When Debate Messages Help or Hurt: Diagnosing Communication in Self-Distillation
Abstract
AbstractMulti-agent debate is widely used to enhance reasoning in language models, yet the exchanged messages are rarely analyzed. We study these messages in a Debate+SFT self-distillation loop by tracking their round-level evolution, intervening on peer messages with T-CIC at inference time, and comparing against matched Solo controls. Across three 5-seed discovery tasks–MATH, GSM-Hard, and 2WikiMultiHopQA–we observe weak refinement on MATH (+0.9pp for Debate versus +9.0pp for Solo), harmful homogenization on GSM-Hard (-11.3pp, with messages 27.5% longer and more self-similar), and mild inflation on 2WikiMHQA (-0.5pp with moderate length growth). A 2-seed T-CIC check on round-4 adapters shows different forms of message dependence: MATH collapses when peer messages are removed, shuffled, or replaced (-16 to -23pp), GSM-Hard is selectively sensitive to mismatched context, and 2WikiMHQA is mostly stable except under cross-problem shuffling. Debate messages should therefore be measured and stress-tested before they are trusted as a training signal.