At a Glance
Pooling information can make a group better at guessing the single best answer but much worse at actually finding the target if everyone repeats that one recommendation; separate how you aggregate information from how you assign actions to preserve discovery.
ON THIS PAGE
What They Found
Pooling eight noisy reports improves the single best recommendation (the posterior mode) but concentrates all effort on that option, cutting the group's chance of success from about 83% to about 38%. A coordinator that assigns the eight highest-probability boxes instead of repeating one answer restores and even slightly improves group success (about 86%). Simple incentive tweaks—splitting rewards or paying only the unique finder—partially or fully recover portfolio-style search, while shared or copied sources of information can collapse independent channels and change which protocol wins. Market-Based Coordination Pattern
Not sure where to start?Get personalized recommendations
Data Highlights
1Eight independent clue-followers: group discovery = 83.22%; if all follow the pooled single recommendation, discovery = 38.35%.
2A planner who assigns eight distinct boxes using the same pooled reports achieves discovery = 85.94% (better than decentralized search).
3Equal-split reward raises discovery to 59.91%; a 'sole-rescue' reward that pays only a unique finder implements the planner optimum; common-source copying flips advantages when copying probability c ≈ 0.788462.
What This Means
Engineers and architects building systems of collaborating agents (or people) should care because information sharing without action allocation can destroy collective performance. Product leaders and org designers should treat evidence aggregation and task assignment as two separate design problems to avoid wasting capacity on the same candidate. Multi-Agent System
Key Figures
Fig 1: Figure 1: The exact pooled-information frontier for the canonical benchmark. The one-action consensus value is only the left endpoint. Seven coordinated actions recover the performance of eight decentralized clue-followers.
Fig 2: Figure 2: Exact ex-ante discovery under four protocols in the independent sixteen-box benchmark. The labels inside the bars report the expected number of distinct boxes searched. The dashed line marks the blind coordinated portfolio at 8 / 16 = 0.50 8/16=0.50 : informed consensus falls below a protocol that ignores the clues entirely. Equal sharing supplies partial dispersion, but not enough to match natural coverage from independent private clues.
Fig 3: Figure 3: Exact canonical planner gain V 8 ( c ) − G private ( c ) V_{8}(c)-G_{\mathrm{private}}(c) . As common-source copying creates more duplicated capacity, identifying and reallocating it becomes more valuable.
Fig 4: Figure 4: Exact canonical comparison under common-source copying. The symmetric equal-split market initially trails decentralized report-following, then overtakes it when channel collapse becomes sufficiently severe. The crossover is a canonical computational result, not a general theorem.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
The model is a deliberately small, idealized benchmark: exactly one valuable target, eight searchers, and symmetric noisy clues, so real systems with multiple targets, costs, or heterogenous agents may behave differently. Coordinator solutions assume the planner can observe pooled information and reassign actions, which may not be feasible in decentralized settings. Implementing the sole-rescue reward may be hard in practice because of attribution, collusion, or risk preferences—it's a design ideal, not a turnkey contract for all contexts. Orchestrator-Worker Pattern Guardrails Pattern
Methodology & More
Start with a simple test: 16 boxes, one jackpot, eight searchers. Each searcher gets a noisy clue that names one box correctly 20% of the time. If every searcher follows her private clue independently, the group finds the jackpot about 83.22% of the time. If everyone pools their clues, picks the single most likely box, and all open that same box, the best single recommendation rises to 38.35% accuracy—but group discovery plunges to 38.35% because all eight attempts are the same. The paradox is that better shared estimation (a stronger single guess) can lead to worse discovery when actions are not diversified. The paper separates the value of information from the value of action allocation. With the same pooled information, a planner who assigns the eight highest-posterior boxes gets discovery up to 85.94%—higher than decentralized clue-following—showing the loss comes from converting a portfolio of actions into one repeated bet. Letting selfish agents split the prize when they tie improves dispersion (raising discovery to about 59.91%), but anonymous symmetric equilibria can still collide. Paying only the unique finder (a "sole-rescue" rule) aligns individual incentives with discovery: every pure equilibrium then covers the best portfolio. Finally, if many agents copy the same unobserved source, effective diversity collapses and the value of reallocation grows; beyond a copying threshold (around 0.788), a dispersing market can beat private report-following. The actionable lesson: keep evidence sharing and task assignment separate—share the evidence, but explicitly diversify who acts and consider incentives that reward unique contributions. Tree of Thoughts Pattern Dynamic Task Routing Pattern
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
Single-author arXiv paper with very low h-index and no listed affiliation. Little recognizable institutional or author reputation — minimal credibility signal.