Beyond Sally-Anne: Evaluating Theory of Mind in LLMs using Epistemic Schelling Points

2026-07-13 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

The Epistemic Asymmetry Schelling Task (EAST) is introduced as a novel two-player dialogue game designed to benchmark robust and generalizable Theory of Mind (ToM) abilities in Large Language Models. This new evaluation method addresses limitations of traditional text-based ToM tests, such as the Sally-Anne task, which can be "gamed" due to LLM exposure during pre-training and may not reflect functional ToM in naturalistic settings. By requiring LLM-LLM dyads to independently converge on semantic Schelling points under varying states of epistemic transparency, EAST reveals a significant capability gap in functional social reasoning. Results indicate that only frontier models successfully navigate these tasks, with coordination failures primarily stemming from epistemic tracking errors, like conflating private and mutual knowledge. This highlights robust social reasoning and epistemic tracking as critical bottlenecks for LLM development, despite high performance on static benchmarks.

Key takeaway

For AI Scientists and Machine Learning Engineers focused on developing more human-like social reasoning in LLMs, you should recognize that traditional ToM benchmarks are insufficient. Your development efforts must prioritize improving epistemic tracking, specifically preventing models from conflating private and mutual knowledge. Integrating dialogue-based evaluation methods like EAST into your testing pipeline will provide a more robust assessment of functional social intelligence, guiding targeted improvements beyond superficial performance.

Key insights

The Epistemic Asymmetry Schelling Task (EAST) reveals LLMs' significant functional social reasoning and epistemic tracking gaps.

Principles

• Traditional ToM benchmarks are insufficient.
• Epistemic tracking is critical for social reasoning.
• Dialogue games test robust ToM capabilities.

Method

EAST is a two-player dialogue game where LLM-LLM dyads independently converge on semantic Schelling points under varying epistemic transparency to evaluate robust Theory of Mind.

In practice

• Prioritize LLM development on epistemic tracking.
• Utilize dialogue games for robust ToM evaluation.
• Analyze coordination failures for epistemic errors.

Topics

• Large Language Models
• Theory of Mind
• LLM Evaluation
• Epistemic Schelling Task
• Social Reasoning
• Coordination Games

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• NICE: A Theory-Grounded Diagnostic Benchmark for Social Intelligence of LLMs — NICE provides a theory-grounded diagnostic benchmark to identify specific social intelligence weaknesses in LLMs, crucial for their application in social contexts.
• Theory of Mind and Persuasion Beyond Conversation: Assessing the Capacity of LLMs to Induce Belief States via Planning and Action — Large Language Models can induce belief states in others through planned actions, extending Theory of Mind beyond conversation.
• Reasoning as Pattern Matching: Shared Mechanisms in Human and LLM Everyday Reasoning — Human and LLM everyday reasoning share pattern-matching mechanisms, leading to similar error patterns.
• Beyond Social Pressure: Benchmarking Epistemic Attack in Large Language Models — LLMs exhibit distinct vulnerabilities to philosophical pressure, requiring targeted diagnostic benchmarks and mitigation strategies.
• Traces of Social Competence in Large Language Models — LLMs' apparent social competence, particularly in False Belief Tests, is driven by specific mental-state vocabulary rather than deep understanding.
• Narration-of-Thought: Inference-Time Scaffolding for Defeasible Ethical Reasoning in Large Language Models — Narration-of-Thought (NoT) enhances LLM ethical reasoning by structuring chain-of-thought to explicitly address stakeholders and uncertainties.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.