SimToM
Simulating Theory of Mind means explicitly modeling what different people know and believe — filtering the context to each individual’s perspective so the model answers perspective-dependent questions correctly, rather than reasoning from everything it knows.
Introduced: SimToM was published in 2023 by Wilf et al. The technique addresses a fundamental limitation in LLMs: they see all information simultaneously and struggle to reason from limited perspectives — a problem known as the “curse of knowledge.” SimToM introduces a two-step approach: first, filter the context to only what a specific person knows, then answer the question from that filtered perspective alone. By forcing the model to explicitly separate “what I know” from “what this person knows,” SimToM transforms perspective-taking from a hidden inference into a structured, auditable process.
Modern LLM Status: Theory of Mind remains challenging for LLMs. While models have improved at basic perspective-taking tasks, they still default to omniscient reasoning when faced with complex multi-agent scenarios. SimToM’s explicit perspective-filtering consistently outperforms direct prompting on perspective-dependent questions. The technique is most valuable when scenarios involve multiple agents with different knowledge states, hidden information, or false beliefs — situations where even advanced models struggle without explicit guidance to separate what each person knows.
Think From Their Shoes, Not Yours
LLMs suffer from an omniscience problem. When given a scenario describing multiple people with different knowledge, the model sees everything at once — every detail, every hidden action, every private conversation. It cannot naturally “forget” what it has read. This means when asked “What does Alice think?”, the model often answers based on what it knows rather than what Alice knows, producing confidently wrong answers to perspective-dependent questions.
SimToM forces explicit perspective separation. The process works in two deliberate steps: first, list only the facts that the target person would know based on the scenario; second, answer the question using only that filtered information. This separation prevents information leakage — the model cannot accidentally use knowledge that the target person does not have, because that knowledge has been removed from the working context entirely.
Think of it like a detective who interviews witnesses separately. Each witness only saw part of the event. To understand what Witness A believes happened, the detective does not consult Witness B’s statement — they reconstruct the event purely from Witness A’s vantage point. SimToM applies this same principle to language models: reconstruct the world from each person’s limited view before reasoning about their beliefs.
Telling a model to “pretend you don’t know X” rarely works reliably. The information is still in the context, and the model’s attention mechanism will attend to it regardless of the instruction. SimToM solves this structurally: by explicitly extracting only the relevant person’s knowledge into a separate, reduced context, the forbidden information is physically absent from the reasoning step. You cannot leak what is not there. This is why perspective filtering consistently outperforms “imagine you are” or “pretend you only know” instructions.
The SimToM Process
Four stages from full scenario to perspective-accurate answer
Read the Full Context
Absorb the complete scenario with all information about all parties. This includes every action taken, every statement made, and every event that occurred — regardless of who witnessed it. At this stage, you have the omniscient view of the entire situation.
“Alice put a toy in the blue box. Alice left the room. Bob entered and moved the toy from the blue box to the red drawer. Bob left. Alice returned.”
Filter to Target Person’s Perspective
Extract only the facts that the target person would know, creating a reduced context. Go through each event in the scenario and ask: “Was the target person present for this? Could they have observed this?” Include only the events they witnessed and discard everything else. This creates a filtered world model specific to that individual.
Alice’s filtered context: “Alice put a toy in the blue box. Alice left the room. Alice returned.” — Alice was not present when Bob moved the toy, so that event is excluded from her perspective entirely.
Answer from Filtered Perspective
Using only the filtered context, answer the question as if you only knew what that person knows. Do not reference, hint at, or incorporate any information from the full scenario that was excluded in the filtering step. The filtered context is your entire world for this answer.
Question: Where does Alice think the toy is?
Answer from filtered context: Alice thinks the toy is in the blue box, because that is where she placed it and she has no information suggesting it was moved.
Verify Perspective Integrity
Check that the answer does not leak information the target person could not know. Review the response for any details, implications, or hedging that reveals awareness of events outside the filtered context. If the answer references anything the person was not present for, the perspective filter has been violated and the answer needs revision.
Leak check: Does the answer mention the red drawer, Bob’s actions, or that the toy was moved? If yes, the perspective filter failed. The correct answer should reference only the blue box with no awareness of any changes made in Alice’s absence.
See the Difference
Why explicit perspective filtering prevents omniscient reasoning errors
Direct Prompt
Alice put a toy in the box and left the room. Bob entered and moved the toy to the drawer. Where does Alice think the toy is?
The toy is in the drawer, since Bob moved it there. Alice would likely check the drawer.
SimToM
Step 1 — Filter to Alice’s perspective: Alice put a toy in the box. Alice left the room. Alice has no information about events during her absence.
Step 2 — Answer from Alice’s filtered context: Based only on what Alice knows, the toy is in the box where she placed it.
Alice thinks the toy is in the box. She placed it there before leaving and was not present when it was moved, so she has no reason to believe it is anywhere else.
Natural Language Works Too
While structured frameworks and contextual labels are powerful tools, LLMs are exceptionally good at understanding natural language. As long as your prompt contains the actual contextual information needed to create, answer, or deliver the response you’re looking for — the who, what, why, and constraints — the AI can produce complete and accurate results whether you use a formal framework or plain conversational language. But even in 2026, with the best prompts, verifying AI output is always a necessary step.
SimToM in Action
See how perspective filtering solves Theory of Mind challenges
“Sally places her marble in the basket and leaves the room. Anne takes the marble from the basket and puts it in her own box. Sally returns. Where will Sally look for her marble?”
Step 1 — Filter to Sally’s perspective:
Sally placed her marble in the basket. Sally left the room. Sally returned to the room. Sally does not know about any events that occurred while she was away.
Step 2 — Answer from Sally’s filtered context:
Sally will look for her marble in the basket. From her perspective, that is the last place she put it, and she has no information suggesting it was moved.
Verification: The answer does not reference Anne’s box or Anne’s action of moving the marble. Perspective integrity is maintained.
“The CEO informed the VP of Operations that layoffs will happen next month. The VP has not told anyone on the team. Employee Dana asks the VP for a long-term project assignment. What does Dana believe about job security at the company?”
Step 1 — Filter to Dana’s perspective:
Dana works at the company. Dana is asking for a long-term project assignment. Dana has not been informed of any upcoming layoffs or organizational changes. Dana has no knowledge of private conversations between the CEO and the VP.
Step 2 — Answer from Dana’s filtered context:
Dana believes job security at the company is stable. Her request for a long-term project indicates she expects to remain employed and sees no reason for concern. She has no information about upcoming layoffs.
Verification: The answer does not reference the CEO’s announcement, the planned layoffs, or the VP’s knowledge. Dana’s perspective is preserved without information leakage.
“Maya is planning a surprise birthday party for Leo. She told Priya and Kai about the plan. Kai accidentally mentioned ‘Saturday plans’ to Leo but did not reveal the party. Priya bought a cake and hid it at her house. Leo asked Maya what she is doing on Saturday, and Maya said she is free. What does each person know about Saturday?”
Filter to Leo’s perspective:
Kai mentioned “Saturday plans” but gave no details. Leo asked Maya about Saturday, and Maya said she is free. Leo does not know about any party, cake, or coordination among his friends.
Leo believes: Saturday is an ordinary day. Maya is free. Kai has some unspecified Saturday plans.
Filter to Maya’s perspective:
Maya organized the surprise party. She told Priya and Kai. She told Leo she is free on Saturday to maintain the surprise.
Maya knows: The party is happening. She successfully deflected Leo’s question. She does not know Kai mentioned Saturday plans to Leo.
Filter to Priya’s perspective:
Maya told Priya about the party. Priya bought a cake and hid it at her house.
Priya knows: The party is happening and she has the cake. She does not know about Kai’s slip or Leo’s conversation with Maya.
Verification: Each person’s knowledge is strictly limited to the events they participated in or were told about directly. No perspective leaks information from another person’s view.
When to Use SimToM
Best for scenarios where who knows what matters for the correct answer
Perfect For
Any question where the correct answer depends on what a specific person knows or believes, rather than the objective truth of the situation.
Classic Theory of Mind tests where one person holds an outdated or incorrect belief because they missed a critical event — the Sally-Anne test and its variants.
Simulating negotiations where each party has private information — understanding what each side knows prevents unrealistic assumptions about shared knowledge.
Situations requiring strict information boundaries — ensuring responses do not reveal information that a specific user, role, or agent should not have access to.
Skip It When
When everyone in the scenario has access to the same facts, perspective filtering adds unnecessary complexity — there is no information asymmetry to manage.
If the question asks for objective facts rather than someone’s belief or expectation, perspective filtering is unnecessary — the answer is the same regardless of viewpoint.
Questions with one subject and no perspective-dependent component — “What is the capital of France?” has no hidden perspective to filter.
Use Cases
Where SimToM delivers the most value
Social Reasoning
Model interpersonal dynamics by tracking what each person knows, believes, and expects — essential for understanding misunderstandings, secrets, and social conflicts in narrative analysis.
Game Theory & Strategy
Analyze strategic interactions where each player has private information. Filter each player’s knowledge state to predict rational decisions based on what they actually know, not the full game state.
Privacy Compliance
Ensure AI responses respect information boundaries by filtering context to what each user role is authorized to know — preventing accidental disclosure of restricted data in multi-user systems.
Customer Intent Modeling
Understand what a customer knows and does not know when they contact support. Filter context to the customer’s visible experience to accurately interpret their questions and frustrations.
Negotiation Prep
Prepare for negotiations by modeling each party’s information state separately. Understand what the other side knows, what they think you know, and where information gaps create leverage or risk.
Educational Assessment
Model what a student knows versus what the curriculum assumes they know. Filter to the student’s demonstrated knowledge to identify genuine gaps rather than assuming omniscient understanding.
Where SimToM Fits
SimToM bridges simple role adoption and full multi-agent perspective modeling
SimToM’s perspective filtering pairs powerfully with Chain-of-Thought reasoning. After filtering context to a specific person’s knowledge, use CoT to reason step-by-step through their decision-making process. This combination gives you both accurate knowledge boundaries and transparent reasoning within those boundaries — producing perspective-aware answers with full reasoning visibility.
Related Techniques
Explore complementary perspective and reasoning techniques
Reason From Their Perspective
Try SimToM perspective filtering on your own scenarios or build perspective-aware prompts with our tools.