False Belief Test (FBT)

• False Belief Test (FBT) is a standard experimental measure of theory of mind that assesses belief attribution by testing if observers can predict behavior based on others’ false beliefs.
• FBT employs canonical paradigms, including location-change, Smarties, and Sally-Anne tasks, using explicit and implicit cues to evaluate inferential reasoning.
• FBT research integrates empirical studies, model-based evaluations, and logical frameworks to advance understanding of cognitive processes and inform AI applications.

The False Belief Test (FBT) is a narrow but standard measure of belief attribution and a core probe of theory of mind (ToM): it asks whether an observer can represent that another individual may hold a belief about an object’s location that differs from reality, and use that belief to predict behavior. In its canonical location-change form, a character places an object in a Start location, another character moves it to an End location, and the diagnostically relevant manipulation is whether the target witnessed the move; common probe formulations are explicit, as in “$X$ thinks…,” or implicit, as in “$X$ goes to get…” or “$X$ looks for…”. Closely related paradigms include the Smarties task, the Sally-Anne task, and the Unexpected Contents Task (Trott et al., 2022, Trott et al., 17 Feb 2026, Brauner, 2013).

1. Canonical paradigms and task structure

Across the literature represented here, the FBT is operationalized through a small set of recurrent paradigms that differ in surface form but share the same inferential core: the target’s epistemic access diverges from the actual state of the world, and the respondent must answer from the target’s perspective rather than from reality. In the canonical linguistic version, the object starts in one location, is moved to another, and the target either sees or does not see the move. In Smarties-style and Unexpected Contents variants, the conflict is between appearance or labeling and hidden contents. In Sally-Anne, the conflict is generated by a location transfer performed outside the protagonist’s awareness (Trott et al., 17 Feb 2026, 2406.14737, Brauner, 2013).

Variant	Core manipulation	Diagnostic response
Location-change FBT	Object moves from Start to End; target is present or absent	Where the target thinks the object is, or where the target will search
Smarties	Tube suggests Smarties but contains pencils	What another person, or one’s earlier self, would think is inside
Sally-Anne	Sally leaves; Anne moves the marble from basket to box	Where Sally will look for the marble
Unexpected Contents	Label suggests one content, actual contents differ	What another person thinks is inside, or which item the character prefers

Several recent LM studies use a balanced linguistic battery of 12 scenarios, each rendered in 16 versions, for 192 passages, crossing Knowledge State, Knowledge Cue, and position controls such as First Mention and Recent Mention. The explicit/implicit cue distinction is central: explicit probes contain a propositional-attitude verb such as thinks or believes, whereas implicit probes ask about action, such as where the agent goes to get or looks for the object (Trott et al., 2022, Trott et al., 17 Feb 2026, Kouwenhoven et al., 4 Mar 2026).

2. Inferential prerequisites and cognitive interpretation

A major recent reformulation treats FBT success as the product of upstream inferential components rather than as a monolithic competence. On that view, false-belief reasoning depends first on perception inference, namely inferring who perceived each event or utterance, and then on perception-to-belief inference, the “seeing leads to knowing” step by which perceptual access is converted into a belief state. In this decomposition, the defining false-belief mechanism is that if a character did not witness an event, that character should not update belief on the basis of that event. The same work identifies inhibitory control as a key precursor: successful performance requires suppressing salient but irrelevant reality-based information when reasoning about another person’s belief (Jung et al., 2024).

This inferential decomposition aligns with formal analyses that describe FBT as a problem of perspective shift. In hybrid-modal treatments, the relevant operation is to reason at another local point of evaluation—another person or another time—and then return to the current perspective. The Smarties task appears in two versions, one involving a shift to another person and the other to an earlier time, and both are shown to have the same underlying logical structure. The Sally-Anne task adds a stronger temporal component: it requires a principle of inertia, according to which a belief is preserved over time unless there is belief to the contrary (Brauner, 2013).

A plausible implication is that the FBT is not exhausted by the final choice between two locations or two contents. It also tests whether the reasoner can keep perception, perspective, and belief persistence distinct while resisting interference from the true state of affairs.

3. Empirical behavior in humans and LLMs

In a preregistered linguistic FBT administered to both humans and GPT-3, GPT-3 text-davinci-002 showed reliable sensitivity to the implied knowledge state of the protagonist, with 74.5% accuracy when prediction was defined by the higher-probability Start-versus-End completion. A linear mixed-effects analysis showed that including Knowledge State improved fit, $\chi^2(1) = 18.6, p < .001$, and adding the Knowledge State $\times$ Knowledge Cue interaction also improved fit, $\chi^2(1) = 20.6, p < .001$. Human retained accuracy was 82.7%, and human responses remained sensitive to Knowledge State even after conditioning on GPT-3 log-odds, $\chi^2(1) = 30.4, p < 0.001$, indicating that GPT-3 did not explain the full extent of human behavior (Trott et al., 2022).

A broader open-weight replication extended this design to 41 open-weight models from five families using the same 192 total stimulus passages. 14 of the 41 models, about 34%, showed statistically reliable sensitivity to Knowledge State after Holm–Bonferroni correction, with likelihood-ratio statistics ranging from 11.9 to 21.4. When all models were analyzed together, Knowledge State remained significant, with $\beta = -1.5$, $SE = 0.37$, $p < .001$, and $X$0. The best models reached 74.5% accuracy, the mean across all 41 models was 56.4%, and the best psychometric predictive power reached $X$1. Parameter count, rather than training tokens or instruction tuning alone, was the significant predictor of both mean accuracy and PPP (Trott et al., 17 Feb 2026).

Developmental analyses over training checkpoints refine this picture. In the Olmo2 and Pythia suites, above-chance FBT performance depends on both model size and sufficient training volume, with a size-by-training interaction of $X$2. The behavior emerges late in pretraining rather than early. Situation modeling—the ability to report factual properties of the described scene—generally precedes and exceeds FBT accuracy, but the resulting representations remain partially incoherent: for example, Olmo2 13b can still be misled by the target agent’s knowledge state and by non-factive verbs when queried about the antagonist, who always knows the true location (Rivière et al., 26 Jun 2026).

Another large-scale open-model study identifies a strong cross-over effect tied to propositional-attitude wording. In a Bayesian multilevel logistic regression, the interaction between knowledge state and knowledge cue was $X$3: explicit cues help False Belief but hurt True Belief, whereas implicit cues show the opposite pattern. Scaling model size improves performance on average, $X$4, but not strictly, since False Belief and True Belief move in opposite directions with scale. Vector steering further isolates a think vector as a causal driver of this behavior (Kouwenhoven et al., 4 Mar 2026).

4. Diagnostic decompositions and targeted interventions

Perception-augmented benchmarks make the upstream structure of FBT explicit. Percept-ToMi is built from 150 story-question pairs for each of the four ToMi question types—first-order true belief, first-order false belief, second-order true belief, and second-order false belief—while Percept-FANToM contains 220 conversations and 735 sets of questions. For perception inference, models must output, in JSON format, which characters perceived each sentence or utterance; for perception-to-belief inference, models receive ground-truth perception annotations and then answer the original ToM question. Across eight state-of-the-art LLMs, average perception inference accuracy was 0.781 on Percept-ToMi and 0.926 on Percept-FANToM, with almost no true-versus-false-belief gap at this precursor stage. By contrast, models remained weak at converting perceptual access into belief, especially in false-belief settings, and ToM performance showed weak or near-zero correlation with perception inference but positive correlation with perception-to-belief inference (Jung et al., 2024).

The same decomposition motivates PercepToM, a three-stage pipeline consisting of Perception inference, Perspective context extraction, and Response generation. The extraction stage filters the full context to sentences or utterances perceived by the target character, so that the answer is grounded in “information symmetry” with the character’s belief state. Empirically, the method improves performance across models and datasets, with particularly large gains on false-belief cases: in Table 1, GPT-4 Turbo reaches 1.000 on false belief in ToMi, and in false-belief FANToM Llama-3 70B Instruct rises to 0.147 from near zero vanilla performance. An oracle ablation shows that merely providing ground-truth perception annotations is insufficient; the perspective extraction step itself is beneficial, which the authors interpret as evidence that irrelevant full-context information is actively harmful and that weak inhibitory control is a central bottleneck (Jung et al., 2024).

A complementary diagnostic method, SCALPEL—Selective Comparison of Adversarial Linguistic Prompts to Explain Lacunae—tests why models fail on modified false-belief prompts by making specific implicit inferences explicit. In the Transparent-Access Variation of the Unexpected Contents Task, changing transparency wording alone did not help: the original condition yielded 22.14% for GPT-3.5 and 20.35% for GPT-4, while see-through and see-inside produced no meaningful improvement. Making looking explicit improved both models into the mid-30% range, and making recognition explicit produced the strongest effect: the recognize condition reached 54.28% for GPT-3.5 and 89.64% for GPT-4. The reported interpretation is that failures on altered FBTs may reflect missing commonsense bridges—such as reading implying looking, or looking implying recognizing—rather than a wholesale inability to represent beliefs (2406.14737).

5. Logical and formal frameworks

Logical work on false belief abstracts away from benchmark wording and treats false belief as a modal or centered reasoning phenomenon. In hybrid-modal logic, the key machinery consists of nominals and the satisfaction operator $X$5, which permit explicit reasoning at a named point—another person, another time, or another local perspective. The crucial proof rule is $X$6, which allows one to reason at a hypothetical point and then shift back. On this analysis, the temporal-shift and person-shift versions of Smarties have exactly the same logical structure, whereas Sally-Anne requires both person shift and temporal belief persistence via the inertia principle (Brauner, 2013).

A separate line of work studies a primitive false-belief operator $X$7, read as “the agent is wrong about $X$8.” Its semantics is given by

$X$9

This program solves the long-open problem of axiomatizing the transitive logic of false belief by using an almost definability schema,

$X$0

and extends the results to Euclidean logics and to radical ignorance via interdefinability of the corresponding operators (Fan, 2024).

In the intuitionistic setting, false belief is modeled on neighborhood frames with the clause

$X$1

The basic system $X$2, containing intuitionistic propositional logic, the axiom $X$3, and an extensionality rule, is proved sound and strongly complete with respect to the class of all $X$4-frames, and stronger variants add closure under intersection or supplementation (Witczak, 2020).

Related work on knowledge-wh and false-belief sensitivity compares four bundled operators corresponding to MS-true belief, MS-true belief with FS, MS-knowledge, and MS-knowledge with FS. The main logical conclusion is not that FS is mandatory, but that it is structurally costly: it breaks monotonicity and weakens positive introspection, even though it captures intuitions about cases where a correct answer is embedded in a broader field of mistaken beliefs (Yang, 2023).

6. Applied false-belief inference and construct-validity debates

FBT-style reasoning also appears in multi-agent applied systems. In a simulated Minecraft urban search-and-rescue scenario with a team of three human players, one player is deliberately given a different legend for marker blocks, inducing a false belief about how markers indicate room state. The resulting AI agent, ToMCAT (Theory of Mind-based Cognitive Architecture for Teams), uses a probabilistic graphical model with latent variables for each player’s original legend, the team’s adopted legend, and time-indexed perception, grounded in field-of-view evidence and marker placements. Approximate online inference is performed with Rao-Blackwellized particle filtering. On false-belief identification, ToMCAT achieved $X$5 in 5-fold cross-validation versus 0.33 chance; on an annotated set it achieved 0.41, while human observers were at 0.33. The system thus extends false-belief inference from toy dyads to real-time coordination with individual and shared mental states (Soares et al., 2023).

At the same time, the FBT is repeatedly described as an imperfect assay. Several papers emphasize that it is a narrow or classic probe rather than a complete measure of ToM; that it may recruit non-ToM capacities; and that, in LM settings, it is vulnerable to contamination, prompt artifacts, lexical cueing, and shallow heuristics. The presence of non-factive verbs such as thinks can bias both humans and models toward false-belief attributions, and in LMs that cue can become a dominant surface feature that outweighs other scenario semantics. Accordingly, high FBT performance is not treated as decisive evidence of robust, human-like mentalizing, and failure on modified tasks is not, by itself, decisive evidence of its absence (Trott et al., 17 Feb 2026, Kouwenhoven et al., 4 Mar 2026, Rivière et al., 26 Jun 2026, 2406.14737, Trott et al., 2022).

A stable conclusion across these lines of work is more modest and more diagnostic. FBT behavior can reveal sensitivity to another agent’s informational access and belief-relevant language, but its interpretation depends on whether perception tracking, perspective filtering, belief updating, and cue sensitivity have been disentangled. On that view, the test remains valuable not as a standalone verdict on ToM, but as a controlled probe whose evidential force depends on how finely its component inferences are analyzed.