Beyond Belief Nesting: Rethinking Higher-Order Theory of Mind in Large Language Models
Authors: Shanshan You, Zhengqi Wang, Hao Liu, Yuqin Liu, Ya He, Xin Hao, Liyuan Wang
Organizations: Department of Psychological and Cognitive Sciences, Tsinghua University · IDG/McGovern Institute for Brain Research, Tsinghua University · Department of Cognitive Sciences, University of California, Irvine
Theory of mind (ToM), the ability to reason about others' mental states, underlies effective communication and coordination. Its higher-order form enables reasoning over nested perspectives in strategic interactions. However, existing evaluations often equate belief-nesting depth with reasoning order, allowing nominally higher-order questions to admit lower-order shortcuts. We formalize effective order through the distinct events supporting successive belief prefixes, revealing widespread order inflation across existing benchmarks. We introduce HopToM, with certified full-order questions and controls for lower-order alternatives. Under direct prompting, mean accuracy across four frontier LLMs falls from 73.5% at first order to 20.0% at fourth order, and human participants also struggle. Causal interventions link a recurring class of errors to evidence for shallower belief prefixes. We introduce HopTrace, a plug-and-play inference-time method that encodes each interaction once into perspective records and resolves nested queries deterministically. On certified narrative questions, it improves fourth- and fifth-order accuracy by 22.0 and 38.5 percentage points over strong prompting baselines, while using 20-39 times fewer tokens than multi-call methods. These results sharpen how higher-order ToM is evaluated and offer a practical way to improve perspective reasoning in interactive agents.
Figures & tables
Figure 1: Nested beliefs behind a fourth-order question.
Figure 2: Belief nesting can overstate effective order. Left: The same fourth-order query has r=1 in A and r=4 in B. Colors mark the event supporting each prefix. Right: HopTrace encodes perspective records and resolves each prefix independently. Sequential backward lookup instead selects t=2 and incorrectly answers “crate”.
Benchmark
N
k0
k
Mean r
Full order
ToMi
1,998
2
2
1.36
36.2%
ExploreToM
3,298
2
2
1.07
6.6%
Hi-ToM
120
4
4
2.07
2.5%
[2pt/2pt] HopToM (Ours)
200
5
5
5.00
100%
Table 1: Effective order at each benchmark’s highest published level. N is the number of questions, k is the belief-chain length. Full order is the share with r=k .
Subset
k
Mean r
Full order
State match
r counts
All
1
1.00
100%
50.0%
120
2
1.40
40.0%
66.7%
72 / 48
3
1.68
9.2%
79.2%
50 / 59 / 11
4
2.07
2.5%
85.0%
26 / 63 / 28 / 3
w/o Claims
2
1.29
29.3%
73.3%
53 / 22
3
1.47
2.7%
86.7%
42 / 31 / 2
Table 2: Hi-ToM audit by belief-chain length. Each order contains 120 unique questions, including 75 without and 45 with claim events. Full order is the share with r=k . State match is the share whose answer equals the answer to a shorter prefix or the target’s actual final location. The last column gives counts for r=1,…,k .
Model
1st
2nd
3rd
4th
5th
All
GPT-6 Sol
92.0
62.0
58.5
32.5
36.0
56.2
Claude Opus 5.5
69.5
33.0
42.5
16.0
21.5
36.5
Gemini 3 Flash
72.5
34.5
41.0
17.0
16.0
36.2
DeepSeek-V4-Pro
60.0
49.0
32.5
14.5
14.0
34.0
Average
73.5
44.6
43.6
20.0
21.9
40.7
Table 3: Direct-prompting accuracy (%) on the 1,000 certified full-order questions of HopToM . Each order contains 200 questions.
Figure 3: Results on certified questions and controls. (a) Fourth- and fifth-order results for human participants, who could not reread the stories, and four frontier LLMs. (b) Results by syntactic order for open-weight models.
HopToM (True Higher-Order)
Hi-ToM
Method
1st
2nd
3rd
4th
5th
All
All
GPT-6 Sol
CoT
94
58
52
42
44
58.0
80.3
TimeToM
78
83
56
58
50
65.0
78.5
RecToM
85
74
52
58
36
61.0
79.2
[2pt/2pt]
HopTrace
89
70
67
76
90
78.4
82.4
DeepSeek
CoT
76
68
35
23
14
43.2
62.5
Table 4: Accuracy (%) on HopToM and Hi-ToM. Models: GPT-6 Sol, DeepSeek-V4-Pro, Gemini 3 Flash, and Claude Opus 5.5. Best per model and column in bold.
Figure 4: Accuracy, inference cost, and ablations on the narrative part of HopToM . (a) Accuracy on 500 certified full-order questions against tokens per story. HopTrace answers all nine questions per story with one encoding call, while TimeToM and RecToM answer the five certified questions. (b) Accuracy of the full method and its ablations on all 900 narrative questions. Hatching denotes individual models and dashed lines denote model averages.
Figure 5: Fourth- and fifth-order accuracy on the certified questions of HopTrace with reasoning enabled.
Figure 6: Direct answers. (a) Wrong answers by order k (%). Shallower prefix : answer to a shorter prefix. Innermost belief : the last-named agent’s belief. (b) Log-probability drop when supporting paragraph is hidden.
Benchmark
Source
Max level
Size
Format
Order convention
ToMi †
Le et al. (2019)
2
5,994 test questions
open-ended
nesting depth
Hi-ToM †
Wu et al. (2023)
4
120 stories, 600 unique questions
15-way choice
nesting depth
FANToM
Kim et al. (2023)
2
256 conversations, 870 short sets
choice, free-form, list
nesting depth
OpenToM
Xu et al. (2024)
2
696 narratives, 16,008 questions
2/3-way choice
nesting depth
ToMBench
Chen et al. (2024)
2
2,860 items
4-way choice
task taxonomy
ExploreToM †
Sclar et al. (2025)
2
619 stories, 13,309 questions (sample)
open-ended, yes/no
nesting depth
Table 5: Benchmarks with nested-belief questions. Maximum levels follow each source’s convention. † Audited in this work; Appendix A describes the audited subsets.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Family
Type
Definition
Asymmetric Epistemic Access
present but inattentive
An agent is physically present during an event but does not attend to it and therefore does not acquire the relevant information.
secret observation
An agent observes an event without one or more other agents knowing that the observation occurred.
Temporal Presence Dynamics
leave and return
An agent leaves before a relevant event and later returns without direct access to what occurred during the absence.
late joining
An agent enters after earlier events have occurred and lacks the information available before arrival.
Strategic Information Presentation
deceptive communication
An agent intentionally provides false or misleading information to influence another agent’s belief.
Unintentional Epistemic Error
observation misperception
An agent observes an event but incorrectly interprets what happened.
Appendix
Table 6: The ten scenario types of HopToM , grouped into six families. A prototype can combine several types.
1st
2nd
3rd
4th
5th
All
Certified
58.3
31.7
58.3
6.7
11.7
33.3
Control
–
76.7
55.0
41.7
11.7
46.3
Overall
58.3
54.2
56.7
24.2
11.7
39.1
Appendix
Table 7: Human accuracy (%) by syntactic order.
Model
Open
MC
p
Certified
Controls
Claude Opus 5.5
34.0
92.4
<10−100
35.2 → 89.2
32.5 → 96.5
GPT-6 Sol
55.4
58.3
0.12
55.4 → 55.6
55.5 → 61.8
Gemini 3 Flash
34.2
44.0
<10−10
34.0 → 45.0
34.5 → 42.8
DeepSeek-V4-Pro
36.2
30.0
2×10−4
34.2 → 33.2
38.8 → 26.0
Appendix
Table 8: Open-ended and seven-option multiple-choice accuracy (%) on the 900 questions of the 100 narrative realizations. Open and MC : accuracy in each format. p : paired McNemar test between the two formats. Certified and Controls : accuracy on each question subset, open-ended → multiple choice.