We test AgentCoMa on 61 contemporary LLMs of different sizes, including reasoning models (both SFT and RL-tuned). While the LLMs perform well on commonsense and math reasoning in isolation, they are far less effective at solving AgentCoMa tasks that require their composition!
