From Belief Inference to Task Action: A Multi-Level Evaluation of Theory of Mind on Verifiable Tasks

Zhenshuai Yin; Yongxin Guo; Mingda Zhang; Zhongze Luo; Haoyue Liu; Lianmin Zhang; Xiaoying Tang

cs.CLOct 26, 2026

From Belief Inference to Task Action: A Multi-Level Evaluation of Theory of Mind on Verifiable Tasks

Authors: Zhenshuai Yin, Yongxin Guo, Mingda Zhang, Zhongze Luo, Haoyue Liu, Lianmin Zhang, Xiaoying Tang

Organizations: School of Science and Engineering, The Chinese University of Hong Kong (Shenzhen), Longgang, Shenzhen, Guangdong 518172, P.R. China · ByteDance Inc. · Shenzhen Research Institute of Big Data, Shenzhen, China

Abstract

Effective collaboration requires language models to understand what others know and believe, and to act accordingly. Models that assume others share their information may omit necessary communication and undermine coordination. Many theory-of-mind (ToM) benchmarks assess mental-state inference through story questions, but correct answers do not establish that models can use these judgments in subsequent tasks. We introduce a multi-level evaluation built on existing, verifiable tasks. We extend household collaboration and calendar scheduling with asymmetric information while retaining their original constraints and validators. Independent interfaces assess world understanding, information access, beliefs, decisions, and execution. Controls supply personal evidence or correct beliefs to distinguish state recovery from state use. Across six model configurations, these abilities diverge. On matched scheduling cases, Claude recovers all agent beliefs correctly but produces subjectively optimal plans in only 67.9% of cases. In household collaboration, supplying beliefs raises Qwen-27B success from 55.0% to 75.0%, yet Qwen-4B answers every fixed supplied-belief source question correctly while completing only 20.0% of the corresponding continuation condition. These results show why ToM evaluation must connect perspective inference to independently verified decisions and task consequences.

Figures & tables

Figure 1: An actual CalBench case connects private information to complete plans. The same five-meeting proposal is optimal for uninformed A but infeasible for B, who knows slot 8 is blocked. Both proposals fail in the actual world. Native replay and subjective solvers distinguish perspective-specific correctness from world success. Appendix M gives all five proposals.

Figure 2: Evaluation from personal evidence to task action. Independent interfaces probe world, access, beliefs, and decisions; native execution supplies action consequences and subsequent observations. Supplied-access and supplied-belief controls diagnose state recovery and use. The right-hand examples illustrate reference states, not measured model answers.

C-WAH: fixed nodes
Model	World	Access	Narrative belief	Struct. belief	Given access belief	Direct decision	Given belief decision
Qwen3.5-4B	93.1	33.8	49.3	38.6	60.2	66.5	100.0
Qwen3.5-9B	88.6	24.2	60.3	49.2	69.0	72.5	99.1
Qwen3.5-27B	96.3	42.3	62.8	63.6	91.4	82.1	100.0
Claude Opus 5.5	99.0	75.8	71.2	73.4	92.3	87.6	100.0
Gemini 3.1 Pro Preview	99.4	99.4	99.4	99.0	99.9	86.2	100.0

Table 1: Results on common support across six configurations (%); bold marks column maxima. C-WAH uses 218 fixed nodes (1,426 world entries; 2,852 access/belief entries). CalBench uses 916 common complete requests, with interface denominators in the appendix. Native comparisons pair 40 continuation and 35 route configurations. Belief entries are pooled; whole tables must be entirely correct. Coverage and all-valid results are separate. Columns measure different targets.

Figure 3: Information boundaries on the same household references: 1446 current, 1159 unknown, and 247 stale entries. Narrative and supplied-access inputs test whether agents’ states remain distinct from world truth. Nodes repeat within 43 configurations.

Figure 4: Correct source choices partitioned by whether the independent narrative belief table is entirely correct, on the same 218 nodes. Correct choices coexist with partial state errors; a task-relevant answer does not certify the entire perspective.

Figure 5: Native task success on 40 common configurations. Circles denote direct continuation and diamonds supplied beliefs. Labels show successful episodes. Lines connect input conditions, not time.

Figure 6: Belief recovery versus planning on the same 91 stories (455 agent cases per axis). Beliefs require all 80 entries; plans require subjective feasibility and optimality. The dashed line marks equal percentages, not equivalent metrics.

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Metric	Qwen-4B	Qwen-9B	Qwen-27B	Claude	Gemini	GPT
World	1327/1426	1263/1426	1373/1426	1412/1426	1418/1426	1355/1426
Access	964/2852	690/2852	1207/2852	2163/2852	2836/2852	2466/2852
Narrative beliefs	1407/2852	1719/2852	1790/2852	2031/2852	2836/2852	2701/2852
Structured beliefs	1102/2852	1404/2852	1813/2852	2094/2852	2823/2852	2624/2852
Given access: beliefs	1717/2852	1969/2852	2606/2852	2631/2852	2850/2852	2840/2852
Direct decisions	145/218	158/218	179/218	191/218	188/218	183/218

Appendix

Table 2: Exact C-WAH counts on the same 218 nodes. Textual and symbolic supplied beliefs are scored separately.

Metric	Qwen-4B	Qwen-9B	Qwen-27B	Claude	Gemini	GPT
World	2281/9600	7365/9600	9592/9600	9598/9600	9520/9600	9592/9600
Access	38/600	124/600	489/600	583/600	594/600	590/600
Narrative beliefs	26/825	88/825	794/825	825/825	800/825	766/825
Structured beliefs	33/875	165/875	842/875	855/875	870/875	840/875
Direct planning	64/485	252/485	430/485	334/485	485/485	393/485
Text beliefs: planning	225/735	394/735	631/735	566/735	735/735	586/735

Appendix

Table 3: Exact CalBench counts on common complete support, corresponding to the main table. Supplied access is listed separately because it jointly requests beliefs and proposals, unlike the separate belief and planning interfaces.

Metric	Qwen-4B	Qwen-9B	Qwen-27B	Claude	Gemini	GPT
Narrative: Current	785/1446	851/1446	1156/1446	1312/1446	1444/1446	1335/1446
Narrative: Unknown	600/1159	851/1159	534/1159	561/1159	1159/1159	1148/1159
Narrative: Stale	22/247	17/247	100/247	158/247	233/247	218/247
Structured: Current	885/1446	937/1446	1240/1446	1222/1446	1426/1446	1317/1446
Structured: Unknown	200/1159	450/1159	466/1159	738/1159	1159/1159	1109/1159
Structured: Stale	17/247	17/247	107/247	134/247	238/247	198/247

Appendix

Table 4: Exact C-WAH counts by reference belief category, using the same frozen entries across configurations.

Interface	Input in addition to task rules	Output target
WORLD_ACCESS	Complete narrative	World state and personal evidence collections
BELIEF_NL	Complete narrative	Each agent’s beliefs
BELIEF_STRUCT	Structured event history	Each agent’s beliefs
GIVEN_NL	World state and textual reference beliefs	Decisions under personal information
GIVEN_SYMBOL	World state and symbolic reference beliefs	Decisions under personal information
DIRECT	Complete narrative	Decisions under personal information

Appendix

Table 5: Shared interfaces. C-WAH decisions specify the believed source for a fixed assignment. CalBench decisions give a complete five-meeting proposal for each of five agents. Supplied states come from the evaluator, not model answers to other interfaces.

Figure 7: Belief retention and updating across C-WAH models. Circles represent 2248 entries that should be retained without new evidence, and squares 170 actual updates. Blue denotes narrative input and green supplied access. The remaining 434 confirmation entries are excluded.

Configuration	Reasoning control	C-WAH limit	CalBench limit	Temperature / top- $p$
Qwen3.5-4B, 9B	Thinking off	4096	12288	0 / 1
Qwen3.5-27B	Thinking off	4096	12288	0 / 1
Claude Opus 5.5	Medium	32768	32768	Unspecified
Gemini 3.1 Pro Preview	Medium	32768	32768	Unspecified
GPT-6.1 Sol	Medium	32768	32768	Unspecified

Appendix

Table 6: Evaluated configurations and output limits in tokens. Closed configurations request the recorded model names and medium reasoning through a compatible API. Equal reasoning computation across services is not assumed.

Metric	Qwen-4B	Qwen-9B	Qwen-27B	Claude	Gemini	GPT
Complete-output coverage	1136/1260	1015/1260	1251/1260	1260/1260	1260/1260	1260/1260
Narrative: complete beliefs	26/895	88/830	864/900	895/900	875/900	836/900
Structured: complete beliefs	33/900	165/875	867/900	880/900	895/900	865/900
Direct planning	107/815	281/545	768/895	582/900	900/900	669/900
Text beliefs: planning	255/855	424/785	759/885	688/900	900/900	712/900
Symbolic beliefs: planning	281/810	400/775	801/890	713/900	900/900	729/900

Appendix

Table 7: CalBench scores and coverage over all complete outputs. The main table compares the common support of all six configurations.

Metric	Qwen-4B	Qwen-9B	Qwen-27B	Claude	Gemini	GPT
Narrative beliefs	0/76	0/76	70/76	76/76	73/76	60/76
Structured beliefs	0/85	2/85	79/85	81/85	84/85	72/85
Direct planning	2/29	10/29	23/29	9/29	29/29	17/29
Text beliefs: planning	6/63	18/63	44/63	33/63	63/63	41/63
Symbolic beliefs: planning	5/53	15/53	40/53	35/53	53/53	34/53

Appendix

Table 8: Strict CalBench source-task accuracy. All five agents must be correct in both PUBLIC and PRIVATE. Denominators count common source tasks with complete outputs in both conditions.

Metric	Qwen-4B	Qwen-9B	Qwen-27B	Claude	Gemini	GPT
WORLD_ACCESS	128/180	170/180	180/180	180/180	180/180	180/180
BELIEF_NL	179/180	166/180	180/180	180/180	180/180	180/180
BELIEF_STRUCT	180/180	175/180	180/180	180/180	180/180	180/180
GIVEN_NL	171/180	157/180	177/180	180/180	180/180	180/180
GIVEN_SYMBOL	162/180	155/180	178/180	180/180	180/180	180/180
DIRECT	163/180	109/180	179/180	180/180	180/180	180/180

Appendix

Table 9: Complete-output coverage by CalBench interface (180 scheduled per interface). Incomplete outputs are excluded from capability scores.

Metric	Qwen-4B	Qwen-9B	Qwen-27B	Claude	Gemini	GPT
Native valid coverage	158/158	154/158	158/158	158/158	158/158	158/158
Direct task success	5/43	2/42	23/43	40/43	39/43	40/43
Given belief task success	8/43	1/41	32/43	39/43	41/43	42/43
Model route entries	264/484	295/484	299/484	453/484	484/484	440/484
Given route entries	256/484	292/474	294/484	333/484	484/484	442/484
Model route whole table	0/36	0/36	1/36	25/36	36/36	22/36

Appendix

Table 10: Exact counts over all technically valid native jobs. Coverage differs by configuration; the main table uses paired common configurations.

Figure 8: Planning outcomes on matched stories. Direct, textual-belief, and symbolic-belief rows each contain 335 proposals. Incomplete structure means the original answer lacks a scorable plan, excluding recoverable fences and surrounding prose.

Metric	Qwen-4B	Qwen-9B	Qwen-27B	Claude	Gemini	GPT
Direct narrative	44/335	182/335	299/335	232/335	335/335	271/335
Given textual beliefs	107/335	177/335	280/335	259/335	335/335	261/335
Given symbolic beliefs	141/335	179/335	302/335	283/335	335/335	283/335

Appendix

Table 11: Matched-story planning controls: 67 stories complete in all three interfaces and all six configurations, with five proposals per story.

Figure 9: Supplied-belief controls on the same 67 stories and 335 agent proposals. All conditions target complete subjectively optimal schedules.

Figure 10: Post-action belief accuracy, averaged equally over 35 common route configurations. Circles denote model-selected routes and diamonds supplied routes. Main-table entry accuracy instead pools all 474 entries.

Figure 11: PRIVATE minus PUBLIC in the number of optimal agents on 29 common source-task pairs. Cells count tasks. Zero change need not mean correctness. Agents are repeated measurements, not independent tasks.

Script	Success/43	Goals met	Steps
Belief-first	42	189/190	4710
World-first	13	122/190	4525
Exploration-first	39	186/190	5369

Appendix

Table 12: Shared deterministic references. Each strategy uses the same tasks and native preparation states for 43 episodes. Step totals include preparation. Scripts are not counted again for each model configuration.

Holder	Reference	Model	Update received
A	1	B	None
B	B	B	Own change
C	1	B	None
D	B	B	Private message
E	1	B	None

Appendix

Table 13: Five agents’ beliefs about B’s slot 8 in the actual case. The number 1 is a displacement cost, and the string B means immovable. Model values come directly from the complete structured response.

Agent	Meeting slots	Cost	Opt.	Belief feasible	World feasible
A	8,3,12,2,5	14	14	Yes	No
B	8,3,12,2,5	—	15	No	No
C	8,3,12,2,5	14	14	Yes	No
D	12,3,2,7,5	22	15	Yes	Yes
E	8,3,12,2,5	14	14	Yes	No

Appendix

Table 14: Complete proposals for b032 PRIVATE. Slots correspond to meetings 1–5. Cost is total displacement under the subjective calendar, and Opt. is the reference minimum.

Figure 12: Belief granularity on 81 common PRIVATE stories: 32400 entries, 405 complete agent states, and 486 stale entries. Copying the world scores 98.5%, 20.0%, and zero, respectively. The construction-wide counts in the text use all 90 source tasks.

Configuration	Optimal proposals
Qwen3.5-4B	0/19
Qwen3.5-9B	0/19
Qwen3.5-27B	13/19
Claude Opus 5.5	10/19
Gemini 3.1 Pro Preview	19/19
GPT-6.1 Sol	10/19

Appendix

Table 15: Subjective optimality on 19 strict-cost-separation agent cases from six source tasks within common DIRECT support. All cases retain their original task constraints and scoring.

Theory of Mind

Theory of Mind studies how people and AI infer others’ beliefs, goals, and knowledge to predict behavior and coordinate interaction.

See all 92 papers

OmniToM: Benchmarking Theory of Mind in LLMs via Explicit Belief Modeling
May 25, 2026Adam Bawatneh, Sagar Sapkota, Amrit Singh Bedi +2LLM Evaluation Theory of Mind
University of Central Florida, Orlando, Florida, USA
EnactToM: An Evolving Benchmark for Functional Theory of Mind in Embodied Agents
May 11, 2026Gurusha Juneja, Dylan Lu, Saaket Agashe +7Multi-Agent Coordination Theory of Mind
University of California, Santa Barbara · 2King’s College London · 3Cisco Research
Who Knows What - and What Does It Support? Perspective-Bounded Evidence Calibration for Multimodal Theory of Mind
Oct 27, 2026Ziqi Liao, Jiaqi Shen, Yiming Na +3Theory of Mind MLLM Reasoning
Jilin University · National Taiwan University

May 25, 2026J/K move · Enter open · S save

OmniToM: Benchmarking Theory of Mind in LLMs via Explicit Belief Modeling

Adam Bawatneh, Sagar Sapkota, Amrit Singh Bedi, Santu Karmaker+1

University of Central Florida, Orlando, Florida, USA

LLM Evaluation Theory of Mind

Open paper AskPDF arXiv

Theory of Mind (ToM), the ability to infer others' knowledge, intentions, and emotions, is commonly evaluated in large language models (LLMs) using end-point question answering, where performance is judged solely by the final answer to a social reasoning query. This paradigm obscures whether the model actually constructs the underlying mental-state representations required for robust reasoning, particularly in scenarios involving divergent, evolving, or mistaken beliefs. In order to address this research gap, we introduce OmniToM, a benchmark that directly evaluates these representations by requiring explicit modeling of belief structures for all relevant actors within a narrative. These structures are composed of belief propositions: minimal statements of what an actor takes to be true about the world or another actor's mental state, allowing knowledge, intentions, emotions, and false beliefs to be analyzed in a common format. Models are evaluated in two stages: Stage 1: Belief Extraction, which extracts from the story the beliefs relevant to its social dynamics, and Stage 2: Belief Labeling, which assigns each belief a seven-dimensional schema label covering recursive order, truth status, knowledge access, explicitness, content type, mental source, and context. Built from 895 stories from the existing ToMBench story corpus and augmented with 22,343 labeled belief propositions, OmniToM uses a human-calibrated LLM-assisted annotation pipeline. Across diverse models in zero-shot evaluation, OmniToM reveals an actor-specific belief-tracking bottleneck: current LLMs struggle with the knowledge-access and representational decisions required to transform narrative facts into actors' beliefs and shared mental states.

Abstract

Theory of Mind

Explore similar work

OmniToM: Benchmarking Theory of Mind in LLMs via Explicit Belief Modeling

EnactToM: An Evolving Benchmark for Functional Theory of Mind in Embodied Agents

Who Knows What - and What Does It Support? Perspective-Bounded Evidence Calibration for Multimodal Theory of Mind