Compare to Correct: Pairwise-Guided Logit Adjustment for LLM Test-Time Scaling

Key points

• Inspired by the Condorcet principle in social choice theory, we introduce pairwise confidence comparisons between paths supporting different answers to capture answer-level comparative advantage.1
• We combine this comparative evidence with voting support through a KL-regularized objective.1
• Theoretically, we show that this comparative-advantage is equivalent to the adjustment to the traditional voting logits.1
• In this way, our proposed Pairwise-Guided Logit Adjustment (PGLA) framework allows comparative evidence to compensate for weaker voting support and correct voting-based ranking errors.1

Sources (1)

• [1]Compare to Correct: Pairwise-Guided Logit Adjustment for LLM Test-Time Scaling
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 03:46 PM

  Sampling-based test-time scaling enhances LLM reasoning by aggregating multiple reasoning paths for the same query.

  Inspired by the Condorcet principle in social choice theory, we introduce pairwise confidence comparisons between paths supporting different answers to capture answer-level comparative advantage.

Extractive summary: sentences quoted from the sources.