Teams that search together iterate faster but explore less: analysis of online machine-learning contests shows joint search raises the number of attempts while narrowing solution diversity, and both higher attempt volume and broader exploration boost contest success.

Searching together versus searching apart: Evidence from Kaggle

Marco S. Minervini, Tianyu He, Phanish Puranam · January 27, 2026 · Strategic Management Journal

openalex quasi_experimental medium evidence 7/10 relevance Summary only summary available; pdf_status=paywall DOI Source PDF

Structured author observations

Linked only from stored provider relations; the raw author line above is never matched by name.

OpenAlex

Latest observation: August 24, 2026

Marco S. Minervini provider ID
Tianyu He provider ID
Phanish Puranam provider ID

Semantic Scholar

Latest observation: August 24, 2026

Marco S. Minervini provider ID
Tianyu He provider ID
P. Puranam provider ID

In online machine-learning competitions, teams that search together submit more attempts but generate less diverse solutions than if members searched separately, and both higher attempt volume and greater exploration independently improve performance.

Citation observations

Cumulative provider counts captured on specific dates; providers are never combined.

0 cumulative citations

OpenAlex · Observed August 24, 2026

View corpus context

0 cumulative citations

Semantic Scholar · Observed August 24, 2026

View corpus context

Abstract Research Summary How does the mode of search—independently or jointly—affect collective search, a central component of organizational adaptation and innovation? Using naturally occurring data from a strongly incentivized online competition platform, we find that compared to their counterfactuals that search apart, groups searching together exhibit less exploration in their search outcomes as noted in prior experimental and computational modeling studies. However, groups searching together stimulate a greater number of search attempts from their members than groups searching apart, an effect that has so far remained unnoticed. Further, both search attempts and exploration contribute positively to search performance. This suggests that the choice of search mode should depend on the demands of strategic contexts that make either the variety or volume of solutions relatively more important for collective search. Managerial Summary How members of a team interact when looking for solutions to problems shapes their success. Using evidence from online machine learning competitions, we find that looking for solutions together encourages members to examing more alternatives but narrows the range of these ideas, while working independently generates greater variety but fewer alternatives being examined. We propose that the choice between the two depends on context. When innovation requires novel, distinctive solutions, as in early‐stage R&D, independent work may be more effective. When speed and volume matter, as in patent races or crowded digital markets, searching for solutions together delivers faster iteration. Managers can improve innovation outcomes by aligning the mode of teamwork with the strategic demands of the problem.

Summary

Main Finding

Groups that search together (collaboratively) produce fewer distinct ideas (less exploration) but generate a larger volume of attempts per member than comparable groups that search independently. Both the number of attempts and the degree of exploration independently improve search performance, so the optimal search mode depends on whether a task values variety (novel solutions) or volume/speed (rapid iteration).

Key Points

Collaborative search reduces exploration: teams that work together converge on a narrower set of solutions than the counterfactual case where the same members search independently.
Collaborative search increases effort/volume: teams searching together produce a greater number of search attempts (submissions) per member.
Both exploration (variety of solutions) and search attempts (trial volume) positively affect final performance.
Trade-off implication: independent search favors novelty/diversity; joint search favors speed/iteration and higher throughput.
Managerial recommendation: align team search mode to strategic needs—use independent search when novelty is paramount (e.g., early-stage R&D), use joint search when iteration speed and volume matter (e.g., patent races, fast-moving digital markets).

Data & Methods

Data source: naturally occurring behavioral data from a strongly incentivized online machine-learning competition platform (real-world setting rather than lab experiments).
Comparison approach: contrasted outcomes for groups searching together with counterfactuals where the same members searched apart (statistical construction of counterfactuals to isolate mode effects).
Key metrics:
- Exploration: measured as the variety/distance among solutions generated by the group.
- Search attempts: counted as the number of solution submissions or trials by group members.
- Performance: success metrics on the competition tasks (e.g., accuracy/score improvements).
Identification: observational analysis exploiting platform incentives and counterfactual comparisons; controlled for confounders typical of platform competition settings.
Limitations (implicit): observational design and platform-specific incentives may limit external validity to other organizational contexts; exact statistical models and robustness checks are in the full paper.

Implications for AI Economics

Organizational design and incentives:
- Firms should choose search/coordination modes to match the economic value of novelty versus iteration speed. This affects team structuring, reward systems, and R&D project design.
- Compensation and incentive schemes can be tailored to promote independent exploration when differentiated solutions are valuable, or joint collaboration when rapid improvement and first-to-market speed are rewarded.
Platform and market design:
- Competition/platform designers (e.g., Kaggle-like marketplaces, crowdsourcing systems) can influence collective output characteristics by enabling or discouraging joint work—shaping the supply of novel vs. iterated solutions.
- Market structure implications for innovation races: when markets prize quick iteration (winner-take-all, fast product cycles), collaborative search may be socially optimal; when markets prize distinct breakthroughs, promoting independent search may yield higher social returns.
Labor and productivity effects:
- The mode of coordination influences the productivity distribution of contributors: joint search amplifies throughput but risks homogenizing solutions, which matters for labor allocation in AI teams and crowd work.
Policy and competition:
- Regulators and policymakers interested in fostering innovation should consider how grants, procurement, and competition design influence whether firms/teams pursue exploratory vs. exploitative search modes.
Modeling and forecasting:
- Economic models of innovation should incorporate both dimensions (volume and variety) as separate channels through which collaboration affects outcomes; predictions about welfare, market dynamics, and optimal firm strategy change depending on which channel is dominant.
Practical guidance for AI projects:
- For projects aiming to produce novel model architectures or breakthrough capabilities, encourage independent parallel search processes with periodic cross-pollination.
- For tasks requiring rapid iteration (hyperparameter tuning, deployment speed, incremental model improvements), organize tighter collaborative workflows to increase submission throughput.

If you want, I can: (a) draft a short managerial checklist for choosing search mode in AI projects, or (b) outline how to incorporate these effects into a simple economic model of innovation. Which would you prefer?

Assessment

Paper Typequasi_experimental Evidence Strengthmedium — The study uses rich, high-frequency, real-world data from incentivized competitions, which supports credible measurement of search attempts, exploration, and performance, but causal claims rest on non-random variation and constructed counterfactuals, leaving open selection and unobserved confounding threats. Methods Rigormedium — Likely strong descriptive and comparative analysis with novel operationalization of search attempts and exploration and use of within-platform variation; however, the abstract provides no details on robustness checks, controls, or techniques (e.g., matching, fixed effects, instrumental variables) that would more fully mitigate endogeneity, so methodological rigor cannot be rated as high. SampleTimestamped submission and leaderboard data from a highly incentivized online machine-learning competition platform covering individual participants and self-organized teams; measures include number of search attempts (submissions), a metric of exploration/diversity of solutions, and performance outcomes (competition scores); sample size and exact platform not specified in the abstract. Themesinnovation org_design productivity human_ai_collab IdentificationObservational comparison using naturally occurring variation on an online machine-learning competition platform: the authors compare observed outcomes for groups that searched together to constructed counterfactuals of those same members searching apart using timestamped submission logs and inferred search modes; identification therefore relies on observational counterfactuals and within-platform controls rather than randomized assignment. GeneralizabilityPlatform- and task-specific: findings are from ML competitions and may not generalize to other R&D or problem-solving contexts, Participant selection: contestants are likely skilled, competitive data scientists, not representative of the broader workforce, Incentive structure: high-stakes contest incentives differ from typical organizational or academic research settings, Group formation: self-selection into teams may bias results and differs from employer-assigned teams, Narrow outcome measures: success measured by competition score may not capture long-term innovation value or commercialization

Claims (6)

Claim	Direction	Outcome	Confidence & Evidence	Details
Compared to their counterfactuals that search apart, groups searching together exhibit less exploration in their search outcomes. Creativity	negative	exploration in search outcomes (variety/range of ideas)	Reading fidelity high Study strength medium	not reported 0.48
Groups searching together stimulate a greater number of search attempts from their members than groups searching apart. Team Performance	positive	number of search attempts	Reading fidelity high Study strength medium	not reported 0.48
Both search attempts and exploration contribute positively to search performance. Innovation Output	positive	search performance (competition performance metric)	Reading fidelity high Study strength medium	not reported 0.48
Looking for solutions together encourages members to examine more alternatives but narrows the range of these ideas, while working independently generates greater variety but fewer alternatives being examined. Creativity	mixed	number of alternatives examined and variety/range of ideas	Reading fidelity high Study strength medium	not reported 0.48
The choice of search mode should depend on the demands of strategic contexts: independent work may be more effective when novel, distinctive solutions are required (e.g., early-stage R&D), while searching together is preferable when speed and volume matter (e.g., patent races or crowded digital markets). Organizational Efficiency	mixed	effectiveness of search mode given contextual strategic demands (novelty vs. speed/volume)	Reading fidelity high Study strength speculative	not reported 0.08
The empirical analysis uses naturally occurring data from a strongly incentivized online machine learning competition platform. Other	null_result	data source / study setting	Reading fidelity high Study strength high	not reported 0.8