Don't Make the LLM Read the Graph: Make the Graph Think

A new paper on arXiv (arXiv:2604.23057) by Yuqi Sun and colleagues investigates whether explicit belief graphs improve LLM performance in cooperative multi-agent reasoning, using the card game Hanabi as a testbed. Across 3,000+ controlled trials with four LLM families (including GPT-4o, Gemini, and Llama 70B), the researchers established four key findings. First, the integration architecture of belief graphs matters enormously: when used as prompt context, graphs are only beneficial for weak models on 2nd-order Theory of Mind tasks (80% vs 10% performance, p<0.0001). However, when graphs gate action selection through ranked shortlists, they become structurally essential even for strong models (100% vs 20% on 2nd-order ToM, p<0.001).

The study also uncovered 'Planner Defiance,' a model-family-specific failure where LLMs override correct planner recommendations at partial competence. Gemini models showed near-zero defiance, while Llama 70B exhibited a 90% override rate (replicated N=20). Notably, models distinguished between factual context (which they deferred to) and advisory recommendations (which they overrode). Full-game evidence confirmed that inter-agent conventions outperform all single-agent interventions by 128% over baseline (p=0.003). Preliminary scaling analysis suggests that graph depth has diminishing returns: shallow graphs offer the best cost-benefit ratio, while deeper Theory of Mind graphs appear harmful at larger player counts (-1.5 points at 5-player, p=0.029).