TL;DR
MuSeToM is a new human-validated benchmark that tests whether Large Language Models (LLMs) can correctly determine when a character should revise their beliefs based on sensory evidence. The study reveals a significant "over-updating" bias: models frequently mistake perceptually salient but irrelevant information for definitive proof, failing to distinguish between what catches an agent's attention and what actually justifies a change of mind.
What Problem Does This Paper Address?
Theory of Mind (ToM)—the ability to reason about the mental states of others—is a cornerstone of social intelligence. While LLMs have shown promise in classical ToM tasks like the "Sally-Anne" false-belief test, these evaluations are often static. They ask what someone believes, but rarely if they should change that belief based on new, potentially ambiguous information.
In the real world, humans don't update their beliefs every time they see something new. We evaluate epistemic sufficiency: is this evidence strong enough to warrant a change? Current LLMs struggle with this. As shown in Figure 1, a model might see a character notice an "Almond Milk" carton nearby and incorrectly conclude the character now believes their pitcher contains almond milk, even if the carton was never poured into the pitcher. This confusion between perceptual salience (what is noticeable) and epistemic sufficiency (what is proof) is the core challenge MuSeToM addresses.
Methods
The authors developed a rigorous pipeline to create MuSeToM:
- Source Selection: They adapted narrative contexts from existing ToM datasets (BigToM and OpenToM).
- Multi-Sensory Cue Generation: Using a suite of LLMs (GPT-4o, DeepSeek-R1, Qwen), they generated cues across six sensory modalities: visual, auditory, olfactory, gustatory, tactile, and proprioception (awareness of body position).
- Human Validation: Three human annotators filtered and labeled every instance to ensure logical consistency and to categorize cues as:
- Positive: Sufficient evidence for belief revision.
- Hard Negative: Salient and relevant, but insufficient for a definitive update.
- Easy Negative: Entirely irrelevant to the belief.
- Evaluation Tasks: Models were tested on belief-change detection (binary) and updated-belief identification (multiple choice) across both "clean" (single cue) and "noisy" (cue + distractor) settings.
Datasets
- MuSeToM (New): Contains 2,415 human-validated instances derived from 220 source stories. It is the first benchmark to focus specifically on epistemic belief revision across six senses.
- BigToM & OpenToM: Used as the foundational narrative sources for story contexts and initial belief states.
Key Insights and Ideas
- The Over-Updating Bias: Models exhibit a "trigger-happy" approach to belief revision. They are much better at accepting sufficient evidence (True Positive Rate) than they are at rejecting plausible but insufficient evidence (True Negative Rate on hard negatives).
- Detection vs. Identification: Models are significantly better at detecting that a belief might change than they are at identifying what the new belief actually is.
- Salience Masks Failure: Aggregate accuracy scores often hide failures. A model might look competent overall, but a modality-level breakdown (Figure 5) shows that visual cues—while having high mean accuracy—often have the lowest rejection rates for insufficient evidence.
- Chain-of-Thought (CoT) is a Double-Edged Sword: While CoT generally improves performance, for some models (like DeepSeek-R1), it actually increased the tendency to over-update, leading to a "negative filtering gain."
Key Results
The researchers evaluated 18 state-of-the-art models, including GPT-4o, Llama-3.1-405B, and DeepSeek-V3.
- Human vs. Model Gap: Humans achieved ~94% accuracy on clean detection tasks. The best-performing models, like MiniMax-M2.5 (86.05%) and GPT-OSS-120B (82.53%), still lag significantly behind, especially in multiple-choice identification where scores drop to the 60-70% range.
- The Salience Gap: As shown in Figure 3, there is a massive gap (up to 74% for Phi-3-Medium) between a model's ability to detect a correct change and its ability to reject a "hard negative" (salient but insufficient) cue.
- Noisy Settings: Adding a distractor cue (Task 3 & 4) proved to be a major stress test. While some models like GPT-3.5-Turbo actually saw an accuracy increase (likely due to a shift in response bias), others like DeepSeek-V3.1 saw performance plummet from 82.57% to 55.97%.
Limitations and Future Work
- Textual Proxy for Senses: The benchmark uses textual descriptions of sensory experiences rather than raw multi-modal inputs (video/audio). Future work could extend this to true multi-modal environments.
- Categorical vs. Probabilistic: MuSeToM treats belief revision as a binary "change/no-change" event. In reality, humans often update their confidence in a belief rather than replacing it entirely.
- Static Narratives: The stories are fixed. Testing belief revision in dynamic, interactive multi-agent simulations would be a natural next step to see how these biases play out in real-time social coordination.
Summary updated