Imagine reading a mystery novel where you, the reader, know exactly who the killer is. If someone asks you, "Where will the detective look for clues next?", you must ignore your own knowledge and think only about what the detective knows. If you fail to separate your comprehensive understanding from the detective's limited perspective, you fall victim to the curse of knowledge.
Large Language Models (LLMs) face exactly this problem when tasked with Theory of Mind (ToM)—the ability to reason about others' mental states, desires, and beliefs. Because an entire narrative or scene is provided within their input context, models default to an omniscient "god's eye view." They struggle to inhibit their own privileged knowledge of the ground truth when predicting the actions or beliefs of an uninformed character.
The Anatomy of Knowledge Leakage
Why do models fail to maintain distinct perspectives? It is not because they lack the capacity to represent different viewpoints entirely. In fact, Zhu demonstrated in Language Models Represent Beliefs of Self and Others that LLMs naturally form distinct internal neural representations for their own knowledge versus the potentially false beliefs of characters.
However, these internal representations are highly fragile and easily disrupted by how a question is framed, as shown by Bortoletto et al. in Brittle Minds, Fixable Activations: Understanding Belief Representations in Language Models. During text generation, the global ground truth acts as a powerful cognitive attractor. The model lacks inhibitory control—the capacity to actively suppress known facts when simulating an uninformed perspective (Jung et al.).
This deficit leads to several distinct failure modes:
- Knowledge Leakage: Privileged information known to the model but hidden from a character inadvertently bleeds into the model's predictions of that character's beliefs (Sarangi et al.).
- Persona Overriding: Even when explicitly prompted to simulate a specific human persona with limited knowledge, the model's internal factual knowledge frequently overrides those constraints, causing it to output facts the human subject did not actually possess (Toubia et al.).
- Asymmetry Neglect: Models struggle to track multi-perspective reasoning, often defaulting to tracking only a primary agent while failing to isolate what secondary characters know (Sap et al.).
Exposing the Illusion of Social Intelligence
Early benchmarks suggested that scaling up models might naturally yield robust social reasoning. However, recent work reveals that this apparent success was largely an artifact of omniscient evaluation settings where the model's perspective and the character's perspective happened to align.
As Zhou et al. argue in Is this the real life? Is this just fantasy? The Misleading Success of Simulating Social Interactions With LLMs, when models are forced into realistic, information-asymmetric roles where they lack access to the private goals of their interlocutors, their performance plummets.
To expose this omniscient observer bias, researchers have introduced benchmarks that explicitly stress-test information asymmetry:
[Global Input Context / Model's View]
Story: Alice puts the book on the table. Bob moves it to the drawer while Alice is away.
-----------------------------------------------------------------------------------------
[Omniscient Default Bias] --> Model predicts: "Alice will look in the drawer." (Fails ToM)
[Perspective Isolation] --> Model predicts: "Alice will look on the table." (Passes ToM)
- BigToM (Gandhi et al.) uses procedurally generated social scenarios to demonstrate that models systematically fail when characters hold false beliefs, defaulting instead to objective reality.
- FANToM (Kim et al.) tests multi-party conversations where characters drop in and out, showing that models incorrectly attribute the full transcript history to characters who were absent for parts of the discussion.
- MMToM-QA (Jin et al.) extends this challenge to multimodal settings, proving that vision-language models adopt a "god's eye view," conflating their complete view of a video scene with an agent's restricted physical perspective.
- DECRYPTO (Lupu et al.) evaluates strategic communication, revealing that models commit critical errors by assuming competitors have access to the model's own private information.
- ATOMS (Ma et al.) transitions evaluations into a situated 2D grid world, highlighting how models fail to separate global prompt information from an agent's limited field of view.
Mechanisms for Information Isolation
To overcome the omniscient observer bias, researchers have developed inference-time interventions designed to enforce strict boundaries between global knowledge and local perspectives. These solutions fall into three main paradigms: context filtering, task decomposition, and representation steering.
Context Filtering
The most direct way to prevent an LLM from using privileged information is to strip it away before the final inference step. Frameworks like SIMTOM (Wilf et al.) and PercepToM (Jung et al.) implement a two-stage prompting mechanism.
First, the model is prompted solely to identify what a specific character witnessed or perceived. Second, the prompt cuts off the rest of the story, forcing the model to reason strictly within that restricted "perspective context." By removing the privileged information from the prompt entirely, the curse of knowledge is bypassed.
Task Decomposition and State Tracking
Instead of relying on a single end-to-end generation, some approaches break down complex, higher-order reasoning ("What does Alice think Bob knows?") into explicit, manageable steps.
In Decompose-ToM: Enhancing Theory of Mind Reasoning in Large Language Models through Simulation and Task Decomposition, Sarangi et al. break down social reasoning into simpler factual sub-tasks to prevent knowledge leakage.
Similarly, TimeToM (Hou et al.) constructs explicit "Temporal Belief State Chains" for each character, tracking how information asymmetries evolve step-by-step over time rather than attempting to resolve the entire narrative state at once.
Neuro-Symbolic Isolation and Activation Steering
Other methods enforce perspective boundaries architecturally or representationally to ensure standard prompting techniques do not break down (Xu et al.).
For example, SYMBOLICTOM (Sclar et al.) maintains explicit, recursive symbolic graphs of multi-character beliefs at decoding time, preventing global knowledge from bleeding into local character nodes. EnigmaToM (Xu et al.) utilizes a neural knowledge base of entity states paired with iterative masking to isolate viewpoints.
Finally, because the representations of self and other beliefs are present but weak within the model's activations, researchers can bypass prompting entirely and manipulate the internal states directly. Both Zhu and Bortoletto et al. demonstrate that techniques like Contrastive Activation Addition (CAA) can steer internal activations during generation. By artificially amplifying the neural circuits representing the "other's belief," they reinforce the character's perspective and successfully override the model's egocentric bias.
Key Foundational Work
Towards A Holistic Landscape of Situated Theory of Mind in Large Language Models (2023) — contributed 2 concept elements to this direction.
Minding Language Models' (Lack of) Theory of Mind: A Plug-and-Play Multi-Character Belief Tracker (2023) — contributed 1 concept elements to this direction.
Understanding Social Reasoning in Language Models with Language Models (2023) — contributed 1 concept elements to this direction.