Wrong but Useful: Trajectory Value Beyond Answer Correctness in Multi-Agent Messages
The paper finds that incorrect agent messages can still improve a multi-agent system’s final reasoning.
The authors test this with Diverse Hypothesis Deliberation, replaying solver runs with individual messages shown or hidden to measure their “trajectory value.” Wrong-but-helpful messages appeared across all five math and science benchmarks and both model families tested. Among wrong-answer messages that changed final correctness, more than four in ten changes were helpful for each model. The authors argue that answer correctness is useful signal, but not enough to decide which agent messages should be kept. ArXiv · AI/CL/LG's note
The authors test this with Diverse Hypothesis Deliberation, replaying solver runs with individual messages shown or hidden to measure their “trajectory value.” Wrong-but-helpful messages appeared across all five math and science benchmarks and both model families tested. Among wrong-answer messages that changed final correctness, more than four in ten changes were helpful for each model. The authors argue that answer correctness is useful signal, but not enough to decide which agent messages should be kept. ArXiv · AI/CL/LG's note
score 5