Do LLMs Act on What They Know? From Partner Representations to Cooperative Actions

Key points

• We study this problem in a controlled Hanabi-derived environment with scripted hint generation, LLM-controlled receiving decisions, and frozen model weights.1
• We compare probe-predicted and ground-truth conventions presented either as general rules or as externally computed action recommendations.1
• In a Qwen3-8B case study, matched-state statement reversals reveal much greater sensitivity to action recommendations than to rule statements.1
• Activation transfers from oracle-action and non-oracle hint-restatement donors improve intent accuracy on both action classes, but the tested alternatives do not reliably reproduce these benefits.1

Sources (1)

• [1]Do LLMs Act on What They Know? From Partner Representations to Cooperative Actions
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 6, 10:44 AM

  Cooperation with unfamiliar partners requires adapting to communication conventions that are not known in advance.

  We study this problem in a controlled Hanabi-derived environment with scripted hint generation, LLM-controlled receiving decisions, and frozen model weights.

Extractive summary: sentences quoted from the sources.