When Debate Helps: Proposal Supply and Verification-Aware Readout in Multi-Agent Reasoning
Debate only helps when agents both surface the right answer and then recognize it despite a bad majority.
The paper separates those two jobs into proposal supply and verification-aware readout. It defines “recoverable headroom” for cases where a correct proposal exists but majority voting would choose wrong. Its Latent Verification Debate model tests whether evidence for a candidate answer changes final decisions while the proposal set stays fixed. The authors report that coverage-selected agent societies improve complementary proposal supply and aggregate accuracy across matched-budget reasoning benchmarks. ArXiv · AI/CL/LG's note
The paper separates those two jobs into proposal supply and verification-aware readout. It defines “recoverable headroom” for cases where a correct proposal exists but majority voting would choose wrong. Its Latent Verification Debate model tests whether evidence for a candidate answer changes final decisions while the proposal set stays fixed. The authors report that coverage-selected agent societies improve complementary proposal supply and aggregate accuracy across matched-budget reasoning benchmarks. ArXiv · AI/CL/LG's note
score 4