Mark Warschauer

Confidence signals in multi-agent debate systems can predict reasoning quality, but they reveal a surprising asymmetry favoring the Constructor over the Auditor.

Early-token confidence can predict reasoning quality in LLM debates better than full-sequence statistics, revealing critical insights into the generation process.

Sequential fine-tuning aligned with discourse structure boosts AES performance, outperforming larger models and revealing the power of task-aware training.