PDDL-Mind: Large Language Models are Capable on Belief Reasoning with Reliable State Tracking

The core insight isn't that LLMs are bad at social reasoning — it's that they conflate two distinct problems: tracking what is objectively true in the world versus modeling what a specific agent believes to be true. PDDL-Mind treats these as separate computational steps, which is what makes the symbolic grounding useful rather than decorative.

This connects directly to the recursive instability finding in 'Generalization in LLM Problem Solving: The Case of the Shortest Path' (arXiv, mid-April), where models failed at longer planning horizons precisely because implicit state tracking degraded with depth. PDDL-Mind is essentially the same diagnosis applied to social cognition: when reasoning chains get long, unstructured token prediction accumulates errors that explicit state representations would prevent. The broader thread running through recent coverage is that pure neural approaches keep hitting ceilings on tasks requiring logical consistency across steps, and symbolic scaffolding keeps re-emerging as the practical patch.

The real test is whether PDDL-Mind's gains on MMToM-QA hold on MuMA scenarios involving more than two agents with conflicting beliefs, since multi-agent state explosion is where PDDL representations historically become unwieldy. If the authors release ablations on that condition, it will clarify whether the framework scales or just tidies up the easy cases.