Filtering multi-agent messages by answer correctness throws away value: over 4 in 10 outcome-changing wrong messages help
Diverse Hypothesis Deliberation caches five independently generated messages per problem, then hides and reveals each one to the same downstream integrator to measure whether availability improves or harms final correctness — separating trajectory value from answer correctness. Across five math and science benchmarks and two model families (gpt-oss-120b, gemma-4-31B-it), wrong-but-helpful messages appear in every benchmark-model combination, and among wrong-answer messages that changed the outcome more than four in ten changes were helpful (p=0.0002). Interventions also showed complete messages beat isolated components, so the practical move is to stop gating agent-to-agent messages on a correctness check and measure marginal contribution to the integrator instead.
↳ Follow the thread