🧀 BigCheese.ai

Social

Prover-Verifier Games improve legibility of LLM outputs

🧀

This paper explores improving the readability of Large Language Models (LLMs) outputs through Prover-Verifier Games. The authors propose a training algorithm which makes use of small verifiers and provers—both helpful and sneaky. Their studies reveal that legibility training aids humans in verifying solution correctness more efficiently.

  • Submitted on 18 Jul 2024
  • Last revised on 1 Aug 2024
  • Authors include Jan Hendrik Kirchner and 5 others
  • Focuses on legibility of LLM outputs
  • Study involved grade-school math problems