QD-Rec's archives during search in three domains
(claude-sonnet-5 proposer). Each map cell shows the highest score among its
recommendations. Red lines extend from earlier recommendations (parents) to the new
recommendation (child) generated from their ideas. The text below each map describes the proposal and whether it is added,
replaces an entry, or is discarded. Companies in the M&A run are labeled by industry.
Open-ended search aims to continually produce new discoveries that support further
exploration
We investigate whether open-ended search can extend to recommendation for decision-making. Unlike open-ended generation, where foundation models create new artifacts, many real-world decisions require choosing from a fixed candidate pool. An acquirer must select among existing companies, and a marketer among available customers. At first glance, a finite, predefined pool seems to leave no room for open-ended exploration.
However, a fixed set of items does not imply a fixed set of reasons for selecting them. Even for the same decision maker, each item can be recommended for different reasons. For example, an acquirer might consider buying a company to gain technology or enter a new market. The search over these reasons remains open-ended even when the candidate set is finite.
We formalize this task as open-ended recommendation. Its unit of search is an item–hypothesis pair: a candidate item and a natural-language explanation of why it fits a decision maker's profile. The profile describes their goals, interests and background. Our objective is to find a diverse, high-quality collection of these pairs for one profile.
These recommendations give the decision maker different options and reasons to compare before
making the final choice. For high-stakes decisions, the goal is to provide several well-reasoned
alternatives that reflect different considerations. This objective follows the
quality–diversity (QD)
approach of finding a set of high-quality, diverse solutions
Two properties of this problem shape our approach.
First, evaluation is costly. A recommendation's value is unknown until its hypothesis has been generated and judged. With candidate pools spanning thousands or tens of thousands of items, evaluating even one hypothesis per item can exceed the available compute budget. The system must use previous evaluations to choose which items to investigate next.
Second, the space of possible hypotheses is much larger than the finite pool of items.
Evaluating every item once would still leave many other reasons for choosing each item unexplored.
One way to explore these possibilities is to adapt and combine ideas from previously generated
hypotheses. Even hypotheses with modest scores can be useful in this process: an idea that works
poorly for one item may lead to better reasoning for another. Such hypotheses can serve as
stepping stones, and QD and novelty search provide ways to retain diverse alternatives as the
search continues
We developed QD-Rec, which takes a decision maker's profile and a fixed pool of candidate items, and maintains an archive of recommendations. Each entry contains an item, a hypothesis and a quality score. The system repeats three steps: select an item, generate a hypothesis, and decide whether to keep the new pair.
QD-Rec uses a contextual bandit to select the next item for hypothesis generation and
evaluation. It uses item embeddings and previous evaluation scores to balance exploitation of
promising items with exploration of
candidates whose potential is less certain
Note that previous work has also combined bandits with QD, but for different purposes, such as
selecting parents (or bins) from the archive
The proposer LLM receives the profile, the selected item and three examples sampled uniformly
from the archive. These examples supply ideas the proposer can adapt to the selected item.
The model uses one of three operators
We call an example used in a proposal its parent, and the new hypothesis its child. As recommendations enter the archive, they become available as parents for later proposals. This lets useful reasoning carry over between different items.
An evaluator scores the new item–hypothesis pair. To decide whether to keep it, QD-Rec compares it with nearby entries in the archive. We embed the item description together with the hypothesis text and retrieve the three entries with the highest cosine similarity. The archive update then asks two separate questions.
This local competition follows the idea of ecological niches. Recommendations with
distinct reasons can coexist in different niches, while similar ones compete within a niche.
A new reason can be retained even if its current score is low. It can then
provide an idea for a later proposal. This archive design follows recent QD systems that use
embeddings to identify similar entries
We test QD-Rec in the following domains.
Our paper also studies book recommendation using recorded human ratings instead of LLM quality
scores
We compare QD-Rec with two ablations, simpler versions that remove components of the method. All three use the same novelty check and local replacement rule, so their differences concern which items they investigate and how they write hypotheses.
| Method | Local novelty filter | Evolutionary operator | Bandit item selection |
|---|---|---|---|
| Novelty | ✓ | × | × |
| Novelty+Evo | ✓ | ✓ | × |
| QD-Rec (full) | ✓ | ✓ | ✓ |
Table 1. QD-Rec and its two ablations. All three use the same archive update: novel recommendations are added, while non-novel ones replace their nearest neighbor only if they win the local quality comparison.
We run each of these six scenarios with two proposers, claude-sonnet-5 and
claude-haiku-4-5. The evaluator is gpt-5.6-terra for quality scoring,
novelty checks and pairwise comparisons, to cope with self-preference
bias
Each run starts with ten seed recommendations, followed by 1,000 evaluations. We repeat each
configuration with five random seeds and give the three methods the same initial archive for
each seed. The seeds control sampling such as item ordering and parent selection; LLM sampling
stays at the provider default.
Candidate and archive embeddings use OpenAI's text-embedding-3-small at 512 dimensions.
Four metrics summarize the final archive of a run.
The scores used to report Quality are the same scores available during search. The experiment measures how well each method uses a fixed evaluation budget to build its archive under that evaluator.
Two examples show how an earlier hypothesis can be adapted to a different candidate.
Both examples use the claude-sonnet-5 proposer
and gpt-5.6-terra judge. In the excerpts, [...] marks text omitted from the generated hypotheses.
M&A (acquirer: JPMorgan Chase & Co)
The Novelty baseline found little reason for JPMorgan Chase & Co to acquire Liberty TripAdvisor Holdings. It argued that Chase Travel and existing partnerships already provided access to similar benefits, and recommended against the acquisition. The judge assigned the proposal a score of 1/10.
Novelty, without exemplars (score 1/10):
This deal lacks strategic logic: JPMorgan is a diversified financial holding company (banking, markets, asset management) while Liberty TripAdvisor is a consumer travel marketplace business with no meaningful overlap in capabilities, technology, or customer needs that JPMorgan doesn't already access via its Chase Travel platform and existing card-linked travel partnerships. [...] Any travel-related synergy JPMorgan seeks (e.g., enhancing Chase Sapphire travel benefits or Chase Travel bookings) can be achieved more efficiently through commercial partnerships or licensing rather than full acquisition and integration of a publicly-troubled, sub-scale asset. We recommend against pursuing this acquisition given the poor strategic fit and absence of compelling synergy rationale.
QD-Rec drew on an idea from an unlikely source: a proposal to acquire Southwest Airlines. This earlier hypothesis argued that owning the airline would connect travel purchases more directly with JPMorgan’s financing, payments and rewards services. However, the judge considered the acquisition impractical, citing airline operating risks and restrictions on bank ownership of nonfinancial businesses. This proposal also received 1/10.
Parent, for Southwest Airlines Co (score 1/10):
Southwest Airlines is unusually well-suited to JPMorgan given the bank's existing role as issuer of the Southwest Rapid Rewards co-brand card portfolio, one of the largest travel rewards programs in the country; full ownership would let JPMorgan capture the entire loyalty-program economics—breakage, points sales, and ancillary fee revenue—rather than sharing them through an issuing partnership, [...] At an EV of $15.3B, Southwest is trading below its levels of two ($15.8B) and four years ago ($17.5B) despite a modest 1.3% YTD gain, offering JPMorgan an opportunity to acquire a top-tier domestic airline brand and its 100+ million-member loyalty base at a valuation depressed by sector-wide fuel and margin pressure rather than by any structural decline in Southwest's franchise. Vertical integration would let JPMorgan embed banking, financing, and payments products directly into the travel purchase funnel—installment financing for fares through Chase's balance-sheet lending, bundled Chase Travel booking, and richer transaction data for underwriting and cross-sell into CCB's card and deposit products. [...] though it would require JPMorgan to underwrite genuine airline operating risk (fuel, labor, capacity cycles) that sits outside its traditional risk models.
Using the Southwest hypothesis as a parent, QD-Rec generated a new rationale for acquiring Liberty TripAdvisor. The proposal connected its travel marketplaces with Chase’s booking, payment and rewards services, and identified opportunities to offer financing for activities, dining and accommodation. It carried over the idea of integrating financial services into travel purchases, without requiring JPMorgan to operate an airline.
Child, QD-Rec (score 5/10):
Liberty TripAdvisor's three marketplaces—Tripadvisor, Viator, and TheFork—represent a global audience of travel researchers and bookers that JPMorgan could plug directly into Chase Travel, its growing proprietary booking platform, and into Chase Offers and Ultimate Rewards redemption, capturing interchange, points-funded bookings, and installment-financing revenue on activities, dining, and accommodations that currently flow to third-party payment rails. [...] At a $2.08B enterprise value—down sharply from $3.39B three years ago and off 49% YTD—the company is trading near a multi-year low, [...] Owning these consumer touchpoints would also generate rich transaction and behavioral data to sharpen underwriting and cross-sell into Chase's card and travel-rewards products, [...]
The judge rated this Liberty TripAdvisor proposal 5/10, compared with 1/10 for the baseline’s independently generated proposal for the same company. Concerns remained about competition and whether partnerships could deliver the same benefits. Even so, the Southwest hypothesis had supplied a useful starting point: an idea from a proposal rated only 1/10 became a stepping stone toward a more highly rated recommendation for a different company.
Customer targeting (product: SkillForge, a fictional online learning subscription)
Here, the task is to find customers for SkillForge, a fictional monthly subscription offering business, technology and creative courses, and explain why it could appeal to them. The candidate is Jesus Montalvo, a fictional 29-year-old nurse in National City, California. He wants to earn a nursing degree, become a charge nurse who leads a team, and start a preventive health screening program in his hometown.
The Novelty baseline wrote its hypothesis from scratch. It recognized that courses such as project management could help with the screening program, but dismissed this as “a peripheral use case”. It also treated his shift work as an obstacle to finding time for study and concluded that the service was a poor strategic fit. The judge gave the recommendation 2/10.
Novelty, without exemplars (score 2/10):
Jesus's near-term career goals are almost entirely clinical—finishing his BSN, becoming an FNP, and building a community health outreach program—none of which SkillForge's business, tech, or creative tracks can substitute for, since it offers no accredited nursing or healthcare credentialing. While he might be mildly drawn to the Business Track's project management or data analysis modules to sharpen his charge-nurse leadership and eventually structure a preventive-screening program, this is a peripheral use case rather than a core need, and his already packed per-shift schedule plus stated preference for unstructured downtime (hiking, guitar, cooking) makes committing another 3-4 hours weekly to a subscription-based platform a hard sell. [...] SkillForge is a poor strategic fit for this persona despite superficial overlap in the 'practical skill-building' ethos.
QD-Rec's archive contained a highly rated hypothesis about a different customer: Sharon English, a fictional 46-year-old kindergarten teacher who wants to move into district-wide leadership.
Parent, for Sharon English (score 7/10):
Sharon's explicit ambition to move from the classroom into a district-wide leadership role—developing STEM curricula and mentoring other teachers—requires competencies her ECE and STEM pedagogy training doesn't fully provide: project management to roll out a district-wide program, data analysis to evaluate learning outcomes, and polished presentation skills to advocate for inclusive policies to administrators and school boards. SkillForge's Business Track maps directly onto these gaps, offering practical frameworks she could immediately apply while still pursuing her master's degree. Her comfort with structure, spreadsheets, and hands-on STEM tools also makes the Tech Track's approachable Python and data courses appealing, letting her build lightweight tools or dashboards to track her STEM curriculum's impact—something she could point to when pitching herself for a curriculum specialist role. [...]
This parent positioned the service as a way to gain skills needed for leadership that her teaching training did not fully cover: project management, data analysis and presentation. The judge gave it 7/10.
QD-Rec's mutate operator adapted this reasoning to Jesus. Like the teacher,
a nurse moving into management or program leadership may need business skills beyond their
professional training. The child used these needs as the reason to recommend the service.
Child, QD-Rec (score 7/10):
Jesus's stated goal of becoming a charge nurse and eventually launching a preventive health outreach program in National City requires competencies his clinical training never covered—project management to coordinate a screening initiative, basic data analysis to track and report health outcomes, and polished presentation skills to pitch the program to hospital administrators or grant funders. SkillForge's Business Track speaks directly to that gap, while the Tech Track's approachable data courses could let him build simple dashboards to demonstrate the impact of his community screenings, reinforcing his evidence-based, protocol-driven mindset. His per-shift, flexible work schedule and self-described tendency toward disorganization make the platform's structured 3-4 hour weekly cadence, dashboard progress tracking, and cancel-anytime plan especially appealing, [...] While SkillForge can't replace the BSN or FNP coursework he ultimately needs, it fills a real and immediate void in the program-management and advocacy skills his broader career vision demands, making him a solid niche fit despite the platform's non-clinical focus.
Skills that the baseline dismissed as peripheral became central to running a screening program and presenting it to hospital administrators. The child also interpreted shift work differently: the service's flexible format could suit an irregular schedule.
The judge found this specific connection between his goals and the skills credible and gave the child 7/10. The teacher's hypothesis served as a stepping stone by providing a new way to assess a different customer. It helped QD-Rec find a persuasive reason to recommend the service to a nurse whom the baseline had judged a poor fit.
Table 2 reports the four metrics for the three variants on both profiles of each of these three domains, under both proposers.
| Domain | Profile | Proposer | QD-Rec | Novelty+Evo | Novelty | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Qual. | Div. | Cov. | n | Qual. | Div. | Cov. | n | Qual. | Div. | Cov. | n | |||
| M&A | Walt Disney Co | Sonnet 5 | 2.91±0.05 | .352±.001 | .26±.00 | 214±2 | 2.51±0.07 | .353±.001 | .18±.01 | 135±5 | 2.40±0.12 | .330±.001 | .12±.00 | 85±1 |
| Haiku 4.5 | 2.55±0.03 | .318±.001 | .27±.00 | 262±5 | 2.11±0.04 | .322±.002 | .23±.01 | 188±4 | 2.12±0.05 | .326±.002 | .19±.01 | 143±4 | ||
| JPMorgan | Sonnet 5 | 3.04±0.04 | .353±.001 | .23±.00 | 183±5 | 2.54±0.08 | .363±.003 | .15±.01 | 108±4 | 2.88±0.04 | .332±.001 | .13±.01 | 95±3 | |
| Haiku 4.5 | 2.71±0.02 | .316±.001 | .28±.00 | 238±7 | 2.17±0.04 | .324±.003 | .22±.01 | 166±9 | 2.25±0.04 | .331±.001 | .20±.00 | 150±3 | ||
| ML Papers | OpenEndedness | Sonnet 5 | 5.32±0.02 | .501±.001 | .44±.01 | 730±8 | 5.26±0.02 | .479±.002 | .47±.01 | 690±8 | 4.96±0.02 | .469±.001 | .43±.01 | 578±5 |
| Haiku 4.5 | 4.87±0.02 | .523±.002 | .30±.01 | 353±34 | 5.03±0.07 | .504±.002 | .26±.02 | 247±31 | 4.26±0.02 | .489±.001 | .39±.01 | 468±9 | ||
| AI Startup | Sonnet 5 | 6.24±0.02 | .521±.001 | .45±.01 | 789±8 | 6.24±0.03 | .502±.001 | .48±.00 | 737±8 | 6.21±0.01 | .495±.001 | .46±.00 | 658±7 | |
| Haiku 4.5 | 5.83±0.02 | .526±.002 | .43±.01 | 720±6 | 5.87±0.03 | .507±.001 | .47±.01 | 674±12 | 5.81±0.02 | .516±.001 | .46±.01 | 616±4 | ||
| Marketing | 3D printer | Sonnet 5 | 7.05±0.05 | .243±.001 | .18±.00 | 164±2 | 6.73±0.01 | .237±.001 | .18±.01 | 156±5 | 6.81±0.04 | .219±.001 | .13±.01 | 96±4 |
| Haiku 4.5 | 6.83±0.03 | .199±.001 | .13±.00 | 104±3 | 6.29±0.02 | .202±.002 | .14±.00 | 110±3 | 6.65±0.05 | .190±.003 | .11±.00 | 80±3 | ||
| Upskilling | Sonnet 5 | 6.89±0.06 | .235±.001 | .11±.00 | 82±3 | 6.70±0.06 | .230±.002 | .10±.00 | 72±3 | 6.47±0.08 | .225±.001 | .09±.00 | 61±1 | |
| Haiku 4.5 | 6.62±0.10 | .245±.002 | .07±.00 | 54±3 | 6.35±0.12 | .238±.005 | .08±.00 | 58±3 | 6.47±0.03 | .221±.001 | .08±.01 | 69±4 | ||
Table 2. Mean ± standard error over five seeds. Sonnet 5 and Haiku 4.5 are the
proposers claude-sonnet-5 and claude-haiku-4-5. All results use
gpt-5.6-terra as the judge. Quality is the judge score from 0 to 10.
For each profile, proposer and metric, shading marks the
best mean and every cell not
significantly different from it (paired two-sided
QD-Rec generally matches or improves on the two ablations in Quality and Diversity. It also builds the largest archive in most settings, retaining more distinct recommendations within the same evaluation budget.
The run histories show how these archives develop. Figure 2 follows one profile from each applied domain with the Sonnet proposer. QD-Rec's mean Quality rises over the course of these runs and finishes highest. Its archive also grows as new hypotheses are admitted.
claude-sonnet-5 proposer, mean ± standard error over five seeds).
Novelty writes each hypothesis from scratch. Novelty+Evo evolves hypotheses using archive
examples. Both select items uniformly at random. QD-Rec uses a diversity-aware bandit to learn
which items to investigate next.
QD-Rec uses open-ended search to build an archive of diverse recommendations for decision-making. Each recommendation includes a reason for choosing the candidate, helping people compare different possibilities. The system combines a contextual bandit for item selection, an LLM for adapting hypotheses, and local comparisons for archive management. Across the domains and two proposers, QD-Rec achieved the highest mean Quality in most settings and usually built the largest archive.
We kept the system relatively simple in this study. Future work could extend its three core components in several ways: