Do LLMs Act on What They Know? From Partner Representations to Cooperative Actions
The paper finds a gap between LLMs encoding a partner’s intent conventions and reliably using that information to cooperate.
In a Hanabi-derived setup, probes could recover sender intent conventions better than target conventions across eight LLMs. But the models’ receiving decisions did not consistently follow the sender’s convention. Direct action recommendations helped more than stating general rules, with a Qwen3-8B case study showing the same pattern under statement reversals. ArXiv · AI/CL/LG's note
In a Hanabi-derived setup, probes could recover sender intent conventions better than target conventions across eight LLMs. But the models’ receiving decisions did not consistently follow the sender’s convention. Direct action recommendations helped more than stating general rules, with a Qwen3-8B case study showing the same pattern under statement reversals. ArXiv · AI/CL/LG's note
score 4