported setting uses β = 1.Source: §4.2 · p. 5

4.3. Implementation details

• The proposer mutates, crosses over, or explores angles using archive hypotheses as exemplars, while batched runs filter near-duplicates and resolve replacement collisions.Batched selection uses independent posterior draws and excludes items already selected in the batch.Source: §4.2 · p. 5, §4.3 · p. 6

5.1. Setup

• The three variants share archive updates; Novelty+Evo adds evolutionary few-shot generation, while QD-Rec also adds diversity-aware bandit item selection.Novelty removes both the bandit and archive exemplars; Novelty+Evo uses uniform-random item selection.Source: §5.1 · p. 7, Table 1 · p. 7
• The evaluation spans Goodreads and three applied domains, using real reader ratings for Goodreads Quality and LLM judging in the other settings.The domains are M&A target discovery, marketing, and paper-abstract exploration; runs use five seeds and matched evaluation budgets.Source: §5 · p. 6, §5.1 · p. 7

5.2. Results

• QD-Rec achieves the best mean Quality in most LLM-judged settings and usually builds the largest archive, with best-or-tied Diversity except in M&A.It is often significantly better than both ablations and never significantly worse than the best variant.Source: Table 2 · p. 8, §5.2 · p. 8
• On Goodreads, the variants achieve similar human star ratings, while QD-Rec matches or beats both ablations on archive size, Diversity, and coverage.Neither training-free reranker reaches any QD variant on Quality; they approach QD-Rec’s Diversity only at their most diversity-favoring settings.Source: §5.2 · pp. 8–9
• QD-Rec costs the same as Novelty+Evo and about 1.3× as much as Novelty; more than 90% of the difference comes from proposer input exemplars.A run costs about $13–$62 across the reported configurations.Source: §5.2 · p. 9
• QD-Rec covers more map space than Novelty, while its Quality rises through the run and ends highest as the archive grows.The map comparison is most visible for Disney and the 3D printer.Source: §5.2 · p. 9, Figure 2 · p. 10
• With Qwen embeddings, Diversity is higher for every variant, while Quality and archive-size orderings remain unchanged across the four profiles.The sensitivity test uses one run per profile, so Diversity and Coverage are comparable only within an embedding space.Source: §5.2 · pp. 9–10, Table 3 · p. 9

6. Conclusion

• QD-Rec uses contextual bandits to choose items while searching open-endedly over the reasons for recommending them.Across four domains, it gave the best mean Quality in most settings and usually built the largest archive.Source: §6 · p. 10
• The method uses a linear reward model over fixed item embeddings, uniform archive exemplars, and a single prompted LLM call for hypothesis generation.The authors identify nonlinear bandits, alternative exemplar selection, and a full research agent as possible extensions.Source: §6 · p. 11

A. Additional results: maps and curves over the run for all scenarios

• Figures 3 and 5 show maps and curves, including semantic Diversity, for all eight scenarios; Figures 4 and 6 give the weaker-proposer counterparts.Source: A. Additional results: maps and curves… · p. 14

A.1. Qualitative maps: all scenarios

• Figure 3 shows MAP-Elites-style heatmaps for all eight scenarios, while Figure 4 repeats them with the weaker proposer.Source: §A.1 · p. 14

A.2. Metrics over the run: all scenarios

• Figures 5 and 6 track mean Quality, archive size n, and semantic Diversity against evaluations across all eight scenarios.Figure 6 uses the weaker proposer claude-haiku-4-5.Source: §A.2 · p. 14
• After about 100 evaluations, Diversity changes little even as archive size n keeps growing.The passage suggests novelty checks against nearest neighbors may admit distinct hypotheses within regions already spanned, increasing archive size without raising mean pairwise distance.Source: §A.2 · p. 14

B. No-learning baselines: relevance-MMR and a quality-weighted DPP

• Relevance-MMR trades item relevance against dissimilarity from already selected items using a weight γ.At γ=1, it recovers plain top-K retrieval.Source: B. No-learning baselines:… · pp. 14–15
• Quality-weighted DPP greedily maximizes weighted relevance plus the log-volume diversity of the selected-item similarity matrix.Its weight γ also trades relevance against diversity, and γ=1 recovers plain top-K retrieval.Source: B. No-learning baselines:… · pp. 14–15
• Neither baseline uses an archive, novelty filter, feedback, or LLM-influenced selection; both select geometrically from precomputed content embeddings.They use relevance cosine similarity and human star ratings for Quality; zero-shot memos are added only after selection for shared Diversity and Coverage measurement.Source: B. No-learning baselines:… · p. 15
• The comparison is not a controlled ablation: QD variants receive ratings for evaluated items, while the baselines see none.Each baseline run is matched to its reader’s QD-Rec archive size, so the gap combines rating feedback and archive machinery.Source: B. No-learning baselines:… · p. 16
• No setting of Relevance-MMR or quality-weighted DPP reaches any QD variant on Goodreads Quality.The baselines approach QD-Rec’s Diversity only at their most diversity-favoring settings, while a few settings exceed its Coverage.Source: B. No-learning baselines:… · p. 16
• DPP gives almost the same results as MMR, so the gap is not specific to MMR’s heuristic.Figures 3 and 4 show 25×25 best-quality-per-cell maps for all three variants; their color ranges are auto-scaled within domains or proposer runs.Source: Figure 3 · p. 15, Figure 4 · p. 16, B. No-learning baselines:… · p. 17

C. Dataset details

• Goodreads uses Children-subset books rated by two readers selected through fixed eligibility thresholds, with ratings supplying the bandit reward and replacement score.The chosen readers have 1,761 and 1,117 books; books are selected without replacement, and ties preserve the incumbent.Source: C. Dataset details · p. 19
• M&A target discovery uses 6,332 USA companies from LSEG Workspace and Walt Disney Co and JPMorgan Chase & Co as acquirer profiles.Candidate embeddings use business descriptions for bandit selection; proposer, judge, and archive embeddings use profiles that also include location, industry, and financial features.Source: C. Dataset details · pp. 19–20
• Marketing uses 10,000 California profiles sampled from the synthetic NVIDIA Nemotron-Personas-USA dataset as a proxy for sensitive real-world customer data.The two product profiles are a desktop 3D printer and an online upskilling subscription.Source: C. Dataset details · p. 20
• Paper-abstract exploration uses 17,511 accepted papers from ICLR 2026, ICML 2026, and NeurIPS 2025, represented by title and abstract metadata.Candidate embeddings support bandit selection, while title, abstract, and memo embeddings support archive-neighbor retrieval, Diversity, and Coverage.Source: C. Dataset details · p. 20

D. Prompt Templates

• The appendix says Goodreads uses only generation and novelty-check prompts, relying on human ratings for quality and replacement; unparseable novelty or duel responses never add archive entries.Prompt placeholders are filled at runtime.Source: D. Prompt Templates · p. 20

D.2. M&A target discovery

• M&A generation prompts ask for concrete acquisition rationales grounded in strategic fit, synergies, market position, financial attractiveness, and timing, using mutate, crossover, or explore operations.They also require reporting the operation and source theses separately from a standalone hypothesis.Source: §D.2 · pp. 23–24
• M&A judges score acquisition hypotheses from 0–10 for strategic fit, synergies, feasibility, risks, and value creation, reserving high scores for compelling, specific, plausible cases.The scoring guide maps 0–2 to no real case, 7–8 to a strong case, and 9–10 to an excellent case.Source: §D.2 · pp. 24–25
• M&A novelty checks require a genuinely distinct strategic direction, while blind pairwise duels must choose one thesis using fit, synergies, feasibility, and value creation.Minor variations or superficially different targets with the same thesis do not count as novel.Source: §D.2 · pp. 25–26

D.3. US customer targeting (Marketing)

• Marketing generation prompts build persona-specific product-interest hypotheses around needs, lifestyle, benefits, and resonant features, using mutate, crossover, or explore operations.The prompt requests an evolution note for inspiration drawn from examples while keeping the hypothesis standalone.Source: §D.3 · p. 27
• Marketing judges score targeting recommendations from 0 to 10 on relevance, specificity, plausibility, insight, and value alignment, with weak or implausible cases scoring low.The scoring guide distinguishes moderate cases with limited specificity or depth from excellent, compelling, specific connections.Source: §D.3 · p. 28
• Marketing novelty assessment focuses on a new hypothesis’s motivation or product use case, not different users, and rejects trivial variations.The prompt requests a short novelty rationale and a Boolean novelty judgment.Source: §D.3 · pp. 28–29
• Marketing pairwise judging requires exactly one hypothesis winner, chosen on content quality using relevance, specificity, plausibility, and value alignment.The prompt requests reasoning that explains the key differentiators and decision.Source: §D.3 · p. 29

D.4. Paper-abstract exploration

• Paper memos must state a researcher-specific relevance hypothesis and one concrete next step, without re-summarizing the abstract.The prompt directs the model to mutate one example, cross over two or more, or explore a new angle, and to report the operator and source examples separately from the standalone memo.Source: §D.4 · p. 30
• Paper recommendation scoring uses a calibrated 0–10 scale assessing researcher relevance, technical quality and novelty, memo insights, and potential research directions.Generic memos and vague directions should score lower; concrete, falsifiable, technically grounded next steps are preferred.Source: §D.4 · p. 31
• The paper novelty check requires a genuinely different memo framing or research direction, not merely a different paper topic or superficial wording.It compares the new recommendation’s reading angle with existing ones and requests a novelty rationale and Boolean decision.Source: §D.4 · p. 32
• Paper pairwise judging requires one winner based on researcher relevance, specificity and depth, concrete research directions, and potential to inspire new work.The judge compares memo content alone and supplies reasoning for the decision.Source: §D.4 · p. 32

D.5. Goodreads-Children (grounding)

• Goodreads uses held-out reader star ratings instead of an LLM quality judge; a non-novel book replaces its nearest neighbor only with a strictly higher rating.The bandit uses rating/5 for seed initialization and updates, while reported Quality stays on the 1–5 scale; previously tried books, including seeds, are excluded from selection.Source: §D.5 · pp. 32–33
• Goodreads memos must connect the book to the reader’s demonstrated tastes and avoid plot restatement, invented details, or guessing a star rating.The prompt also uses mutate, crossover, and explore as memo-building operators and asks the model to report the operator and source examples.Source: §D.5 · p. 33
• The grounding prompt includes the reader profile as {profile}.Source: §D.5 · p. 34
• It presents the reader’s existing shelf through {recommendations_text}.Source: §D.5 · p. 34
• It identifies the candidate item as the NEW recommendation using {new_recommendation_label}.Source: §D.5 · p. 34
• The prompt asks whether the new recommendation differs meaningfully from the existing ones and adds genuine variety to the reader’s shelf.Source: §D.5 · p. 34
• Novelty is judged by the kind of book and its appeal, not by quality.Source: §D.5 · p. 34
• A superficially different title is not novel if it offers the same reading experience, and judgments must rely on content and appeal rather than guessed ratings or scores.Source: §D.5 · p. 34
• The response must follow an exact JSON format with 1–2 sentences of novelty reasoning and an is_novel true-or-false value.Source: §D.5 · p. 34

Researcher profiles

• The ML researcher profile centers on divergent search and its stepping stones, including novelty measurement, archive mechanics, model-based operators, auto-curricula, exploration, and recursive self-improvement.It is skeptical of pure scaling and benchmark-chasing and looks for specific connections to open-ended search.Source: §D.1 · pp. 20–21
• The AI startup researcher evaluates research for practical leverage, such as cheaper, safer, faster, or more reliable products and potential effects on a 3-to-12-month roadmap.The profile also recognizes that major-conference work may be theoretical, exploratory, or several steps removed from company use.Source: §D.1 · p. 21

Acquirer profiles

• The M&A experiments use Walt Disney Co and JPMorgan Chase & Co as acquirers, with their company profiles describing location, business activity, and financial snapshots.Disney’s profile spans entertainment, sports, and experiences; JPMorgan’s profile covers consumer and community banking, commercial and investment banking, and asset and wealth management.Source: §D.1 · p. 21

Product profiles

• The 3D-printer marketing profile presents LAYERCRAFT as a $500–$900 desktop brand for makers, with assembled setup and three models for different printing needs.Its offerings include an enclosed compact model, a larger engineering-materials model, and a four-spool color model.Source: §D.1 · p. 22
• The SkillForge profile describes a fictional learning platform priced at $89–$249 monthly, with business, technology, and creative tracks aimed at practical, job-ready skills.Courses include expert instruction, weekly live Q&As, learner community, feedback on assignments, and progress tracking.Source: §D.1 · pp. 22–23

M&A example

• 5/10 for QD-Rec’s Liberty TripAdvisor acquisition thesis, versus 1/10 for Novelty, after adapting a Southwest Airlines hypothesis about linking travel purchases with financial services.The child proposed connecting activities, dining, and accommodation bookings to Chase Travel, rewards redemption, and installment financing.Source: M&A target discovery: JPMorgan Chase &… · pp. 35–36

Example caveat

• The M&A example warns that generated recommendations can include claims absent from the input and may rely on stale or incomplete company information.The authors say the example used basic financial data and short business descriptions and that actual decisions would require more detailed, up-to-date information.Source: M&A target discovery: JPMorgan Chase &… · p. 36

Marketing example

• The nurse example’s parent hypothesis concerned a preschool teacher’s move toward district-wide curriculum leadership, supplying a skills-based rationale that the child adapted to the nurse’s screening-program goals.The parent received 7/10; the child retained project management, data analysis, and presentation skills as the relevant bridge.Source: Marketing: Upskilling (SkillForge) · pp. 37–38
• 7/10 for QD-Rec’s SkillForge recommendation to a nurse, versus 2/10 for Novelty, as the adapted hypothesis connected his leadership goals to business and data skills.The child cited project management for coordinating screenings, data analysis for tracking outcomes, and presentation skills for securing support; it also noted that SkillForge cannot replace nursing coursework.Source: Marketing: Upskilling (SkillForge) · pp. 37–38