arXiv 2604.230572026-06-04cs.AI

不要让大语言模型读图：让图思考

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

Mindoverflow；University of Waterloo（多伦多大学）；Carnegie Mellon University（卡内基梅隆大学）；Foothill College（foothill学院）；Purdue University（普渡大学）；University of Wisconsin（威斯康星大学）；Apple（苹果公司）

Yuqi Sun, Tianqin Meng, George Liu, Yashraj Panwar, Lakshya Chaudhry, Munasib Ilham, Aman Chadha

备注：: main body has 9 pages, 4 figures, under review for COLM 2026 conference

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

通过3000多次对照实验，研究显式信念图在合作多智能体推理中是否提升LLM性能，发现集成架构决定图的价值，识别出“规划者违抗”现象，并证明图深度收益递减。

摘要：

我们研究了显式信念图是否提升大语言模型在合作多智能体推理中的性能。通过在合作纸牌游戏Hanabi中跨越四个LLM家族的3000多次对照实验，我们建立了四个发现。首先，集成架构决定了信念图是否提供价值：作为提示上下文，图对强模型是装饰性的，仅对弱模型在二阶心智理论上有益（80% vs 10%，p<0.0001，OR=36.0）；当图通过排序短列表门控动作选择时，即使对强模型也变得结构上必要（二阶ToM上100% vs 20%，p<0.001）。其次，我们识别出“规划者违抗”，一种模型家族特定的失败，即LLM在部分能力下覆盖正确的规划者建议（90%覆盖，重复N=20）；Gemini模型表现出接近零的违抗，而Llama 70B表现出90%，模型区分事实上下文（被延迟）和咨询建议（被覆盖）。第三，完整游戏证据证实智能体间约定（比基线高+128%，p=0.003）优于所有单智能体干预，且单个信念图组件必须组合才能产生收益。第四，初步规模分析（N=10/单元，探索性）表明图深度收益递减：浅层图提供最佳成本效益比，而更深层的ToM图在更大玩家数量下似乎有害（5玩家时-1.5分，p=0.029）。

英文摘要：

We investigate whether explicit belief graphs improve LLM performance in cooperative multi-agent reasoning. Through 3,000+ controlled trials across four LLM families in the cooperative card game Hanabi, we establish four findings. First, integration architecture determines whether belief graphs provide value: as prompt context, graphs are decorative for strong models and beneficial only for weak models on 2nd-order Theory of Mind (80% vs 10%, p<0.0001, OR=36.0); 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). Second, we identify "Planner Defiance," a model-family-specific failure where LLMs override correct planner recommendations at partial competence (90% override, replicated N=20); Gemini models show near-zero defiance while Llama 70B shows 90%, and models distinguish factual context (deferred to) from advisory recommendations (overridden). Third, full-game evidence confirms inter-agent conventions (+128% over baseline, p=0.003) outperform all single-agent interventions, and individual belief-graph components must be combined to produce gains. Fourth, preliminary scaling analysis (N=10/cell, exploratory) suggests graph depth has diminishing returns: shallow graphs provide the best cost-benefit ratio, while deeper ToM graphs appear harmful at larger player counts (-1.5 pts at 5-player, p=0.029).

机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。