LaTeX editing, your library, and agents that actually read the literature — in one window. This is the real interface: click around.
perspective-taking-paper
Share
File Tree
figs
fig1.png
fig2.png
literati.bib
main.tex
neurips_2026.sty
File Outline
§Introduction8
§Background14
Papers 4
Wilf et al. (2023) · arXiv:2311.10227Think Twice: Perspective-Taking Improves Large Language Models’ Theory-of-Mind Capabilities
Sclar et al. (2023) · arXiv:2306.00924Minding Language Models' (Lack of) Theory of Mind
Zhou et al. (2023) · arXiv:2310.03051How FaR Are Large Language Models From Agents with Theory-of-Mind?
Kosinski (2023) · arXiv:2302.02083Evaluating Large Language Models in Theory of Mind Tasks
Ask anything…
chat (shift+tab to cycle)esc to interrupt
Aaab.*
Type to search main.tex
Changes
Message (⌘↵ to commit)
Staged Changes1
results.texsections/A
Working Changes2
main.texM
abstract.texsections/M
Nothing to commit — working tree clean
History
›Draft intro + method2hRevert
›Initial skeleton1dRevert
main.tex×
Editing ⌄
8\section{Introduction}
9
10Large language models increasingly exhibit theory-of-mind
11(ToM) capabilities, yet their accuracy on false-belief
12tasks remains inconsistent across benchmarks.
13
14\section{Background}
15
16
17
literati.bib is generated from your library — read-only
1/ 1
Perspective-Taking Improves LLM Theory-of-Mind
Anonymous Author(s) Affiliation Address email
Abstract
1 We study theory-of-mind in language models and propose a method.
2 Across five benchmarks it yields consistent gains.
1 Introduction
3 Large language models increasingly exhibit theory-of-mind (ToM)
4 capabilities, yet their accuracy on false-belief tasks remains
5 inconsistent across benchmarks.
Submitted to 40th Conference on Neural Information Processing Systems (NeurIPS 2026). Do not distribute.
All logsErrorsWarningsInfo
Compilation failed. See errors below.
click anything — Compile, the logs icon, the files (figs, literati.bib, the .sty), your papers, source control, find & replace
10Large language models increasingly exhibit theory-of-mind
11(ToM) capabilities, yet their accuracy on false-belief
12-tasks remains inconsistent.
12+tasks remains inconsistent across benchmarks. We show
13+that a simple perspective-taking prompt closes much of
14+this gap \citep{wilf2023think}.
15
37We evaluate on the ToMI benchmark.
38-Our method improves accuracy.
38+Our method improves accuracy by 47\% on ToMI
39+and 14\% on BigToM over the GPT-3.5 baseline
40+\citep{wilf2023think}.
41We report mean over 5 seeds.
42-% TODO: add ablation table
42+Ablations in Table 2 isolate each component.
1We study theory-of-mind in language models and
2-propose a method.
2+propose Think-Twice, a perspective-taking prompt that
3+reframes each question from the target agent's view.
4Across five benchmarks it yields consistent gains.
Think Twice: Perspective-Taking Improves Large Language Models’ Theory-of-Mind Capabilities
Alex Wilf · Sihyun Shawn Lee · Paul Pu Liang · Louis-Philippe Morency
Carnegie Mellon University · arXiv:2311.10227
AbstractHuman interactions are deeply rooted in the interplay of thoughts, beliefs, and desires made possible by Theory of Mind (ToM). Although ToM may come naturally to us, emulating it presents a challenge to even the most advanced Large Language Models. We turn to the cognitive science theory "Simulation Theory" and introduce SimToM, a two-stage prompting framework based on perspective-taking: first filter context based on what the character in question knows, then answer a question about their mental state. Requiring no additional training and minimal prompt-tuning, SimToM shows substantial improvement over existing methods.
1 Introduction
Large language models increasingly succeed on tasks once thought to require social reasoning, yet their performance on false-belief tasks remains brittle. We hypothesize that much of this brittleness stems from a failure to anchor reasoning to a specific agent's knowledge state, rather than from a lack of underlying capability.
We test this hypothesis with a lightweight intervention that asks the model to restate what each character can and cannot observe before answering. The result is a large, consistent improvement that requires no additional parameters.
41Base
67CoT
88Ours
95Human
Figure 1. Accuracy on ToMI (%). SimToM (Ours) closes most of the gap to human performance.
2 Method
Given a scenario and a question about agent a, we prompt the model to enumerate the events a witnessed, mark events hidden from a, and only then resolve the query conditioned on a's partial observation. This decomposition mirrors the sub-steps humans perform implicitly.
[1] Sclar et al. Minding Language Models' (Lack of) Theory of Mind. 2023.
[2] Zhou et al. How FaR Are Large Language Models From Agents with Theory-of-Mind? 2023.
[3] Kosinski. Evaluating Large Language Models in Theory of Mind Tasks. 2023.
Minding Language Models' (Lack of) Theory of Mind: A Plug-and-Play Multi-Character Belief Tracker
Melanie Sclar · Sachin Kumar · Peter West · Alane Suhr · Yejin Choi · Yulia Tsvetkov
University of Washington · Allen Institute for AI · arXiv:2306.00924
AbstractTheory of mind — tracking what others believe — is a hallmark of social intelligence, but language models struggle to maintain consistent beliefs for multiple characters. We present SymbolicToM, a plug-and-play inference-time tracker that maintains an explicit belief graph per character. It substantially improves robustness on ToMI, especially under adversarial reorderings that break end-to-end models.
1 Motivation
End-to-end prompting conflates the beliefs of different characters, causing errors that compound as stories grow longer. We argue that an explicit, symbolic representation of "who knows what" is both interpretable and more robust than relying on the model's latent state.
49Vanilla
58CoT
93SymbolicToM
Figure 1. Robust accuracy under story reordering. The symbolic tracker degrades far less.
2 Belief Graphs
For each character we instantiate a graph whose nodes are propositions and whose edges encode observation events. Queries are answered by reading off the relevant character's sub-graph, sidestepping the interference that plagues monolithic prompting.
[1] Nematzadeh et al. Evaluating Theory of Mind in Question Answering. 2018.
[2] Le et al. Revisiting the Evaluation of Theory of Mind through Question Answering. 2019.
How FaR Are Large Language Models From Agents with Theory-of-Mind?
Pei Zhou · Aman Madaan · Srividya Pranavi Potharaju · Aditya Gupta · Kevin R. McKee · Ari Holtzman · Jay Pujara · Xiang Ren · Swaroop Mishra · Aida Nematzadeh · Shyam Upadhyay · Manaal Faruqui
University of Southern California · Google DeepMind · arXiv:2310.03051
Abstract"Thinking is for Doing." Humans infer other people’s mental states and act pragmatically on those inferences. Benchmarks such as ToMi ask models about characters’ beliefs, but do not test whether models can use those inferences to guide their actions. We propose Thinking for Doing (T4D), which requires connecting inferences about others’ mental states to actions in social scenarios, and introduce Foresee and Reflect (FaR), a zero-shot prompting framework that encourages LLMs to anticipate future challenges and reason about potential actions.
1 Thinking is for Doing
LLMs such as GPT-4 seemingly excel at tracking characters’ beliefs in stories, yet struggle to translate this capability into strategic action. The core challenge lies in identifying the implicit inferences about mental states — without being explicitly asked, as in ToMi — that lead to choosing the correct action.
50GPT-4
71+ FaR
Figure 1. Accuracy on T4D (%). FaR boosts GPT-4 from 50% to 71%, outperforming Chain-of-Thought and Self-Ask.
2 Foresee and Reflect
The foresee step prompts the model to anticipate likely future events and the challenges facing each character; the reflect step reasons over candidate actions against those futures before committing. FaR generalizes to diverse out-of-distribution story structures and scenarios.
[1] Yao et al. ReAct: Synergizing Reasoning and Acting in Language Models. 2022.
[2] Shinn et al. Reflexion: Language Agents with Verbal Reinforcement Learning. 2023.
1% literati.bib — generated by Literati from your library
2% papers are added via the arXiv API · do not edit by hand
3
4@article{wilf2023think,
5 title = {Think Twice: Perspective-Taking Improves Large Language
6Models' Theory-of-Mind Capabilities},
7 author = {Wilf, Alex and Lee, Sihyun Shawn and Liang, Paul Pu
8and Morency, Louis-Philippe},
9 year = {2023},
10 eprint = {2311.10227},
11}
12
13@article{sclar2023minding,
14 title = {Minding Language Models' (Lack of) Theory of Mind:
16 author = {Sclar, Melanie and Kumar, Sachin and West, Peter and
17Suhr, Alane and Choi, Yejin and Tsvetkov, Yulia},
18 year = {2023},
19 eprint = {2306.00924},
20}
21
22@article{zhou2023far,
23 title = {How FaR Are Large Language Models From Agents
24with Theory-of-Mind?},
25 author = {Zhou, Pei and Madaan, Aman and Potharaju, Srividya Pranavi
26and Gupta, Aditya and McKee, Kevin R. and Holtzman, Ari
27and Pujara, Jay and Ren, Xiang and Mishra, Swaroop
28and Nematzadeh, Aida and Upadhyay, Shyam and Faruqui, Manaal},
29 year = {2023},
30 eprint = {2310.03051},
31}
32
33@article{kosinski2023theory,
34 title = {Evaluating Large Language Models in Theory of Mind Tasks},
35 author = {Kosinski, Michal},
36 year = {2023},
37 eprint = {2302.02083},
38}
39
1% neurips_2026.sty — NeurIPS 2026 official style
2% do not modify — conference formatting
3\NeedsTeXFormat{LaTeX2e}
4\ProvidesPackage{neurips_2026}[2026/01/15 v1.0]
5\RequirePackage{lineno}
6\RequirePackage[colorlinks=true]{hyperref}
7
8\newcommand{\@noticestring}{%
9Submitted to 40th Conference on Neural
10Information Processing Systems (NeurIPS 2026).
11Do not distribute.}
12
13\setlength{\textwidth}{5.5in}
14\setlength{\textheight}{9in}
15\widowpenalty=10000
16\clubpenalty=10000
17\flushbottom
18\sloppy
Perspective-Taking Improves LLM Theory-of-Mind
Anonymous Author(s) Affiliation Address email
Abstract
1 We study theory-of-mind in language models and propose a method.
2 Across five benchmarks it yields consistent gains.
1 Introduction
3 Large language models increasingly exhibit theory-of-mind (ToM)
4 capabilities, yet their accuracy on false-belief tasks remains
5 inconsistent across benchmarks.
2 Method
ℒ = −Σi yi log ŷi (1)
Submitted to 40th Conference on Neural Information Processing Systems (NeurIPS 2026). Do not distribute.
Evaluating Large Language Models in Theory of Mind Tasks
Michal Kosinski
Stanford University · arXiv:2302.02083
AbstractEleven LLMs were assessed using a custom-made battery of false-belief tasks — a gold standard in testing Theory of Mind in humans. Smaller and older models solved no tasks; GPT-3.5-turbo solved 20%, and GPT-4 solved 75% of the tasks, matching the performance of six-year-old children observed in past studies. We discuss whether this reflects genuine mentalizing or the statistics of language, and argue the distinction is increasingly hard to draw.
1 Tasks
We use the Sally–Anne unexpected-transfer task and the Smarties unexpected-contents task, each with carefully constructed controls to rule out shallow lexical shortcuts. Items are presented one sentence at a time to probe incremental belief updates.
Emergence of ToM-like behavior may be an unintended by-product of training on human-generated text. We caution against both over-attribution of mental states and premature dismissal, and call for adversarial benchmarks that separate the two.
[1] Wimmer & Perner. Beliefs about beliefs. 1983.
[2] Ullman. Large Language Models Fail on Trivial Alterations to Theory-of-Mind Tasks. 2023.
Gemini finds. Claude builds. Your library keeps them honest.
Verified citations
Hallucination-proof, by design.
Agents skim and read the papers in your library, then edit main.tex. New papers enter only through the arXiv API — so agents can only cite papers added that way, or by you.
perspective-taking-paper
Share
File Tree
figs
fig1.png
fig2.png
literati.bib
main.tex
neurips_2026.sty
File Outline
§Introduction8
§Background14
Papers 4
Wilf et al. (2023) · arXiv:2311.10227Think Twice: Perspective-Taking Improves Large Language Models’ Theory-of-Mind Capabilities
Sclar et al. (2023) · arXiv:2306.00924Minding Language Models' (Lack of) Theory of Mind
Zhou et al. (2023) · arXiv:2310.03051How FaR Are Large Language Models From Agents with Theory-of-Mind?
Kosinski (2023) · arXiv:2302.02083Evaluating Large Language Models in Theory of Mind Tasks
Ask anything…
chat (shift+tab to cycle)esc to interrupt
Aaab.*
Type to search main.tex
Changes
Message (⌘↵ to commit)
Staged Changes1
results.texsections/A
Working Changes2
main.texM
abstract.texsections/M
Nothing to commit — working tree clean
History
›Draft intro + method2hRevert
›Initial skeleton1dRevert
main.tex×
Editing ⌄
8\section{Introduction}
9
10Large language models increasingly exhibit theory-of-mind
11(ToM) capabilities, yet their accuracy on false-belief
12tasks remains inconsistent across benchmarks.
13
14\section{Background}
15
16
17
literati.bib is generated from your library — read-only
1/ 1
Perspective-Taking Improves LLM Theory-of-Mind
Anonymous Author(s) Affiliation Address email
Abstract
1 We study theory-of-mind in language models and propose a method.
2 Across five benchmarks it yields consistent gains.
1 Introduction
3 Large language models increasingly exhibit theory-of-mind (ToM)
4 capabilities, yet their accuracy on false-belief tasks remains
5 inconsistent across benchmarks.
Submitted to 40th Conference on Neural Information Processing Systems (NeurIPS 2026). Do not distribute.
All logsErrorsWarningsInfo
Compilation failed. See errors below.
hover to play — and click around: files, search, git · papers open in the PDF viewer
10Large language models increasingly exhibit theory-of-mind
11(ToM) capabilities, yet their accuracy on false-belief
12-tasks remains inconsistent.
12+tasks remains inconsistent across benchmarks. We show
13+that a simple perspective-taking prompt closes much of
14+this gap \citep{wilf2023think}.
15
37We evaluate on the ToMI benchmark.
38-Our method improves accuracy.
38+Our method improves accuracy by 47\% on ToMI
39+and 14\% on BigToM over the GPT-3.5 baseline
40+\citep{wilf2023think}.
41We report mean over 5 seeds.
42-% TODO: add ablation table
42+Ablations in Table 2 isolate each component.
1We study theory-of-mind in language models and
2-propose a method.
2+propose Think-Twice, a perspective-taking prompt that
3+reframes each question from the target agent's view.
4Across five benchmarks it yields consistent gains.
Think Twice: Perspective-Taking Improves Large Language Models’ Theory-of-Mind Capabilities
Alex Wilf · Sihyun Shawn Lee · Paul Pu Liang · Louis-Philippe Morency
Carnegie Mellon University · arXiv:2311.10227
AbstractHuman interactions are deeply rooted in the interplay of thoughts, beliefs, and desires made possible by Theory of Mind (ToM). Although ToM may come naturally to us, emulating it presents a challenge to even the most advanced Large Language Models. We turn to the cognitive science theory "Simulation Theory" and introduce SimToM, a two-stage prompting framework based on perspective-taking: first filter context based on what the character in question knows, then answer a question about their mental state. Requiring no additional training and minimal prompt-tuning, SimToM shows substantial improvement over existing methods.
1 Introduction
Large language models increasingly succeed on tasks once thought to require social reasoning, yet their performance on false-belief tasks remains brittle. We hypothesize that much of this brittleness stems from a failure to anchor reasoning to a specific agent's knowledge state, rather than from a lack of underlying capability.
We test this hypothesis with a lightweight intervention that asks the model to restate what each character can and cannot observe before answering. The result is a large, consistent improvement that requires no additional parameters.
41Base
67CoT
88Ours
95Human
Figure 1. Accuracy on ToMI (%). SimToM (Ours) closes most of the gap to human performance.
2 Method
Given a scenario and a question about agent a, we prompt the model to enumerate the events a witnessed, mark events hidden from a, and only then resolve the query conditioned on a's partial observation. This decomposition mirrors the sub-steps humans perform implicitly.
[1] Sclar et al. Minding Language Models' (Lack of) Theory of Mind. 2023.
[2] Zhou et al. How FaR Are Large Language Models From Agents with Theory-of-Mind? 2023.
[3] Kosinski. Evaluating Large Language Models in Theory of Mind Tasks. 2023.
Minding Language Models' (Lack of) Theory of Mind: A Plug-and-Play Multi-Character Belief Tracker
Melanie Sclar · Sachin Kumar · Peter West · Alane Suhr · Yejin Choi · Yulia Tsvetkov
University of Washington · Allen Institute for AI · arXiv:2306.00924
AbstractTheory of mind — tracking what others believe — is a hallmark of social intelligence, but language models struggle to maintain consistent beliefs for multiple characters. We present SymbolicToM, a plug-and-play inference-time tracker that maintains an explicit belief graph per character. It substantially improves robustness on ToMI, especially under adversarial reorderings that break end-to-end models.
1 Motivation
End-to-end prompting conflates the beliefs of different characters, causing errors that compound as stories grow longer. We argue that an explicit, symbolic representation of "who knows what" is both interpretable and more robust than relying on the model's latent state.
49Vanilla
58CoT
93SymbolicToM
Figure 1. Robust accuracy under story reordering. The symbolic tracker degrades far less.
2 Belief Graphs
For each character we instantiate a graph whose nodes are propositions and whose edges encode observation events. Queries are answered by reading off the relevant character's sub-graph, sidestepping the interference that plagues monolithic prompting.
[1] Nematzadeh et al. Evaluating Theory of Mind in Question Answering. 2018.
[2] Le et al. Revisiting the Evaluation of Theory of Mind through Question Answering. 2019.
How FaR Are Large Language Models From Agents with Theory-of-Mind?
Pei Zhou · Aman Madaan · Srividya Pranavi Potharaju · Aditya Gupta · Kevin R. McKee · Ari Holtzman · Jay Pujara · Xiang Ren · Swaroop Mishra · Aida Nematzadeh · Shyam Upadhyay · Manaal Faruqui
University of Southern California · Google DeepMind · arXiv:2310.03051
Abstract"Thinking is for Doing." Humans infer other people’s mental states and act pragmatically on those inferences. Benchmarks such as ToMi ask models about characters’ beliefs, but do not test whether models can use those inferences to guide their actions. We propose Thinking for Doing (T4D), which requires connecting inferences about others’ mental states to actions in social scenarios, and introduce Foresee and Reflect (FaR), a zero-shot prompting framework that encourages LLMs to anticipate future challenges and reason about potential actions.
1 Thinking is for Doing
LLMs such as GPT-4 seemingly excel at tracking characters’ beliefs in stories, yet struggle to translate this capability into strategic action. The core challenge lies in identifying the implicit inferences about mental states — without being explicitly asked, as in ToMi — that lead to choosing the correct action.
50GPT-4
71+ FaR
Figure 1. Accuracy on T4D (%). FaR boosts GPT-4 from 50% to 71%, outperforming Chain-of-Thought and Self-Ask.
2 Foresee and Reflect
The foresee step prompts the model to anticipate likely future events and the challenges facing each character; the reflect step reasons over candidate actions against those futures before committing. FaR generalizes to diverse out-of-distribution story structures and scenarios.
[1] Yao et al. ReAct: Synergizing Reasoning and Acting in Language Models. 2022.
[2] Shinn et al. Reflexion: Language Agents with Verbal Reinforcement Learning. 2023.
1% literati.bib — generated by Literati from your library
2% papers are added via the arXiv API · do not edit by hand
3
4@article{wilf2023think,
5 title = {Think Twice: Perspective-Taking Improves Large Language
6Models' Theory-of-Mind Capabilities},
7 author = {Wilf, Alex and Lee, Sihyun Shawn and Liang, Paul Pu
8and Morency, Louis-Philippe},
9 year = {2023},
10 eprint = {2311.10227},
11}
12
13@article{sclar2023minding,
14 title = {Minding Language Models' (Lack of) Theory of Mind:
16 author = {Sclar, Melanie and Kumar, Sachin and West, Peter and
17Suhr, Alane and Choi, Yejin and Tsvetkov, Yulia},
18 year = {2023},
19 eprint = {2306.00924},
20}
21
22@article{zhou2023far,
23 title = {How FaR Are Large Language Models From Agents
24with Theory-of-Mind?},
25 author = {Zhou, Pei and Madaan, Aman and Potharaju, Srividya Pranavi
26and Gupta, Aditya and McKee, Kevin R. and Holtzman, Ari
27and Pujara, Jay and Ren, Xiang and Mishra, Swaroop
28and Nematzadeh, Aida and Upadhyay, Shyam and Faruqui, Manaal},
29 year = {2023},
30 eprint = {2310.03051},
31}
32
33@article{kosinski2023theory,
34 title = {Evaluating Large Language Models in Theory of Mind Tasks},
35 author = {Kosinski, Michal},
36 year = {2023},
37 eprint = {2302.02083},
38}
39
1% neurips_2026.sty — NeurIPS 2026 official style
2% do not modify — conference formatting
3\NeedsTeXFormat{LaTeX2e}
4\ProvidesPackage{neurips_2026}[2026/01/15 v1.0]
5\RequirePackage{lineno}
6\RequirePackage[colorlinks=true]{hyperref}
7
8\newcommand{\@noticestring}{%
9Submitted to 40th Conference on Neural
10Information Processing Systems (NeurIPS 2026).
11Do not distribute.}
12
13\setlength{\textwidth}{5.5in}
14\setlength{\textheight}{9in}
15\widowpenalty=10000
16\clubpenalty=10000
17\flushbottom
18\sloppy
Perspective-Taking Improves LLM Theory-of-Mind
Anonymous Author(s) Affiliation Address email
Abstract
1 We study theory-of-mind in language models and propose a method.
2 Across five benchmarks it yields consistent gains.
1 Introduction
3 Large language models increasingly exhibit theory-of-mind (ToM)
4 capabilities, yet their accuracy on false-belief tasks remains
5 inconsistent across benchmarks.
2 Method
ℒ = −Σi yi log ŷi (1)
Submitted to 40th Conference on Neural Information Processing Systems (NeurIPS 2026). Do not distribute.
Evaluating Large Language Models in Theory of Mind Tasks
Michal Kosinski
Stanford University · arXiv:2302.02083
AbstractEleven LLMs were assessed using a custom-made battery of false-belief tasks — a gold standard in testing Theory of Mind in humans. Smaller and older models solved no tasks; GPT-3.5-turbo solved 20%, and GPT-4 solved 75% of the tasks, matching the performance of six-year-old children observed in past studies. We discuss whether this reflects genuine mentalizing or the statistics of language, and argue the distinction is increasingly hard to draw.
1 Tasks
We use the Sally–Anne unexpected-transfer task and the Smarties unexpected-contents task, each with carefully constructed controls to rule out shallow lexical shortcuts. Items are presented one sentence at a time to probe incremental belief updates.
Emergence of ToM-like behavior may be an unintended by-product of training on human-generated text. We caution against both over-attribution of mental states and premature dismissal, and call for adversarial benchmarks that separate the two.
[1] Wimmer & Perner. Beliefs about beliefs. 1983.
[2] Ullman. Large Language Models Fail on Trivial Alterations to Theory-of-Mind Tasks. 2023.
If it isn't in your library, it can't be cited. Nothing to hallucinate.
Compile · read · fix
LaTeX errors, debugged for you.
Every agent edit gets compiled. When the build breaks, the agent opens the logs, reads the error, and fixes its own mistake — then the PDF refreshes, like Overleaf with a mechanic inside.
perspective-taking-paper
Share
File Tree
figs
fig1.png
fig2.png
literati.bib
main.tex
neurips_2026.sty
File Outline
§Introduction8
§Method14
Papers 4
Wilf et al. (2023) · arXiv:2311.10227Think Twice: Perspective-Taking Improves Large Language Models’ Theory-of-Mind Capabilities
Sclar et al. (2023) · arXiv:2306.00924Minding Language Models' (Lack of) Theory of Mind
Zhou et al. (2023) · arXiv:2310.03051How FaR Are Large Language Models From Agents with Theory-of-Mind?
Kosinski (2023) · arXiv:2302.02083Evaluating Large Language Models in Theory of Mind Tasks
Ask anything…
chat (shift+tab to cycle)esc to interrupt
Aaab.*
Type to search main.tex
Changes
Message (⌘↵ to commit)
Staged Changes1
results.texsections/A
Working Changes2
main.texM
abstract.texsections/M
Nothing to commit — working tree clean
History
›Draft intro + method2hRevert
›Initial skeleton1dRevert
main.tex×
Editing ⌄
8\section{Introduction}
9
10Large language models increasingly exhibit theory-of-mind
11(ToM) capabilities, yet their accuracy on false-belief
12tasks remains inconsistent across benchmarks.
13
14\section{Method}
15
16
17
literati.bib is generated from your library — read-only
1/ 1
Perspective-Taking Improves LLM Theory-of-Mind
Anonymous Author(s) Affiliation Address email
Abstract
1 We study theory-of-mind in language models and propose a method.
2 Across five benchmarks it yields consistent gains.
1 Introduction
3 Large language models increasingly exhibit theory-of-mind (ToM)
4 capabilities, yet their accuracy on false-belief tasks remains
5 inconsistent across benchmarks.
Submitted to 40th Conference on Neural Information Processing Systems (NeurIPS 2026). Do not distribute.
All logsErrorsWarningsInfo
Compilation failed. See errors below.
hover to play — and click around: files, search, git · papers open in the PDF viewer
8\section{Introduction}
9
10Large language models increasingly exhibit theory-of-mind
11(ToM) capabilities, yet their accuracy on false-belief
12tasks remains inconsistent across benchmarks.
13
14\section{Method}
15
16\begin{equation}
17 \mathcal{L} = -\sum_i y_i \log \haty_i
18\end{equation}
19
8\section{Introduction}
9
10Large language models increasingly exhibit theory-of-mind
11(ToM) capabilities, yet their accuracy on false-belief
12tasks remains inconsistent across benchmarks.
13
14\section{Method}
15
16\begin{equation}
17 \mathcal{L} = -\sum_i y_i \log \hat{y}_i
18\end{equation}
19
10Large language models increasingly exhibit theory-of-mind
11(ToM) capabilities, yet their accuracy on false-belief
12-tasks remains inconsistent.
12+tasks remains inconsistent across benchmarks. We show
13+that a simple perspective-taking prompt closes much of
14+this gap \citep{wilf2023think}.
15
37We evaluate on the ToMI benchmark.
38-Our method improves accuracy.
38+Our method improves accuracy by 47\% on ToMI
39+and 14\% on BigToM over the GPT-3.5 baseline
40+\citep{wilf2023think}.
41We report mean over 5 seeds.
42-% TODO: add ablation table
42+Ablations in Table 2 isolate each component.
1We study theory-of-mind in language models and
2-propose a method.
2+propose Think-Twice, a perspective-taking prompt that
3+reframes each question from the target agent's view.
4Across five benchmarks it yields consistent gains.
Think Twice: Perspective-Taking Improves Large Language Models’ Theory-of-Mind Capabilities
Alex Wilf · Sihyun Shawn Lee · Paul Pu Liang · Louis-Philippe Morency
Carnegie Mellon University · arXiv:2311.10227
AbstractHuman interactions are deeply rooted in the interplay of thoughts, beliefs, and desires made possible by Theory of Mind (ToM). Although ToM may come naturally to us, emulating it presents a challenge to even the most advanced Large Language Models. We turn to the cognitive science theory "Simulation Theory" and introduce SimToM, a two-stage prompting framework based on perspective-taking: first filter context based on what the character in question knows, then answer a question about their mental state. Requiring no additional training and minimal prompt-tuning, SimToM shows substantial improvement over existing methods.
1 Introduction
Large language models increasingly succeed on tasks once thought to require social reasoning, yet their performance on false-belief tasks remains brittle. We hypothesize that much of this brittleness stems from a failure to anchor reasoning to a specific agent's knowledge state, rather than from a lack of underlying capability.
We test this hypothesis with a lightweight intervention that asks the model to restate what each character can and cannot observe before answering. The result is a large, consistent improvement that requires no additional parameters.
41Base
67CoT
88Ours
95Human
Figure 1. Accuracy on ToMI (%). SimToM (Ours) closes most of the gap to human performance.
2 Method
Given a scenario and a question about agent a, we prompt the model to enumerate the events a witnessed, mark events hidden from a, and only then resolve the query conditioned on a's partial observation. This decomposition mirrors the sub-steps humans perform implicitly.
[1] Sclar et al. Minding Language Models' (Lack of) Theory of Mind. 2023.
[2] Zhou et al. How FaR Are Large Language Models From Agents with Theory-of-Mind? 2023.
[3] Kosinski. Evaluating Large Language Models in Theory of Mind Tasks. 2023.
Minding Language Models' (Lack of) Theory of Mind: A Plug-and-Play Multi-Character Belief Tracker
Melanie Sclar · Sachin Kumar · Peter West · Alane Suhr · Yejin Choi · Yulia Tsvetkov
University of Washington · Allen Institute for AI · arXiv:2306.00924
AbstractTheory of mind — tracking what others believe — is a hallmark of social intelligence, but language models struggle to maintain consistent beliefs for multiple characters. We present SymbolicToM, a plug-and-play inference-time tracker that maintains an explicit belief graph per character. It substantially improves robustness on ToMI, especially under adversarial reorderings that break end-to-end models.
1 Motivation
End-to-end prompting conflates the beliefs of different characters, causing errors that compound as stories grow longer. We argue that an explicit, symbolic representation of "who knows what" is both interpretable and more robust than relying on the model's latent state.
49Vanilla
58CoT
93SymbolicToM
Figure 1. Robust accuracy under story reordering. The symbolic tracker degrades far less.
2 Belief Graphs
For each character we instantiate a graph whose nodes are propositions and whose edges encode observation events. Queries are answered by reading off the relevant character's sub-graph, sidestepping the interference that plagues monolithic prompting.
[1] Nematzadeh et al. Evaluating Theory of Mind in Question Answering. 2018.
[2] Le et al. Revisiting the Evaluation of Theory of Mind through Question Answering. 2019.
How FaR Are Large Language Models From Agents with Theory-of-Mind?
Pei Zhou · Aman Madaan · Srividya Pranavi Potharaju · Aditya Gupta · Kevin R. McKee · Ari Holtzman · Jay Pujara · Xiang Ren · Swaroop Mishra · Aida Nematzadeh · Shyam Upadhyay · Manaal Faruqui
University of Southern California · Google DeepMind · arXiv:2310.03051
Abstract"Thinking is for Doing." Humans infer other people’s mental states and act pragmatically on those inferences. Benchmarks such as ToMi ask models about characters’ beliefs, but do not test whether models can use those inferences to guide their actions. We propose Thinking for Doing (T4D), which requires connecting inferences about others’ mental states to actions in social scenarios, and introduce Foresee and Reflect (FaR), a zero-shot prompting framework that encourages LLMs to anticipate future challenges and reason about potential actions.
1 Thinking is for Doing
LLMs such as GPT-4 seemingly excel at tracking characters’ beliefs in stories, yet struggle to translate this capability into strategic action. The core challenge lies in identifying the implicit inferences about mental states — without being explicitly asked, as in ToMi — that lead to choosing the correct action.
50GPT-4
71+ FaR
Figure 1. Accuracy on T4D (%). FaR boosts GPT-4 from 50% to 71%, outperforming Chain-of-Thought and Self-Ask.
2 Foresee and Reflect
The foresee step prompts the model to anticipate likely future events and the challenges facing each character; the reflect step reasons over candidate actions against those futures before committing. FaR generalizes to diverse out-of-distribution story structures and scenarios.
[1] Yao et al. ReAct: Synergizing Reasoning and Acting in Language Models. 2022.
[2] Shinn et al. Reflexion: Language Agents with Verbal Reinforcement Learning. 2023.
1% literati.bib — generated by Literati from your library
2% papers are added via the arXiv API · do not edit by hand
3
4@article{wilf2023think,
5 title = {Think Twice: Perspective-Taking Improves Large Language
6Models' Theory-of-Mind Capabilities},
7 author = {Wilf, Alex and Lee, Sihyun Shawn and Liang, Paul Pu
8and Morency, Louis-Philippe},
9 year = {2023},
10 eprint = {2311.10227},
11}
12
13@article{sclar2023minding,
14 title = {Minding Language Models' (Lack of) Theory of Mind:
16 author = {Sclar, Melanie and Kumar, Sachin and West, Peter and
17Suhr, Alane and Choi, Yejin and Tsvetkov, Yulia},
18 year = {2023},
19 eprint = {2306.00924},
20}
21
22@article{zhou2023far,
23 title = {How FaR Are Large Language Models From Agents
24with Theory-of-Mind?},
25 author = {Zhou, Pei and Madaan, Aman and Potharaju, Srividya Pranavi
26and Gupta, Aditya and McKee, Kevin R. and Holtzman, Ari
27and Pujara, Jay and Ren, Xiang and Mishra, Swaroop
28and Nematzadeh, Aida and Upadhyay, Shyam and Faruqui, Manaal},
29 year = {2023},
30 eprint = {2310.03051},
31}
32
33@article{kosinski2023theory,
34 title = {Evaluating Large Language Models in Theory of Mind Tasks},
35 author = {Kosinski, Michal},
36 year = {2023},
37 eprint = {2302.02083},
38}
39
1% neurips_2026.sty — NeurIPS 2026 official style
2% do not modify — conference formatting
3\NeedsTeXFormat{LaTeX2e}
4\ProvidesPackage{neurips_2026}[2026/01/15 v1.0]
5\RequirePackage{lineno}
6\RequirePackage[colorlinks=true]{hyperref}
7
8\newcommand{\@noticestring}{%
9Submitted to 40th Conference on Neural
10Information Processing Systems (NeurIPS 2026).
11Do not distribute.}
12
13\setlength{\textwidth}{5.5in}
14\setlength{\textheight}{9in}
15\widowpenalty=10000
16\clubpenalty=10000
17\flushbottom
18\sloppy
Perspective-Taking Improves LLM Theory-of-Mind
Anonymous Author(s) Affiliation Address email
Abstract
1 We study theory-of-mind in language models and propose a method.
2 Across five benchmarks it yields consistent gains.
1 Introduction
3 Large language models increasingly exhibit theory-of-mind (ToM)
4 capabilities, yet their accuracy on false-belief tasks remains
5 inconsistent across benchmarks.
2 Method
ℒ = −Σi yi log ŷi (1)
Submitted to 40th Conference on Neural Information Processing Systems (NeurIPS 2026). Do not distribute.
Evaluating Large Language Models in Theory of Mind Tasks
Michal Kosinski
Stanford University · arXiv:2302.02083
AbstractEleven LLMs were assessed using a custom-made battery of false-belief tasks — a gold standard in testing Theory of Mind in humans. Smaller and older models solved no tasks; GPT-3.5-turbo solved 20%, and GPT-4 solved 75% of the tasks, matching the performance of six-year-old children observed in past studies. We discuss whether this reflects genuine mentalizing or the statistics of language, and argue the distinction is increasingly hard to draw.
1 Tasks
We use the Sally–Anne unexpected-transfer task and the Smarties unexpected-contents task, each with carefully constructed controls to rule out shallow lexical shortcuts. Items are presented one sentence at a time to probe incremental belief updates.
Emergence of ToM-like behavior may be an unintended by-product of training on human-generated text. We caution against both over-attribution of mental states and premature dismissal, and call for adversarial benchmarks that separate the two.
[1] Wimmer & Perner. Beliefs about beliefs. 1983.
[2] Ullman. Large Language Models Fail on Trivial Alterations to Theory-of-Mind Tasks. 2023.
You never have to decode an Undefined control sequence again.
Version control, built in
Version control, built in.
Every agent edit gets compiled. When the build breaks, the agent opens the logs, reads the error, and fixes its own mistake — then the PDF refreshes, like Overleaf with a mechanic inside.
perspective-taking-paper
Share
File Tree
figs
fig1.png
fig2.png
literati.bib
main.tex
neurips_2026.sty
File Outline
§Introduction8
§Background14
Papers 4
Wilf et al. (2023) · arXiv:2311.10227Think Twice: Perspective-Taking Improves Large Language Models’ Theory-of-Mind Capabilities
Sclar et al. (2023) · arXiv:2306.00924Minding Language Models' (Lack of) Theory of Mind
Zhou et al. (2023) · arXiv:2310.03051How FaR Are Large Language Models From Agents with Theory-of-Mind?
Kosinski (2023) · arXiv:2302.02083Evaluating Large Language Models in Theory of Mind Tasks
Ask anything…
chat (shift+tab to cycle)esc to interrupt
Aaab.*
Type to search main.tex
Changes
Message (⌘↵ to commit)
Staged Changes1
results.texsections/A
Working Changes2
main.texM
abstract.texsections/M
Nothing to commit — working tree clean
History
›Draft intro + method2hRevert
›Initial skeleton1dRevert
main.tex×
Editing ⌄
8\section{Introduction}
9
10Large language models increasingly exhibit theory-of-mind
11(ToM) capabilities, yet their accuracy on false-belief
12tasks remains inconsistent across benchmarks.
13
14\section{Background}
15
16
17
literati.bib is generated from your library — read-only
1/ 1
Perspective-Taking Improves LLM Theory-of-Mind
Anonymous Author(s) Affiliation Address email
Abstract
1 We study theory-of-mind in language models and propose a method.
2 Across five benchmarks it yields consistent gains.
1 Introduction
3 Large language models increasingly exhibit theory-of-mind (ToM)
4 capabilities, yet their accuracy on false-belief tasks remains
5 inconsistent across benchmarks.
Submitted to 40th Conference on Neural Information Processing Systems (NeurIPS 2026). Do not distribute.
All logsErrorsWarningsInfo
Compilation failed. See errors below.
hover to play — and click around: files, search, git · papers open in the PDF viewer
10Large language models increasingly exhibit theory-of-mind
11-(ToM) capabilities, yet accuracy on false-belief tasks varies.
11+(ToM) capabilities, yet their accuracy on false-belief
12+tasks remains inconsistent across benchmarks. A simple
13+perspective-taking prompt closes much of this gap.
14
11(ToM) capabilities, yet their accuracy on false-belief
12tasks remains inconsistent across benchmarks.
13+\paragraph{Contributions.} We propose
14+a perspective-taking prompt, evaluated
15+across five ToM benchmarks.
16
11(ToM) capabilities, yet their accuracy on false-belief
12tasks remains inconsistent across benchmarks.
13\paragraph{Contributions.} We propose
14a perspective-taking prompt, evaluated
15across five ToM benchmarks.
16
8\section{Introduction}
9
10Large language models increasingly exhibit theory-of-mind
11(ToM) capabilities, yet their accuracy on false-belief
12tasks remains inconsistent across benchmarks.
13\paragraph{Contributions.} We propose
14a perspective-taking prompt, evaluated
15across five ToM benchmarks.
16
17\section{Background}
18
10Large language models increasingly exhibit theory-of-mind
11(ToM) capabilities, yet their accuracy on false-belief
12tasks remains inconsistent across benchmarks. A simple
13perspective-taking prompt closes much of this gap.
14
10Large language models increasingly exhibit theory-of-mind
11(ToM) capabilities, yet their accuracy on false-belief
12-tasks remains inconsistent.
12+tasks remains inconsistent across benchmarks. We show
13+that a simple perspective-taking prompt closes much of
14+this gap \citep{wilf2023think}.
15
37We evaluate on the ToMI benchmark.
38-Our method improves accuracy.
38+Our method improves accuracy by 47\% on ToMI
39+and 14\% on BigToM over the GPT-3.5 baseline
40+\citep{wilf2023think}.
41We report mean over 5 seeds.
42-% TODO: add ablation table
42+Ablations in Table 2 isolate each component.
1We study theory-of-mind in language models and
2-propose a method.
2+propose Think-Twice, a perspective-taking prompt that
3+reframes each question from the target agent's view.
4Across five benchmarks it yields consistent gains.
Think Twice: Perspective-Taking Improves Large Language Models’ Theory-of-Mind Capabilities
Alex Wilf · Sihyun Shawn Lee · Paul Pu Liang · Louis-Philippe Morency
Carnegie Mellon University · arXiv:2311.10227
AbstractHuman interactions are deeply rooted in the interplay of thoughts, beliefs, and desires made possible by Theory of Mind (ToM). Although ToM may come naturally to us, emulating it presents a challenge to even the most advanced Large Language Models. We turn to the cognitive science theory "Simulation Theory" and introduce SimToM, a two-stage prompting framework based on perspective-taking: first filter context based on what the character in question knows, then answer a question about their mental state. Requiring no additional training and minimal prompt-tuning, SimToM shows substantial improvement over existing methods.
1 Introduction
Large language models increasingly succeed on tasks once thought to require social reasoning, yet their performance on false-belief tasks remains brittle. We hypothesize that much of this brittleness stems from a failure to anchor reasoning to a specific agent's knowledge state, rather than from a lack of underlying capability.
We test this hypothesis with a lightweight intervention that asks the model to restate what each character can and cannot observe before answering. The result is a large, consistent improvement that requires no additional parameters.
41Base
67CoT
88Ours
95Human
Figure 1. Accuracy on ToMI (%). SimToM (Ours) closes most of the gap to human performance.
2 Method
Given a scenario and a question about agent a, we prompt the model to enumerate the events a witnessed, mark events hidden from a, and only then resolve the query conditioned on a's partial observation. This decomposition mirrors the sub-steps humans perform implicitly.
[1] Sclar et al. Minding Language Models' (Lack of) Theory of Mind. 2023.
[2] Zhou et al. How FaR Are Large Language Models From Agents with Theory-of-Mind? 2023.
[3] Kosinski. Evaluating Large Language Models in Theory of Mind Tasks. 2023.
Minding Language Models' (Lack of) Theory of Mind: A Plug-and-Play Multi-Character Belief Tracker
Melanie Sclar · Sachin Kumar · Peter West · Alane Suhr · Yejin Choi · Yulia Tsvetkov
University of Washington · Allen Institute for AI · arXiv:2306.00924
AbstractTheory of mind — tracking what others believe — is a hallmark of social intelligence, but language models struggle to maintain consistent beliefs for multiple characters. We present SymbolicToM, a plug-and-play inference-time tracker that maintains an explicit belief graph per character. It substantially improves robustness on ToMI, especially under adversarial reorderings that break end-to-end models.
1 Motivation
End-to-end prompting conflates the beliefs of different characters, causing errors that compound as stories grow longer. We argue that an explicit, symbolic representation of "who knows what" is both interpretable and more robust than relying on the model's latent state.
49Vanilla
58CoT
93SymbolicToM
Figure 1. Robust accuracy under story reordering. The symbolic tracker degrades far less.
2 Belief Graphs
For each character we instantiate a graph whose nodes are propositions and whose edges encode observation events. Queries are answered by reading off the relevant character's sub-graph, sidestepping the interference that plagues monolithic prompting.
[1] Nematzadeh et al. Evaluating Theory of Mind in Question Answering. 2018.
[2] Le et al. Revisiting the Evaluation of Theory of Mind through Question Answering. 2019.
How FaR Are Large Language Models From Agents with Theory-of-Mind?
Pei Zhou · Aman Madaan · Srividya Pranavi Potharaju · Aditya Gupta · Kevin R. McKee · Ari Holtzman · Jay Pujara · Xiang Ren · Swaroop Mishra · Aida Nematzadeh · Shyam Upadhyay · Manaal Faruqui
University of Southern California · Google DeepMind · arXiv:2310.03051
Abstract"Thinking is for Doing." Humans infer other people’s mental states and act pragmatically on those inferences. Benchmarks such as ToMi ask models about characters’ beliefs, but do not test whether models can use those inferences to guide their actions. We propose Thinking for Doing (T4D), which requires connecting inferences about others’ mental states to actions in social scenarios, and introduce Foresee and Reflect (FaR), a zero-shot prompting framework that encourages LLMs to anticipate future challenges and reason about potential actions.
1 Thinking is for Doing
LLMs such as GPT-4 seemingly excel at tracking characters’ beliefs in stories, yet struggle to translate this capability into strategic action. The core challenge lies in identifying the implicit inferences about mental states — without being explicitly asked, as in ToMi — that lead to choosing the correct action.
50GPT-4
71+ FaR
Figure 1. Accuracy on T4D (%). FaR boosts GPT-4 from 50% to 71%, outperforming Chain-of-Thought and Self-Ask.
2 Foresee and Reflect
The foresee step prompts the model to anticipate likely future events and the challenges facing each character; the reflect step reasons over candidate actions against those futures before committing. FaR generalizes to diverse out-of-distribution story structures and scenarios.
[1] Yao et al. ReAct: Synergizing Reasoning and Acting in Language Models. 2022.
[2] Shinn et al. Reflexion: Language Agents with Verbal Reinforcement Learning. 2023.
1% literati.bib — generated by Literati from your library
2% papers are added via the arXiv API · do not edit by hand
3
4@article{wilf2023think,
5 title = {Think Twice: Perspective-Taking Improves Large Language
6Models' Theory-of-Mind Capabilities},
7 author = {Wilf, Alex and Lee, Sihyun Shawn and Liang, Paul Pu
8and Morency, Louis-Philippe},
9 year = {2023},
10 eprint = {2311.10227},
11}
12
13@article{sclar2023minding,
14 title = {Minding Language Models' (Lack of) Theory of Mind:
16 author = {Sclar, Melanie and Kumar, Sachin and West, Peter and
17Suhr, Alane and Choi, Yejin and Tsvetkov, Yulia},
18 year = {2023},
19 eprint = {2306.00924},
20}
21
22@article{zhou2023far,
23 title = {How FaR Are Large Language Models From Agents
24with Theory-of-Mind?},
25 author = {Zhou, Pei and Madaan, Aman and Potharaju, Srividya Pranavi
26and Gupta, Aditya and McKee, Kevin R. and Holtzman, Ari
27and Pujara, Jay and Ren, Xiang and Mishra, Swaroop
28and Nematzadeh, Aida and Upadhyay, Shyam and Faruqui, Manaal},
29 year = {2023},
30 eprint = {2310.03051},
31}
32
33@article{kosinski2023theory,
34 title = {Evaluating Large Language Models in Theory of Mind Tasks},
35 author = {Kosinski, Michal},
36 year = {2023},
37 eprint = {2302.02083},
38}
39
1% neurips_2026.sty — NeurIPS 2026 official style
2% do not modify — conference formatting
3\NeedsTeXFormat{LaTeX2e}
4\ProvidesPackage{neurips_2026}[2026/01/15 v1.0]
5\RequirePackage{lineno}
6\RequirePackage[colorlinks=true]{hyperref}
7
8\newcommand{\@noticestring}{%
9Submitted to 40th Conference on Neural
10Information Processing Systems (NeurIPS 2026).
11Do not distribute.}
12
13\setlength{\textwidth}{5.5in}
14\setlength{\textheight}{9in}
15\widowpenalty=10000
16\clubpenalty=10000
17\flushbottom
18\sloppy
Perspective-Taking Improves LLM Theory-of-Mind
Anonymous Author(s) Affiliation Address email
Abstract
1 We study theory-of-mind in language models and propose a method.
2 Across five benchmarks it yields consistent gains.
1 Introduction
3 Large language models increasingly exhibit theory-of-mind (ToM)
4 capabilities, yet their accuracy on false-belief tasks remains
5 inconsistent across benchmarks.
2 Method
ℒ = −Σi yi log ŷi (1)
Submitted to 40th Conference on Neural Information Processing Systems (NeurIPS 2026). Do not distribute.
Evaluating Large Language Models in Theory of Mind Tasks
Michal Kosinski
Stanford University · arXiv:2302.02083
AbstractEleven LLMs were assessed using a custom-made battery of false-belief tasks — a gold standard in testing Theory of Mind in humans. Smaller and older models solved no tasks; GPT-3.5-turbo solved 20%, and GPT-4 solved 75% of the tasks, matching the performance of six-year-old children observed in past studies. We discuss whether this reflects genuine mentalizing or the statistics of language, and argue the distinction is increasingly hard to draw.
1 Tasks
We use the Sally–Anne unexpected-transfer task and the Smarties unexpected-contents task, each with carefully constructed controls to rule out shallow lexical shortcuts. Items are presented one sentence at a time to probe incremental belief updates.
Emergence of ToM-like behavior may be an unintended by-product of training on human-generated text. We caution against both over-attribution of mental states and premature dismissal, and call for adversarial benchmarks that separate the two.
[1] Wimmer & Perner. Beliefs about beliefs. 1983.
[2] Ullman. Large Language Models Fail on Trivial Alterations to Theory-of-Mind Tasks. 2023.
You never have to decode an Undefined control sequence again.
Gemini-powered discovery
Find relevant papers with Gemini. Never copy a BibTeX again.
Ask what's missing. The find agent reads your draft, skims your library, and searches with Gemini. You pick the papers — literati.bib updates itself and they're instantly citeable.
perspective-taking-paper
Share
File Tree
figs
fig1.png
fig2.png
literati.bib
main.tex
neurips_2026.sty
File Outline
§Introduction8
§Background14
Papers 4
Wilf et al. (2023) · arXiv:2311.10227Think Twice: Perspective-Taking Improves Large Language Models’ Theory-of-Mind Capabilities
Sclar et al. (2023) · arXiv:2306.00924Minding Language Models' (Lack of) Theory of Mind
Zhou et al. (2023) · arXiv:2310.03051How FaR Are Large Language Models From Agents with Theory-of-Mind?
Kosinski (2023) · arXiv:2302.02083Evaluating Large Language Models in Theory of Mind Tasks
Ask anything…
chat (shift+tab to cycle)esc to interrupt
Aaab.*
Type to search main.tex
Changes
Message (⌘↵ to commit)
Staged Changes1
results.texsections/A
Working Changes2
main.texM
abstract.texsections/M
Nothing to commit — working tree clean
History
›Draft intro + method2hRevert
›Initial skeleton1dRevert
main.tex×
Editing ⌄
8\section{Introduction}
9
10Large language models increasingly exhibit theory-of-mind
11(ToM) capabilities, yet their accuracy on false-belief
12tasks remains inconsistent across benchmarks.
13
14\section{Background}
15
16
17
literati.bib is generated from your library — read-only
1/ 1
Perspective-Taking Improves LLM Theory-of-Mind
Anonymous Author(s) Affiliation Address email
Abstract
1 We study theory-of-mind in language models and propose a method.
2 Across five benchmarks it yields consistent gains.
1 Introduction
3 Large language models increasingly exhibit theory-of-mind (ToM)
4 capabilities, yet their accuracy on false-belief tasks remains
5 inconsistent across benchmarks.
Submitted to 40th Conference on Neural Information Processing Systems (NeurIPS 2026). Do not distribute.
All logsErrorsWarningsInfo
Compilation failed. See errors below.
hover to play — and click around: files, search, git · papers open in the PDF viewer
10Large language models increasingly exhibit theory-of-mind
11(ToM) capabilities, yet their accuracy on false-belief
12-tasks remains inconsistent.
12+tasks remains inconsistent across benchmarks. We show
13+that a simple perspective-taking prompt closes much of
14+this gap \citep{wilf2023think}.
15
37We evaluate on the ToMI benchmark.
38-Our method improves accuracy.
38+Our method improves accuracy by 47\% on ToMI
39+and 14\% on BigToM over the GPT-3.5 baseline
40+\citep{wilf2023think}.
41We report mean over 5 seeds.
42-% TODO: add ablation table
42+Ablations in Table 2 isolate each component.
1We study theory-of-mind in language models and
2-propose a method.
2+propose Think-Twice, a perspective-taking prompt that
3+reframes each question from the target agent's view.
4Across five benchmarks it yields consistent gains.
Think Twice: Perspective-Taking Improves Large Language Models’ Theory-of-Mind Capabilities
Alex Wilf · Sihyun Shawn Lee · Paul Pu Liang · Louis-Philippe Morency
Carnegie Mellon University · arXiv:2311.10227
AbstractHuman interactions are deeply rooted in the interplay of thoughts, beliefs, and desires made possible by Theory of Mind (ToM). Although ToM may come naturally to us, emulating it presents a challenge to even the most advanced Large Language Models. We turn to the cognitive science theory "Simulation Theory" and introduce SimToM, a two-stage prompting framework based on perspective-taking: first filter context based on what the character in question knows, then answer a question about their mental state. Requiring no additional training and minimal prompt-tuning, SimToM shows substantial improvement over existing methods.
1 Introduction
Large language models increasingly succeed on tasks once thought to require social reasoning, yet their performance on false-belief tasks remains brittle. We hypothesize that much of this brittleness stems from a failure to anchor reasoning to a specific agent's knowledge state, rather than from a lack of underlying capability.
We test this hypothesis with a lightweight intervention that asks the model to restate what each character can and cannot observe before answering. The result is a large, consistent improvement that requires no additional parameters.
41Base
67CoT
88Ours
95Human
Figure 1. Accuracy on ToMI (%). SimToM (Ours) closes most of the gap to human performance.
2 Method
Given a scenario and a question about agent a, we prompt the model to enumerate the events a witnessed, mark events hidden from a, and only then resolve the query conditioned on a's partial observation. This decomposition mirrors the sub-steps humans perform implicitly.
[1] Sclar et al. Minding Language Models' (Lack of) Theory of Mind. 2023.
[2] Zhou et al. How FaR Are Large Language Models From Agents with Theory-of-Mind? 2023.
[3] Kosinski. Evaluating Large Language Models in Theory of Mind Tasks. 2023.
Minding Language Models' (Lack of) Theory of Mind: A Plug-and-Play Multi-Character Belief Tracker
Melanie Sclar · Sachin Kumar · Peter West · Alane Suhr · Yejin Choi · Yulia Tsvetkov
University of Washington · Allen Institute for AI · arXiv:2306.00924
AbstractTheory of mind — tracking what others believe — is a hallmark of social intelligence, but language models struggle to maintain consistent beliefs for multiple characters. We present SymbolicToM, a plug-and-play inference-time tracker that maintains an explicit belief graph per character. It substantially improves robustness on ToMI, especially under adversarial reorderings that break end-to-end models.
1 Motivation
End-to-end prompting conflates the beliefs of different characters, causing errors that compound as stories grow longer. We argue that an explicit, symbolic representation of "who knows what" is both interpretable and more robust than relying on the model's latent state.
49Vanilla
58CoT
93SymbolicToM
Figure 1. Robust accuracy under story reordering. The symbolic tracker degrades far less.
2 Belief Graphs
For each character we instantiate a graph whose nodes are propositions and whose edges encode observation events. Queries are answered by reading off the relevant character's sub-graph, sidestepping the interference that plagues monolithic prompting.
[1] Nematzadeh et al. Evaluating Theory of Mind in Question Answering. 2018.
[2] Le et al. Revisiting the Evaluation of Theory of Mind through Question Answering. 2019.
How FaR Are Large Language Models From Agents with Theory-of-Mind?
Pei Zhou · Aman Madaan · Srividya Pranavi Potharaju · Aditya Gupta · Kevin R. McKee · Ari Holtzman · Jay Pujara · Xiang Ren · Swaroop Mishra · Aida Nematzadeh · Shyam Upadhyay · Manaal Faruqui
University of Southern California · Google DeepMind · arXiv:2310.03051
Abstract"Thinking is for Doing." Humans infer other people’s mental states and act pragmatically on those inferences. Benchmarks such as ToMi ask models about characters’ beliefs, but do not test whether models can use those inferences to guide their actions. We propose Thinking for Doing (T4D), which requires connecting inferences about others’ mental states to actions in social scenarios, and introduce Foresee and Reflect (FaR), a zero-shot prompting framework that encourages LLMs to anticipate future challenges and reason about potential actions.
1 Thinking is for Doing
LLMs such as GPT-4 seemingly excel at tracking characters’ beliefs in stories, yet struggle to translate this capability into strategic action. The core challenge lies in identifying the implicit inferences about mental states — without being explicitly asked, as in ToMi — that lead to choosing the correct action.
50GPT-4
71+ FaR
Figure 1. Accuracy on T4D (%). FaR boosts GPT-4 from 50% to 71%, outperforming Chain-of-Thought and Self-Ask.
2 Foresee and Reflect
The foresee step prompts the model to anticipate likely future events and the challenges facing each character; the reflect step reasons over candidate actions against those futures before committing. FaR generalizes to diverse out-of-distribution story structures and scenarios.
[1] Yao et al. ReAct: Synergizing Reasoning and Acting in Language Models. 2022.
[2] Shinn et al. Reflexion: Language Agents with Verbal Reinforcement Learning. 2023.
1% literati.bib — generated by Literati from your library
2% papers are added via the arXiv API · do not edit by hand
3
4@article{wilf2023think,
5 title = {Think Twice: Perspective-Taking Improves Large Language
6Models' Theory-of-Mind Capabilities},
7 author = {Wilf, Alex and Lee, Sihyun Shawn and Liang, Paul Pu
8and Morency, Louis-Philippe},
9 year = {2023},
10 eprint = {2311.10227},
11}
12
13@article{sclar2023minding,
14 title = {Minding Language Models' (Lack of) Theory of Mind:
16 author = {Sclar, Melanie and Kumar, Sachin and West, Peter and
17Suhr, Alane and Choi, Yejin and Tsvetkov, Yulia},
18 year = {2023},
19 eprint = {2306.00924},
20}
21
22@article{zhou2023far,
23 title = {How FaR Are Large Language Models From Agents
24with Theory-of-Mind?},
25 author = {Zhou, Pei and Madaan, Aman and Potharaju, Srividya Pranavi
26and Gupta, Aditya and McKee, Kevin R. and Holtzman, Ari
27and Pujara, Jay and Ren, Xiang and Mishra, Swaroop
28and Nematzadeh, Aida and Upadhyay, Shyam and Faruqui, Manaal},
29 year = {2023},
30 eprint = {2310.03051},
31}
32
33@article{kosinski2023theory,
34 title = {Evaluating Large Language Models in Theory of Mind Tasks},
35 author = {Kosinski, Michal},
36 year = {2023},
37 eprint = {2302.02083},
38}
39
1% neurips_2026.sty — NeurIPS 2026 official style
2% do not modify — conference formatting
3\NeedsTeXFormat{LaTeX2e}
4\ProvidesPackage{neurips_2026}[2026/01/15 v1.0]
5\RequirePackage{lineno}
6\RequirePackage[colorlinks=true]{hyperref}
7
8\newcommand{\@noticestring}{%
9Submitted to 40th Conference on Neural
10Information Processing Systems (NeurIPS 2026).
11Do not distribute.}
12
13\setlength{\textwidth}{5.5in}
14\setlength{\textheight}{9in}
15\widowpenalty=10000
16\clubpenalty=10000
17\flushbottom
18\sloppy
Perspective-Taking Improves LLM Theory-of-Mind
Anonymous Author(s) Affiliation Address email
Abstract
1 We study theory-of-mind in language models and propose a method.
2 Across five benchmarks it yields consistent gains.
1 Introduction
3 Large language models increasingly exhibit theory-of-mind (ToM)
4 capabilities, yet their accuracy on false-belief tasks remains
5 inconsistent across benchmarks.
2 Method
ℒ = −Σi yi log ŷi (1)
Submitted to 40th Conference on Neural Information Processing Systems (NeurIPS 2026). Do not distribute.
Evaluating Large Language Models in Theory of Mind Tasks
Michal Kosinski
Stanford University · arXiv:2302.02083
AbstractEleven LLMs were assessed using a custom-made battery of false-belief tasks — a gold standard in testing Theory of Mind in humans. Smaller and older models solved no tasks; GPT-3.5-turbo solved 20%, and GPT-4 solved 75% of the tasks, matching the performance of six-year-old children observed in past studies. We discuss whether this reflects genuine mentalizing or the statistics of language, and argue the distinction is increasingly hard to draw.
1 Tasks
We use the Sally–Anne unexpected-transfer task and the Smarties unexpected-contents task, each with carefully constructed controls to rule out shallow lexical shortcuts. Items are presented one sentence at a time to probe incremental belief updates.
Emergence of ToM-like behavior may be an unintended by-product of training on human-generated text. We caution against both over-attribution of mental states and premature dismissal, and call for adversarial benchmarks that separate the two.
[1] Wimmer & Perner. Beliefs about beliefs. 1983.
[2] Ullman. Large Language Models Fail on Trivial Alterations to Theory-of-Mind Tasks. 2023.
Understanding Social Reasoning in Language Models with Language Models
Kanishk Gandhi, Jan-Philipp Fränken, Tobias Gerstenberg, Noah D. Goodman
Stanford University · NeurIPS 2023 · arXiv:2306.15448
AbstractWe introduce BigToM, a benchmark of theory-of-mind reasoning built with a causal template that composes true- and false-belief scenarios procedurally. Populating the template with a language model yields controllable, contamination-resistant evaluations. We find performance is highly sensitive to how beliefs are elicited, and many apparent successes collapse under matched controls.
1 A causal template for social reasoning
Each item is generated from a causal graph over desires, percepts, and beliefs, letting us hold reasoning structure fixed while varying surface content. This isolates belief tracking from lexical shortcuts and supports large, balanced test sets.
34Zero-shot
52CoT
71+Persp.
88Human
Figure 1. Accuracy on BigToM false-belief items across prompting conditions.
2 Findings
Perspective-taking prompts close much, but not all, of the gap to humans. Robustness varies sharply across scenario types, suggesting current models track beliefs opportunistically rather than systematically.
[1] Baker et al. Rational quantitative attribution of beliefs, desires, and percepts. 2017.
[2] Sap et al. Neural Theory-of-Mind? On the Limits of Social Intelligence in LLMs. 2022.
