Perspective Filtering for Theory of Mind Reasoning

This direction is merged into another direction: higher-order theory of mind. Its last version stays readable and citable here.

Version 1 of 115 papers26 conceptsfrozen 2026-10-04

Large Language Models (LLMs) are effectively omniscient readers. When they process a story, they have access to every event, dialogue, and hidden action across the entire narrative. This "all-knowing" perspective, however, is a liability when solving Theory of Mind (ToM) tasks—the ability to reason about the mental states of others.

If a model knows that a ball was moved from a basket to a box while Sally was out of the room, it struggles to remember that Sally still thinks the ball is in the basket. This "reality bias" or "curse of knowledge" causes models to conflate their own global knowledge with a character’s limited perspective.

The Curse of Knowledge as a Context Problem

The core failure in social reasoning is not necessarily a lack of logic, but an inability to ignore privileged information. As Alex Wilf and colleagues argue in SIMTOM, the primary obstacle is identifying what information is known to a specific agent. When models are shown a full story, the ground-truth facts "pollute" their reasoning about a character's false beliefs.

Research across several frameworks—including PercepToM by C. G. Jung et al. and TIMETOM by Guiyang Hou et al.—suggests a shared solution: Perspective Filtering. Instead of asking a model to "imagine" what a character knows, we should physically restrict the model's input to only the events that character witnessed.

Transforming False Beliefs into True Beliefs

The intuition behind perspective filtering is a reduction in task complexity. A "false-belief" task (reasoning about someone who is wrong) is cognitively harder than a "true-belief" task (reasoning about someone who is right).

By filtering the narrative context to include only a specific character's observations, we transform the problem:

  1. Stage 1 (Filtering): The system identifies and extracts only the sentences or events witnessed by Character A.
  2. Stage 2 (Inference): The model is given this "perspective context" and asked to predict Character A's belief.

In this restricted view, Character A’s belief is no longer "false"—it is the only "truth" available in the context window. As noted in UniToMBench, this bypasses the model's inability to ignore its own knowledge, allowing it to solve complex benchmarks that previously required massive neural scaling.

Decomposing Social Complexity

Social reasoning often involves nested beliefs (e.g., "Alice thinks that Bob believes..."). This higher-order reasoning is notoriously difficult for LLMs. Sneheel Sarangi and the creators of Decompose-ToM address this by recursively simulating perspectives.

By breaking a complex social question into simpler factual sub-tasks—"What did Alice see?" and "What did Bob see in Alice's presence?"—the system treats perspective-taking as a data-filtering task rather than a generative reasoning task. Hainiu Xu et al. demonstrate in EnigmaToM that this iterative masking allows smaller models to outperform much larger ones by focusing on "entity states" rather than raw text.

Symbolic and Temporal Tracking

While simple prompting can filter context, more robust systems maintain explicit "belief graphs." Melanie Sclar and the team behind SYMBOLICTOM use a "plug-and-play" tracker that builds recursive symbolic representations of what each character knows.

This approach is further refined by modeling time. In EvolvTrip, researchers use temporal knowledge graphs to track how mental states evolve. By modeling beliefs as structured triples that change over time, the system ensures that a character's perspective is always grounded in their specific chronological experience, preventing the "leakage" of future events into a character's current mental state.

Why It Works

The success of these methods reveals a fundamental insight into LLM behavior: Models are better at selecting relevant information than they are at suppressing irrelevant information.

When we provide a global context and ask a model to "ignore the fact that the ball moved," we are asking it to perform a high-effort inhibitory task. When we filter the context so the model never sees the ball move, we align the task with the model's core strength—pattern matching within a provided context. By treating Theory of Mind as a context management problem rather than a pure reasoning problem, these frameworks unlock human-level social intelligence in existing language models.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_perspective_filtering_for_theory_of_mind_61d033a1,
  title = {Perspective Filtering for Theory of Mind Reasoning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/perspective-filtering-for-theory-of-mind-reasoning/rdr_61d033a147234293bfc38d2f0d901fe5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_perspective_filtering_for_theory_of_mind_61d033a1_v1,
  title = {Perspective Filtering for Theory of Mind Reasoning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/perspective-filtering-for-theory-of-mind-reasoning/rdr_61d033a147234293bfc38d2f0d901fe5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}