The evaluates LLMs at the process level rather than only checking final answers. It organizes agentic mathematical problem solving into reusable atomic capabilities, then creates planning, action, and feedback tasks across textual and multimodal settings. Trajectories are synthesized automatically, and controlled LLM rewriting produces fine-grained annotations.
The key result is that final-answer accuracy can be misleading: models with similar end-to-end accuracy can show markedly different profiles of planning, action, and feedback ability. This makes more useful for diagnosing where models fail and for guiding work on more robust mathematical agents.