{"id":"52b7a1eb-797a-44d7-85c0-83bd6f9186b3","arxiv_id":"2412.13631","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"LLM Theory of Mind benchmarks focus on logical inference at a fixed mentalizing depth, but overlook the prior step of deciding whether and how deep to mentalize, which the paper argues should be measured with interactive, dynamic evaluation.","lead":"This position paper argues that AI benchmarks for Theory of Mind test only the reasoning half of the skill, ignoring the harder question of whether a model should mentalize at all. The authors support this with a small experiment showing that language models can predict the required mentalizing depth much better than they can answer the belief questions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pilot's proxy task does not measure step I: predicting a question's ToM-order label is a separate meta-classification, solvable from syntax alone (identical with/without story), so it cannot establish that failures are Type C rather than Type B.","rationale":"The conceptual core of the paper — that existing LLM ToM benchmarks are static, make the need for ToM obvious from the question, and therefore concentrate on inference at a fixed Depth of Mentalizing while overlooking the invocation step — is defensible and well grounded in the cognitive science literature. The weakness lies in the empirical support offered for the stronger claim that benchmark failures are Type C rather than Type B. The reader's weakest assumption is exactly right: the pilot equates invoking ToM with predicting a question's ToM-order label. I sharpen this by noting that the identical scores with and without the story (Table 1) demonstrate the proxy is solvable from syntax alone, which makes it even less a measure of invocation and more a template-recognition task. A direct way to settle whether errors are Type C or Type B is to condition answers on the correct DoM label; this is the experiment the paper does not run. Because the paper is a position piece, the conceptual argument can survive a weak pilot if the empirical claims are modestly restated, so the reader's CONDITIONAL verdict remains appropriate. I see no reason to move the verdict; the load-bearing concern affects the strength of the supporting experiment, not the viability of the central distinction.","tokens_in":12279,"tokens_out":5665,"duration_ms":54250,"concrete_test":"Take a random subset of HiToM/FANToM questions and condition the answer prompt on the gold ToM order, e.g., 'This is a K-th order ToM question. Answer accordingly.' Compare accuracy against the original unconditioned answers. If accuracy does not improve, Type C (reasoning at the given depth) is supported; if it improves substantially, the original failures are partly Type B and the pilot's proxy is insufficient. As a secondary check, rephrase questions so explicit recursive mental verbs are paraphrased away while semantic depth is preserved; if DoM-label prediction collapses to chance, the proxy is syntax-driven and not a measure of mentalizing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim is that benchmark failures are due to reasoning (Type C), not ToM invocation (Type B). The supporting pilot asks models to predict the ToM-order label of a question, and high accuracy on this label task is taken to mean invocation is not the bottleneck. This proxy is not faithful to step I as defined. Step I is the decision, within a social reasoning task, of whether and at what depth to mentalize; the proxy is an isolated classification task where the prompt explicitly instructs the model to report a ToM order, so it never tests spontaneous invocation. Worse, Table 1 shows identical accuracy with and without the story for every model and setting, indicating the label can be recovered from the question's surface syntax (e.g., counting nested 'thinks that' clauses). This is a shallow template-recognition skill, not evidence of mental-state tracking. The experiment also leaves disconnected: a model can state the correct order in a separate meta-task yet fail to apply it when answering, so observed errors could still be undermentalization during actual reasoning. Appendix A.3 further concedes that gold labels are contestable (e.g., questions about 'dealing with' could plausibly be first- or second-order), weakening the ground truth. Consequently, the conclusion that 'the challenging part of these benchmarks is not the ToM order (=invoking ToM)' is not supported by the reported experiment.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Wagner et al. argue that current AI and LLM Theory of Mind (ToM) research focuses on the second step of a two-step process—performing inference at a given Depth of Mentalizing (DoM)—while largely ignoring the first step, the decision of whether and at what depth to invoke ToM. They survey benchmarks, ToM add-ons, and formal models, and present a pilot experiment on HiToM and FANToM in which LLMs predict the ToM-order label of questions. They conclude from this pilot that failures on these benchmarks are due to reasoning (Type C) rather than to ToM invocation (Type B), and they recommend dynamic, interactive evaluations of ToM invocation.","tokens_in":12570,"tokens_out":5026,"duration_ms":42710,"significance":"If the two-step framing is accepted, the paper makes a valuable conceptual contribution by identifying a gap in LLM ToM evaluation and by connecting the AI literature to cognitive science. The survey of benchmarks, add-ons, and formal models is informative, and the proposed dynamic evaluation is a sensible direction. The paper also gives explicit, falsifiable predictions about where benchmark difficulty lies. However, the central empirical claim is not currently supported by the reported pilot, so the paper's main contribution rests on its conceptual argument and recommendations rather than on the experimental evidence.","major_comments":[{"comment":"The experiment's proxy task does not measure step I as the paper defines it. Step I is the 'meta-decision regarding the DoM' that an agent makes while solving a social reasoning task; the pilot, however, explicitly instructs the model to output the ToM order as an answer, so the model never faces the spontaneous decision of whether to invoke ToM. The fact that Table 1 shows identical accuracy with and without the story for every model and setting strongly suggests that the order labels are predicted from the question's surface syntax rather than from any mental-state tracking. Consequently, the experiment cannot distinguish Type B from Type C errors, and the conclusion that benchmark failures are not due to ToM invocation is not supported by this experiment.","section":"Section 3.2, Appendix A.1–A.2"},{"comment":"The strength of the conclusion is disproportionate to the evidence. The pilot uses 100 examples per benchmark, reports no confidence intervals, provides no code or data, and relies on gold labels that Appendix A.3 itself describes as contestable for some FANToM questions. At most the pilot demonstrates that a closely related meta-classification (label prediction) is easier than the original task; it does not show that invocation is not the bottleneck. The text should either present this as preliminary, suggestive evidence or be substantially expanded, for example with a control condition where the model must spontaneously decide whether a question requires ToM.","section":"Section 3.2, Appendix A.2–A.3"}],"minor_comments":[{"comment":"The title contains a formatting artifact: 'Mind Y our Theory' should read 'Mind Your Theory.'","section":"Title"},{"comment":"The reported FANToM baseline accuracies list '0.426 with GPT-4-turbo, and 0.497 with GPT-4-turbo'; the second model name appears to be a typo.","section":"Appendix A.1"},{"comment":"The assertion that 'the necessity of ToM is made obvious' in existing benchmarks would benefit from concrete examples or a citation to an analysis that demonstrates this property.","section":"Section 3.2"},{"comment":"Making the code, prompts, and sampled data publicly available would allow readers to verify the pilot's results and assess the validity of the proxy task.","section":"Appendix A"},{"comment":"The Limitations section does not mention the pilot's limitations, such as the small sample size, the lack of confidence intervals, and the potential invalidity of the proxy task; a sentence acknowledging these would improve the paper.","section":"Limitations"}],"recommendation":"major_revision","confidential_remarks":"The paper's conceptual contribution is real and the survey is broad, but the empirical pilot is currently too weak to support the paper's central conclusion about Type C errors. If the authors are willing to reframe the pilot as an illustration and soften the claims, a minor revision might suffice; if they insist on the Type C conclusion, the experiment would need to be substantially redesigned. The paper may be more appropriate as a position piece than as an empirical study, and the editor may wish to weigh that fit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things worth knowing about this paper. The conceptual core is solid: separating ToM into step I (deciding whether and at what depth to mentalize) from step II (applying the inference) is a genuine and useful reframing, and the paper is right that most LLM benchmarks and add-ons assume the invocation is already done for you. The survey of benchmarks, add-ons, and formal models is broad and fair, and the formalization via do-calculus is thoughtful.\n\nWhere it wobbles is the experiment. The pilot on HiToM and FANToM asks models to predict the ToM-order label of a question, and then infers that because they can do that while failing the belief questions, the hard part is reasoning, not invocation. That doesn't follow. Predicting the label from the question alone, which the results show is just as accurate as with the story, is a meta-classification task solvable from surface syntax (nested 'thinks that' clauses). It never tests spontaneous invocation, because the prompt explicitly asks for the order. The stress-test note is right on this: a model can state the correct order in a separate meta-task and still under-mentalize during the actual question. Appendix A.3's own admission that gold labels are contestable makes the ground truth shakier.\n\nThat said, the weakness is in the empirical support, not the conceptual argument. The paper is honest about the pilot being a case study, and the recommendations for interactive, dynamic benchmarks are reasonable and align with calls from other groups (Wang et al. 2025; Sclar et al. 2022), which are cited. The lack of code, data, and confidence intervals on a 100-example pilot is a minor additional concern, but the core issue is the proxy's validity.\n\nFor a reader, the value is in the framing: if you are building or evaluating ToM benchmarks, the Type A/B/C error taxonomy and the invocation-vs-inference distinction are worth taking seriously. The pilot should be treated as illustrative only. The paper deserves a serious referee because the conceptual claim is important and the writing is clear; the referee should ask for a better experiment or a more modest conclusion.","headline":"A useful two-step ToM framing for LLM evaluation, but the pilot's proxy task doesn't measure invocation, so the empirical conclusion is overstated.","tokens_in":13077,"tokens_out":2143,"would_cite":true,"duration_ms":18902,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that LLM theory-of-mind benchmarks measure reasoning, not the decision to mentalize, and that a two-step view of ToM (invoke, then infer) is needed.","keywords":["theory of mind","depth of mentalizing","LLM benchmarks","false belief","reasoning vs mentalizing","interactive evaluation"],"falsifier":"Find or build a benchmark in which the question does not reveal the required depth — the model must infer from the social context whether mentalizing is needed — and show that models with high ToM-order prediction accuracy on HiToM/FANToM nevertheless fail to invoke ToM when the context demands it; alternatively, show a model that answers belief questions correctly while unable to predict their ToM order, which would break the claimed link between step I and label prediction.","tokens_in":12090,"feed_emoji":"🧠","tokens_out":5450,"duration_ms":42386,"temperature":0.7,"pith_summary":"This paper argues that current evaluations of theory of mind (ToM) in large language models measure only the second half of a two-step cognitive process. Step I is deciding whether to invoke ToM and at what Depth of Mentalizing (DoM) — the number of nested beliefs to consider; step II is applying the correct belief inference once that depth is fixed. The authors claim that benchmarks like HiToM and FANToM, along with prompting add-ons, treat ToM as a static logic problem, so a model that fails them may be failing at reasoning rather than at mentalizing. A small pilot shows that models can predict the required ToM order of a question (a proxy for step I) while still answering the belief questions incorrectly, which the authors read as evidence that the hard part in these benchmarks is step II. If the distinction holds, the debate about emergent ToM in LLMs needs to be reframed around invocation and adaptive depth, not raw accuracy on false-belief vignettes.","feed_headline":"ToM benchmarks test logic, not mentalizing itself","feed_subtitle":"Models can name the needed mental depth while failing the belief questions, so the hard part is inference.","key_machinery":"The load-bearing object is the two-step decomposition of a ToM task: step I, invocation, sets the Depth of Mentalizing (DoM) — the level $k$ of recursive belief nesting, from $k=-1$ (no mentalizing) to $k\\geq 1$ (nested beliefs about beliefs) — and step II performs the inference conditional on that depth. The empirical instrument is a label-prediction task: given a question (with or without its story), the model must output the ToM order that the benchmark's metadata assigns to it. Accuracy on this sub-task is used as a proxy for whether step I can be performed; the near-perfect performance of GPT-4o on the proxy, combined with much lower accuracy on the original belief questions, is the evidence that benchmark failures are Type C reasoning errors rather than failures to invoke ToM.","core_discovery":"The paper's central claim is that the dominant research paradigm for LLM theory of mind addresses only what the authors call step II — inference at a fixed depth of mentalizing — and therefore mistakes failures of logical reasoning for failures of social cognition. Formally, a ToM task is a distribution $P(Y \\mid X)$ over answers $Y$ given a prompt $X$, decomposed into a meta-decision $do(\\mathrm{DoM}(k))$ that sets the recursion depth, followed by the conditional $P(Y \\mid do(\\mathrm{DoM}(k)), X)$. Because current benchmarks announce the mentalizing demand through templatized questions about clearly described scenarios, the model never has to decide whether ToM is needed or how deep to go. The paper offers a case study on HiToM and FANToM in which GPT-4o and GPT-4o-mini predict the required ToM order nearly perfectly while struggling with the belief questions themselves, supporting the conclusion that the benchmark difficulty lies in step II. The authors therefore call for benchmarks that test whether and at what depth ToM is invoked, in dynamic interactive settings.","pith_inferences":["The paper's proxy equates invocation with predicting a question's metadata label; a stronger test would require the model to choose the depth from the social situation itself — e.g., competition vs. cooperation — without the question revealing it, and then act on that choice.","If the distinction generalizes, earlier 'ToM emergence' claims (e.g., on classic false-belief tasks) may overstate social cognition, since those tasks announce the need for mentalizing; conversely, failures on altered vignettes may be reasoning brittleness rather than absent ToM.","A testable extension: measure whether models adjust their DoM in a dialogue when a speaker is revealed as deceptive or when stakes change; the paper predicts such adaptation is currently missing.","The resource-cost argument suggests a concrete behavioral signature: an interactive agent that can conserve mentalizing in cooperative settings and escalate in competitive ones should outperform a fixed-DoM agent at equal parameter count."],"forward_implications":["If the two-step distinction is correct, high scores on existing ToM benchmarks should not be read as evidence that a model decides when to mentalize; they show inference within an already-announced depth.","Low scores on the same benchmarks should not be read as ToM deficits, since they may be ordinary logical failures; error taxonomies need a Type B (under- or over-mentalizing) category alongside Type C.","ToM add-ons that fix a DoM in advance cannot improve invocation, so their gains are limited to step II; a model that never chooses to mentalize will not be helped by these modules.","Evaluation should move to dynamic, interactive settings where the agent must re-assess its DoM as new information arrives, mirroring cognitive tasks with biological agents.","Because mentalizing has a resource cost, correct invocation is also an efficiency question: models that over-mentalize waste compute, and models that under-mentalize miss deception."],"supporting_citations":[{"why":"HiToM, the higher-order false-belief benchmark whose ToM-order metadata supplies the label-prediction task and whose reported accuracy forms the baseline.","marker":"Wu et al., 2023"},{"why":"FANToM, the conversational benchmark whose belief questions and metadata anchor the empirical pilot.","marker":"Kim et al., 2023"},{"why":"Formalizes Depth of Mentalizing as the recursion level of nested beliefs, the central quantity of step I.","marker":"Barnby et al., 2023"},{"why":"Source of the two-step view of ToM (invocation and inference) that the paper adopts.","marker":"Schaafsma et al., 2015"},{"why":"Cognitive account of ToM as a capacity with separable selection and reasoning components.","marker":"Leslie et al., 2004"},{"why":"Argument that benchmark failures may confound ToM errors with poor or fuzzy logic, directly motivating the paper's error taxonomy.","marker":"Pi et al., 2024"}],"fun_headline_variants":["LLMs know when to mentalize, still fail belief tasks","Theory of Mind benchmarks skip the mentalizing step","Why LLMs fail ToM: it's the inference, not the depth","Benchmarks test logic, not theory of mind's hard part","ToM requires deciding depth, benchmarks give it away"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The empirical argument equates succeeding at step I with predicting the ToM-order label of a question, so if label prediction is not a faithful measure of whether a model would actually decide to mentalize in an interactive setting, the pilot's support for the paper's central claim weakens.","fun_headline_variants_meta":{"raw":{"variants":["LLMs know when to mentalize, still fail belief tasks","Theory of Mind benchmarks skip the mentalizing step","Why LLMs fail ToM: it's the inference, not the depth","Benchmarks test logic, not theory of mind's hard part","ToM requires deciding depth, benchmarks give it away"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000174,"raw_usage":{"total_tokens":1261,"prompt_tokens":900,"completion_tokens":361,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":516,"completion_tokens_details":{"reasoning_tokens":277}},"tokens_in":516,"tokens_out":361,"duration_ms":3971,"temperature":1.0,"reasoning_tokens":277,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:56:07.452535+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Find or build a benchmark in which the question does not reveal the required depth — the model must infer from the social context whether mentalizing is needed — and show that models with high ToM-order prediction accuracy on HiToM/FANToM nevertheless fail to invoke ToM when the context demands it; alternatively, show a model that answers belief questions correctly while unable to predict their ToM order, which would break the claimed link between step I and label prediction.","supporting_citations":[],"review_version":1}