Recognizing Atari and Assessing Sente: Controlled Evaluation Tasks for Language-Based Go Agents

Key points

• We construct a controlled evaluation set centered on atari and its relationship to sente in Go, examining terminology recognition under local board changes and move evaluation under different strategic conditions.1
• The model struggles to recognize changes in the atari relationship and assess its strategic implications.1
• Its terminology predictions are associated with move preferences even when those preferences contradict the reference evaluation.1
• Our controlled tests provide a framework for examining whether language-based Go agents connect domain terminology with precise board relationships and strategic move values.1

Sources (1)

• [1]Recognizing Atari and Assessing Sente: Controlled Evaluation Tasks for Language-Based Go Agents
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 16, 05:23 AM

  Domain terminology provides concise descriptions of game states and strategic relationships, but whether game-playing language models can accurately recognize these concepts and relate them to move evaluation remains unclear.

  We construct a controlled evaluation set centered on atari and its relationship to sente in Go, examining terminology recognition under local board changes and move evaluation under different strategic conditions.

Extractive summary: sentences quoted from the sources.