PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2610.34759v1cs.CL

TL;DR

Many ToM evaluations test whether models infer others’ beliefs, but do not establish whether they use those beliefs in tasks with verifiable outcomes. This paper extends household collaboration and scheduling with asymmetric information and separates state recovery from state use. Across configurations, belief inference, fixed decisions, and task success diverge.

  • Problem

    The evaluation gap is connecting personal-state recovery to decision use and independently verifiable task outcomes without treating one as a substitute for another.

    Source: §1 · p. 1

  • Method

    The authors extend household and scheduling tasks with asymmetric information and use independent interfaces and supplied-state controls to separate recovery, decisions, and execution.

    Source: §1 · p. 2, §8 · p. 9

  • Results

    Belief inference, fixed decisions, and task success diverge across configurations, including near-perfect fixed decisions alongside weak long-horizon coordination.

    Source: §1 · p. 2, §7 · p. 8

  • Takeaways & Limitations

    Evaluating collaborative ToM requires checking what models believe others know, the plans they construct from those beliefs, and the consequences of execution.

    Source: §8 · p. 9

    Each evaluation uses one fixed seed without independent repeated runs, so results establish neither statistical significance nor pure parameter-scaling effects.

    Source: Limitations · p. 9

Abstract

from arXiv · show

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.

1 Introduction

The paper connects theory-of-mind inference to decisions and independently verified task outcomes. It extends household collaboration and scheduling tasks with asymmetric information and separates state recovery from state use.

1 Introduction

  • In the illustrated CalBench case, the same proposal is optimal for uninformed A but infeasible for B, and both proposals fail in the actual world.Native replay and subjective solvers distinguish perspective-specific correctness from success in the world.Source: Figure 1 · p. 1
  • The evaluation problem is to connect personal-state recovery, decision use, and independently verifiable outcomes without treating any one as a substitute for the others.Story-question benchmarks test perspective preservation, while producing a complete constraint-satisfying plan requires a different capability.Source: §1 · p. 1
  • The authors extend C-WAH household collaboration and CalBench scheduling with asymmetric information while preserving their goals, constraints, and validators.The extensions add personal observation and communication histories.Source: §1 · p. 2
  • The protocol separates five targets—world, access, belief, decision, and execution—and uses supplied evidence or beliefs to diagnose recovery versus use.Independent requests prevent errors in one answer from mechanically propagating to another.Source: §1 · p. 2
  • Strong belief inference can coexist with non-optimal planning, while near-perfect fixed decisions can coexist with weak long-horizon coordination.The introduction presents both patterns as results of its six-configuration analysis.Source: §1 · p. 2

2 Related Work

Prior work spans mental-state inference, functional and interactive theory of mind, and verifiable task evaluation. This paper combines these strands through complete plans assessed under personal constraints and native execution.

Mental-state inference

  • Mental-state benchmarks test belief, information access, perception, and related states, while this work adds complete plans under personal constraints and their execution.Examples include ToMi, FANToM [Kim et al., 2023], and BigToM [Gandhi et al., 2023].Source: §2 · p. 2

Functional and interactive ToM

  • Functional and interactive ToM work studies prediction, collaboration, planning, or task completion; this paper pairs interaction with solver-verified scheduling and supplied-state comparisons.The cited examples include LLM-Coordination, EnactToM [Juneja et al., 2026], and household collaboration tasks [Puig et al., 2021; Zhang et al., 2024].Source: §2 · pp. 2–3

Verifiable task evaluation

  • Verifiable-task benchmarks check plans or interactive outcomes, while CalBench adds private calendars and displacement-cost optimization; this study retains native constraints and validators.The discussion distinguishes personal evidence from world knowledge when assessing perspective-dependent plans.Source: §2 · p. 3

3 Problem Formulation

The problem is first-order belief attribution and use under asymmetric information: agents can hold different beliefs about the same world because they receive different evidence. The formulation separates perspective-sensitive planning from successful joint action.

3 Problem Formulation

  • A model must distinguish its own knowledge from each agent’s, retain beliefs without new evidence, and use those perspectives in decisions.The focus is first-order belief attribution under asymmetric information.Source: §3 · p. 3
  • C-WAH tests belief maintenance through actions and observations, while CalBench checks complete plans against explicit constraints and optimal costs.Both tasks supply standards independent of the model’s verbal explanation.Source: §3 · p. 3
  • A plan can respect an outdated belief yet fail physically; subjective planning and native collaboration therefore test perspective use and progress toward a shared goal separately.Silently replacing an agent’s belief with world truth tests neither operation.Source: §3 · p. 3

4 Evaluation Framework

The framework separates behavioral targets, reference states, diagnostic controls, and task execution. It uses independent questions and native task validators to distinguish what models know from how they act on it.

4.1 Why Five Targets?

  • World, access, belief, decision, and execution locate different behavioral errors rather than representing five internal modules or a difficulty ordering.The sequence follows information dependencies from understanding events through acting and receiving new evidence.Source: §4.1 · p. 3

4.2 Evidence and Reference States

  • Reference states follow task observation and delivery rules: the model reads all events, but an agent’s access state contains only evidence that agent obtained.In C-WAH, native physical graphs, personal observations, action feedback, and message delivery support rule-based references.Source: §4.2 · pp. 3–4
  • C-WAH access retains obtained positive location assertions even when later superseded, whereas belief reflects the current state after events.CalBench updates owners and actual recipients after public or private calendar changes; uninformed agents retain earlier calendars.Source: §4.2 · p. 4

4.3 Independent Interfaces and Controls

  • Independent interfaces test world/access, beliefs, decisions, and supplied-access conditions, with each request receiving its own context and reference target.C-WAH also includes spatial facts and public-map routes.Source: §4.3 · p. 4
  • Supplied access tests belief recovery while retaining temporal integration; supplied beliefs remove state recovery but preserve perspective-specific decisions.Narrative versus structured histories and textual versus symbolic beliefs test representation changes.Source: §4.3 · p. 4
  • Independent controls isolate whether models recover beliefs from evidence or select appropriate actions when beliefs are supplied; native continuation tests sustained coordination.Continuation lets decisions change subsequent history while models take responsibility for information acquisition and action selection.Source: §4.3 · p. 4

4.4 Task Decisions and Execution

  • C-WAH fixed decisions ask for an object’s source in an actor’s belief, while continuation allows assignments, messages, and exploration within the task budget.Route interventions compare model-selected and supplied public-graph paths from matched checkpoints, then score beliefs against each trajectory.Source: §4.4 · p. 4
  • CalBench evaluates five-meeting schedules across five agents and sixteen slots, checking proposals against subjective displacement-cost optima and actual-world feasibility.Tied optima are accepted, and a subjectively correct plan need not be feasible in the actual world.Source: §4.4 · p. 4

5 Experimental Setup

The experiments evaluate six model configurations on household collaboration and calendar scheduling, using matched tasks and interfaces to separate perspective recovery from decision-making and execution. Scoring distinguishes complete beliefs and feasible, minimum-cost plans from incomplete outputs.

Configurations

  • Six configurations share frozen inputs, references, and semantic scoring, while Qwen sampling is fixed and closed-service sampling controls are unspecified.Qwen output limits differ between C-WAH and CalBench.Source: §5 · p. 4

Tasks and common support

  • C-WAH comparisons match 40 continuation and 35 route configurations, with a 250-step episode budget and shared deterministic-script references.The task set has 43 configurations, 218 event-selected nodes, and nine fixed-request interfaces.Source: §5 · p. 5
  • CalBench comparisons use 916 common complete requests, while public/private analyses additionally match both variants of each source task.The 90 tasks comprise uniform-cost and tiered-cost cases; coverage varies across models and is reported separately.Source: §5 · p. 5

Scoring

  • C-WAH scores person–object entries and full decisions; CalBench requires all 80 belief entries or all five meetings for a feasible minimum-cost plan.Incomplete outputs affect coverage rather than capability scores, and repeated entries are summarized with paired counts and configuration means.Source: §5 · p. 5

6 Results and Analysis

Across household collaboration and calendar scheduling, belief recovery, decisions, and task execution do not reliably coincide. Results show distinct gaps in attributing evidence, using beliefs to act, tracking perspectives after actions, and planning under subjective constraints.

6.1 Recovering Personal Evidence

  • Qwen-4B recovers 93.1% of household world entries but only 33.8% of personal access entries, showing that scene understanding and evidence attribution differ.Supplied access raises household belief accuracy from 49.3% to 60.2% for 4B, 60.3% to 69.0% for 9B, and 62.8% to 91.4% for 27B.Source: §6.1 · p. 5, Table 1 · p. 6
  • Claude errs on 35.7% of no-new-evidence entries and 7.6% of updates, while GPT errs on 4.4% and 15.9%, respectively.Retention and updating expose different residual weaknesses despite both models benefiting from supplied evidence.Source: §6.1 · pp. 5–6

6.2 Using Beliefs in Decisions and Action

  • With supplied beliefs, Qwen-27B household success rises from 22/40 to 30/40, but Qwen-4B rises only from 5/40 to 8/40; Qwen-9B succeeds in 2/40 and 1/40.Despite near-perfect fixed supplied-belief source decisions, 9B issues no assignments in 35/40 episodes in each continuation condition and instead explores extensively.Source: §6.2 · p. 6, Figure 5 · p. 7
  • On 43 shared household tasks, belief-first scripts complete 42 episodes, world-first scripts 13, and exploration-first scripts 39.Closed configurations complete 37–40 paired episodes, but supplied beliefs do not uniformly help: Claude changes from 38 to 37.Source: §6.2 · p. 7

6.3 Tracking Perspectives After Action

  • All supplied-route transports succeed on 35 common configurations, yet no Qwen configuration has an entirely correct post-action belief table; pooled-entry accuracy is 53.0%, 61.6%, and 61.0%.Claude’s belief accuracy is 93.9% after model-selected routes versus 69.0% after supplied routes, while Gemini is perfect in both.Source: §6.3 · p. 7

6.4 Planning Under Subjective Constraints

  • Claude recovers 455/455 agent beliefs but produces only 309/455 subjectively optimal proposals (67.9%), while Qwen-27B recovers 440 beliefs and produces 400 optimal proposals.Across matched stories, models do not share one ranking for belief recovery and planning.Source: §6.4 · p. 7, Figure 6 · p. 7
  • For Qwen-4B, correct plans rise from 44/335 directly to 107/335 with textual beliefs and 141/335 with symbolic beliefs, while Qwen-9B scores 182, 177, and 179.For 9B, textual beliefs increase feasible plans from 192 to 237 while optimal plans fall from 182 to 177; 60 feasible plans are suboptimal.Source: §6.4 · p. 8
  • Copying world calendars yields 98.5% of entries but only 20.0% of complete agent states and zero stale entries on common PRIVATE cases.A privileged solver reference is subjectively optimal in 466/485 direct-planning cases (96.1%), so aggregate success alone does not establish perspective use.Source: §6.4 · p. 8

7 Discussion

The discussion identifies separate gaps in perspective recovery and effective planning, and argues that evaluation must pair state questions with perspective-sensitive plans and native task outcomes.

7 Discussion

  • A correct state or fixed decision does not ensure effective planning and coordination, while successful actions do not establish complete mental-state understanding.Independent state questions, perspective-sensitive plans, and native consequences provide complementary evidence.Source: §7 · pp. 8–9
  • Reference beliefs remove state uncertainty, but models still need to preserve subjective constraints, choose goals, and execute productive commands.In household continuation, communication and exploration can change what an executor knows; in scheduling, adding an unreceived update changes the problem being solved.Source: §7 · p. 8
  • Entry accuracy, whole-agent accuracy, subjective feasibility, optimality, and native outcomes measure distinct levels of recovery, planning, and task success.A successful action does not establish complete mental-state understanding, and a correct mental-state answer does not establish competent collaboration.Source: §7 · pp. 8–9

8 Conclusion

The conclusion presents a task-grounded evaluation that connects asymmetric-information inference to decisions and execution. Across six configurations, these abilities diverge in both directions.

8 Conclusion

  • Across six configurations, belief inference, decision use, and task success diverge in both directions.The protocol supplies evidence to probe state recovery, beliefs to probe state use, and native execution to test whether either advances task progress.Source: §8 · p. 9

Limitations

The evaluation uses six service configurations, 43 household configurations across seven houses, and 90 calendar tasks, with one fixed seed and no independent repeated runs.

Limitations

  • Six service configurations, 43 household configurations across seven houses, and 90 calendar tasks use one fixed seed without independent repeated runs.Reasoning budgets, sampling controls, and serving hardware differ; the results establish neither pure parameter-scaling effects nor statistical significance across independent runs.Source: Limitations · p. 9

Ethical Considerations

The paper describes simulated tasks and a perspective example, then details belief categories and model settings, including how answers are scored.

Ethical Considerations

  • The evaluation uses simulated households and controlled calendars, without new human-subject experiments or personal calendar collection.Task scores cannot establish reliability in real collaboration or replace consent and verification.Source: Ethical Considerations · p. 9

A A Complete Perspective Example

  • In the example, A receives both private updates while B and C know only their own changes, so their earliest feasible proposals differ despite a single globally feasible slot.The reference proposals are A: 16:00, B: 15:00, and C: 14:00.Source: A A Complete Perspective Example · pp. 10–11

B Layer-wise Results and Belief Categories

  • Table 4 separates current, unknown, and stale reference beliefs, and an unknown state is a target to predict rather than a missing sample.Figure 7 distinguishes 2,248 entries to retain without new evidence from 170 actual updates; it excludes 434 confirmation entries.Source: B Layer-wise Results and Belief… · p. 11, Figure 7 · p. 11

C Model Settings and Output Targets

  • The model settings specify task-specific inputs and targets, while semantic scoring distinguishes missing target structures from wrong content.For CalBench, belief correctness requires all 80 entries; scheduling requires five valid meeting slots, native feasibility, and minimum subjective cost.Source: C Model Settings and Output Targets · p. 11

D Complete Coverage and Source-task Aggregation

This section defines the task coverage, matching rules, and scoring procedures for the household and calendar evaluations. It also specifies how private information, subjective plans, and native task execution are represented and checked.

D Complete Coverage and Source-task Aggregation

  • All six configurations completed 1,962 fixed C-WAH requests; five covered 158/158 native jobs and Qwen-9B covered 154/158.Main native comparisons use 40 configurations with valid continuation pairs and 35 with valid route pairs across all models.Source: D Complete Coverage and Source-task… · pp. 11, 13

E Task Semantics of Planning Errors

  • Planning outcomes distinguish subjective optimality, feasible but suboptimal plans, subjective infeasibility, and incomplete structure.A feasible plan meets constraints but misses cost minimization; subjective infeasibility violates constraints known to the target agent.Source: E Task Semantics of Planning Errors · p. 13

F Matched-story Supplied-belief Controls

G Joint Decisions, Routes, and Communication

  • Joint-decision and route comparisons use distinct outcomes and matched statistical units, including 35 common route configurations and 29 complete source-task pairs.The route analysis tracks post-action beliefs; communication comparisons count changes in optimal agents, where zero change need not mean correctness.Source: G Joint Decisions, Routes, and… · p. 13, Figure 10 · p. 15, Figure 11 · p. 16

I.2 CalBench Story Construction

  • Calendar stories preserve each source task’s final world calendar while varying who receives private updates; PUBLIC delivers both updates to all agents, whereas PRIVATE limits recipients.PRIVATE leaves owners and the intermediary with different updates, while two other agents retain their initial calendars.Source: §I.2 · p. 16

J.1 C-WAH Observations, Messages, and Search

  • C-WAH beliefs change through actions, observations, and delivered messages; direct observation can override a conflicting message, and an unsuccessful completed search can yield UNKNOWN.Without new evidence, beliefs persist; UNKNOWN means the source is undetermined, not that the object does not exist.Source: §J.1 · p. 17, §J.2 · p. 17

J.2 CalBench Reference States and Subjective Planning

  • CalBench subjective planning groups identical agent calendars, then checks proposals against subjective constraints and costs as well as actual-world feasibility.Native replay validates relocations and meetings, while independent optimization and replay must agree on reference costs; tied optima count as optimal.Source: §J.2 · p. 17, §L.1 · pp. 21–22

L.1 Strict State and Plan Metrics

  • Scores use complete outputs and paired task support; repeated agents and observations are not treated as independent tasks.Entry accuracy pools scored entries, while configuration averages weight configurations equally; wrong complete answers remain in paired-analysis denominators.Source: §L.1 · p. 21, §L.2 · p. 22

M An Actual Case: Beliefs and Task Verification

An example from a PRIVATE calendar story shows that high entry accuracy can coexist with incorrect personal perspectives. Native plan validation further distinguishes infeasible proposals from feasible but suboptimal ones.

M.1 Information Attribution in Belief Answers

M.2 Complete Plans and Native Verification

  • A, C, and E proposed cost-14 plans that were subjectively optimal, while B’s identical plan violated a known constraint and D’s feasible plan cost 22 against an optimum of 15.Native replay rejects moving B’s immovable slot-8 commitment under B’s information, although that move remains valid under A’s calendars.Source: §M.2 · p. 23, Table 14 · p. 23
  • Belief and planning outputs were independent requests, so the belief response was not passed to the direct-planning response.The example therefore separates belief answers, perspective-specific planning, and actual execution conditions.Source: §M.2 · p. 23

N.1 Copying the World Does Not Establish Perspective Recovery

Copying the actual world calendar can produce high entry accuracy without recovering complete personal beliefs. World-optimal plans are often subjectively optimal, but strict cost differences show cases where they cannot be.

N.1 Copying the World Does Not Establish Perspective Recovery

  • Copying world calendars scores 99.25% across 72,000 entries but only 540/900 complete agent states, including 90/450 in PRIVATE.In PRIVATE, only the intermediary receives both updates; shared facts dominate entry accuracy while complete states and stale entries test information boundaries.Source: §N.1 · p. 23

N.2 World Plans and Perspective-sensitive Planning

  • Across 180 stories, world-optimal plans are subjectively feasible in 900/900 cases and optimal in 863/900 (95.9%); on PRIVATE stories they are optimal in 413/450 (91.8%).These plan checks do not establish recovery of each agent’s full personal state.Source: §N.2 · p. 23
  • On six-model common DIRECT support, world plans are optimal in 466/485 cases; in the remaining 19 cases, subjective optimal cost is strictly lower than world-optimal cost.This subset shows where the world-plan reference fails, not that every successful plan requires an explicit complete belief representation.Source: §N.2 · p. 23, Table 15 · p. 24

N.3 State Representation and Planning

  • Matched textual and symbolic supplied-belief controls both provide correct states, so their comparison concerns representation and downstream use rather than direct measurement of internal reasoning.Both controls leave the planner to construct each agent’s subjective feasible set and minimize cost.Source: §N.3 · p. 24
Loading 2610.34759v1…