Technical Summary: Reasoning Promotes Robustness in Theory of Mind Tasks

Problem Statement

Large Language Models (LLMs) have demonstrated strong performance on Theory of Mind (ToM) tests, sparking debate regarding whether this indicates genuine reasoning capabilities or merely pattern matching. While early models showed promise, subsequent criticism highlighted their fragility; minor prompt perturbations often caused failures, suggesting a lack of "real" understanding. Recently, a new class of "reasoning models" (e.g., GPT-5, Claude, DeepSeek R1) trained via Reinforcement Learning with Verifiable Rewards (RLVR) has emerged. These models are designed to generate complex chains of thought (CoT) before answering. The central problem addressed by this paper is determining whether these reasoning models possess fundamentally new ToM capabilities or if their improved performance stems from increased robustness in finding correct solutions under varying conditions.

Methodology

The study employs a mixed-method approach combining original psychological experiments with an analysis of existing benchmark data.

1. Experimental Setup

The authors conducted experiments on several state-of-the-art models, including GPT-5, Claude (with and without thinking enabled), DeepSeek R1, and Grok-3-mini. Key experimental controls included:

• Prompting Strategies: The study utilized established psychological tests from the [31] battery and newly designed modifications based on [30] to test robustness.
• Model Variations: A critical component was the ability to toggle "thinking" on and off for the Claude model, allowing for a direct comparison of reasoning vs. non-reasoning performance within the same architecture.
• Scoring: Performance was evaluated on a 0-2 scale (normalized to [0, 1]), accounting for both the correctness of the final answer and the quality of the reasoning trace.

2. Test Batteries

• Sally-Anne Tests (1st and 2nd Order): Standard false-belief tests to assess basic and recursive mental state attribution.
• Strange Stories: Seven categories (lies, pretend play, jokes, white lies, misunderstanding, sarcasm, double bluff) requiring inference of intentions.
• Imposing Memory: Tests involving multiple levels of recursion in mental states (e.g., A believes B thinks C wants D).
• Modifications on Simple ToM Tasks: Novel tasks designed to disrupt standard semantic heuristics, specifically testing "Transparent Access" and "Relationship Change" scenarios where visual mental depiction is required.

3. Benchmark Analysis

The authors analyzed results from [10] covering benchmarks such as FANToM, BigToM, MMToM-QA, and ParaphrasedToMi. These datasets were used to corroborate findings across a broader range of models and tasks, including distilled and non-distilled reasoning models.

Key Contributions

• Empirical Evidence of Robustness: The paper provides empirical data showing that reasoning models consistently exhibit higher robustness to prompt variations and task perturbations compared to non-reasoning counterparts.
• Reinterpretation of "Real" ToM: The authors challenge the view that prompt sensitivity equates to a lack of ToM skills. Instead, they argue that the upper bound for success exists in the base models, but reasoning training (RLVR) stabilizes the inference path, making the model less susceptible to distractors.
• Emergent Perspective Taking: The analysis of reasoning traces reveals that reasoning models spontaneously employ "perspective-taking" strategies (filtering context to character knowledge) without explicit prompting, a technique previously required as a manual prompt engineering intervention.
• Identification of Visual Limitations: The study identifies specific failure modes where models struggle with tasks requiring a "mental visual depiction" of a scene (e.g., transparent containers), suggesting that current reasoning improvements are primarily semantic/logical rather than spatial-visual.

Results

• Psychological Tests:
  • Sally-Anne: Models achieved near-perfect scores (1.00) on 1st and 2nd-order tests. The only notable error involved a model merging character knowledge (Grok-3-mini).
  • Strange Stories: Performance was flawless in five of seven categories. Minor errors occurred in "sarcasm" and "double bluff," often where models failed to visualize the physical state of a scene (e.g., a messy kitchen hidden from view).
  • Imposing Memory: Claude (with thinking) outperformed its non-thinking counterpart (100% vs. 92%). Errors were primarily due to hallucination or failing to distinguish between a character's belief and reality.
  • Modifications: Current models outperformed GPT-3 significantly. However, two specific tasks (2A and 2B) remained difficult for all models, likely due to the necessity of visualizing spatial relationships that are trivial to humans but hard to derive semantically.
• Benchmark Results:
  • Reasoning models significantly outperformed non-reasoning models across FANToM, BigToM, and ParaphrasedToMi.
  • The discrepancy was most pronounced in FANToM, a benchmark requiring consistency across multiple questions, highlighting the stability gains of reasoning models.
  • Distilled models showed significantly worse performance in FANToM compared to other reasoning models, though the authors note this requires further investigation.

Significance and Claims

The paper concludes that the primary advantage of RLVR-trained reasoning models in ToM tasks is increased robustness, not the emergence of fundamentally new reasoning capabilities.

• Stability over Capacity: The gains are attributed to the model's ability to find the correct solution path more reliably when faced with perturbations, rather than an expansion of representational capacity. The reasoning paths are bounded by the base model's capabilities but are accessed more consistently.
• Re-evaluating Fragility: The authors argue that previous claims of LLMs lacking "real" ToM skills due to prompt sensitivity may be overstated. The models likely possessed the necessary skills, but the reasoning training provides the stability to access them reliably.
• Implications for Evaluation: The findings suggest that evaluating social-cognitive behavior in LLMs requires testing for robustness against prompt variations, not just accuracy on static benchmarks. The "power" of RLVR training lies in narrowing the reasoning coverage to stable, high-probability paths, which is desirable for robust application.

The authors remain modest, acknowledging that they cannot definitively rule out spurious correlations or meta-knowledge heuristics in all cases, and that the specific contribution of RLVR versus other inference-time scaling techniques requires further comparative study.