{"id":"7afd91ab-d7e8-4d00-ba51-292e5f7023cb","arxiv_id":"1909.00197","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A conceptual review proposing that adaptive Theory of Mind in robots should combine simulation-based and teleological computational models, inspired by infant development.","lead":"This paper argues that robots will interact better with people if they can infer hidden mental states, such as beliefs and intentions, and proposes combining two existing psychological models to give robots this ability. It is a review and research agenda, not a new experiment, that draws on how human infants learn to understand others.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The proposed integration of simulation and teleological ToM lacks a defined representational interface, so the central claim that combining them improves adaptive ToM is underdetermined.","rationale":"I read the paper in good faith as a position/review paper, not as a claim of a working implementation. The summaries of the teleological and simulation theories and of existing robotic implementations are accurate and appropriately hedged. The central claim, however, is not merely that both models are useful, but that their integration will improve adaptive ToM in robots. The support for this is limited to a qualitative symmetry argument: simulation provides action sequences, teleological reasoning provides goals and candidate trajectories, and simulation selects among them. This argument does not specify a computational interface between the two representations, so the claimed improvements are not derivable from the cited evidence. My concern refines the reader's 'no concrete mechanism' point into a specific, testable gap: the missing mapping between motor-simulation outputs and the priors or proposals of a Bayesian inverse planner. There is no reason to reject the paper, because the proposal is a legitimate research direction and the paper explicitly frames it as a suggestion for future work. But the conditional verdict is appropriate, because the central claim cannot yet be evaluated as a technical contribution until the integration is formalized or demonstrated in a proof-of-concept experiment. The verdict should therefore remain unchanged.","tokens_in":7994,"tokens_out":3721,"duration_ms":37153,"concrete_test":"Build a minimal integration on the false-belief/repositioning task from Milliez et al. (Sec 2.1): use Baker et al.'s Bayesian inverse-planning model as the teleological component and a learned kinematic forward model of the observed agent's arm as the simulation component. Define the interface explicitly, for example by letting simulated trajectories form the proposal distribution for action explanations in the inverse planner, while the planner's posterior over goals reweights future simulation. Run three systems (teleological alone, simulation alone, and integrated) on both standard rational false-belief scenarios and scenarios where the agent acts inefficiently or non-rationally, which is the case where Sec 3.1 says teleological reasoning fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.3 asserts that simulation and teleological models are complementary: mirrored action sequences can support teleological goal inference, and teleological reasoning can supply candidate trajectories that simulation then evaluates. The central claim depends on this division of labor being realizable, but no mechanism is given. The two component models operate on different representations: Baker et al.'s teleological model (Sec 3.1) is a Bayesian inverse planner over beliefs, desires, and actions, while simulation models (Sec 3.2) use the observer's own motor representations and predictive activation. Section 3.3 does not specify how a simulated motor trajectory becomes a prior or proposal for inverse planning, how an inferred goal posterior modulates simulation, or what update rule closes the loop. Without this interface, the proposed integration is a conceptual hope, not a testable architecture. The paper's own review shows the weaknesses may not be complementary: teleological reasoning cannot handle subjective or non-rational mental states, and simulation alone fails on novel situations and quarantine. It is not shown that combining them fixes these specific failure modes rather than carrying both sets of limitations. Sec 3.1 also notes Baker et al.'s model is too computationally demanding for online interaction, yet Sec 3.3's adaptive-robotics goal requires online operation; no approximation or complexity reduction is proposed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper is a perspective/review article aiming to transfer insights from developmental psychology on infant Theory of Mind (ToM) to social robotics. It argues that current social robots mostly rely on passive perception, prewired knowledge, and deep neural networks that struggle with generalization and online interaction, and that an \"adaptive ToM\"—the ability to autonomously attribute and reason about mental states—would provide four functional advantages: mentalizing for belief understanding, proactivity and preparation, active perception, and learning. The paper reviews two computational accounts of ToM: the teleological approach (rational-action-based goal inference, including Baker et al.'s Bayesian model) and the simulation approach (motor resonance and predictive activation), and surveys robotic implementations (e.g., Milliez et al., Görür et al., Rabinowitz et al.) to identify current limitations. The central proposal, in Section 3.3, is that simulation and teleological models should be integrated as complementary rather than competing, with simulation providing action trajectories and teleological reasoning providing goal states and candidate inferences. The paper does not present a formal model, experiments, or a concrete computational architecture; it concludes with open questions and a call for future cross-talk between developmental psychology and robotics.","tokens_in":8348,"tokens_out":5207,"duration_ms":37715,"significance":"The paper is a well-referenced synthesis of two research streams and usefully articulates a research agenda. Its strengths include a clear enumeration of functional ToM benefits for robotics, concrete examples of current robotic shortcomings (e.g., hard-coded hypotheses in Milliez et al., the all-observable assumption in Rabinowitz et al.), and a balanced discussion of the limitations of both teleological and simulation accounts. The proposal that these accounts are complementary is plausible and timely, and the developmental-robotics framing is appropriate. However, the significance of the central claim is currently limited by the absence of a mechanistic account or a proof-of-concept demonstration. As a position paper it can stimulate discussion, but as a contribution to a computational journal it needs to specify the integration at a level that could be implemented or at least modeled. The paper makes no falsifiable predictions and ships no code or data, so its value lies in framing rather than demonstration. If the proposed integration were later realized, the contribution could be an important stepping stone for adaptive social robots.","major_comments":[{"comment":"The central proposal that simulation and teleological models are complementary is not supported by a concrete computational mechanism. The teleological model described in Sec. 3.1 is a Bayesian inverse planner over beliefs, desires, and actions, whereas the simulation models in Sec. 3.2 operate on motor representations and predictive activation. Section 3.3 does not specify how a simulated motor trajectory becomes a prior or proposal for inverse planning, how an inferred goal posterior modulates simulation, or what update rule closes the loop. Without such an interface, the claim that integration yields improved adaptive ToM is underdetermined and non-testable. The paper also does not address whether the two models' known failure modes are complementary: teleological reasoning struggles with subjective and non-rational mental states, while simulation fails on novel situations and quarantine; no argument is given that combining them overcomes these specific problems rather than inheriting both sets of limitations.","section":"Sec. 3.3"},{"comment":"The paper acknowledges that Baker et al.'s teleological model is \"computationally demanding and could not be directly used to support online interactions\" (Sec. 3.1), yet the proposed adaptive-robotics applications (proactivity, active perception, learning) require online operation. No approximation scheme, hierarchical inference strategy, or complexity reduction is proposed to bridge this gap. As a result, the claim that integrating the two models will address current robotic limitations is not yet justified.","section":"Secs. 3.1 and 3.3"},{"comment":"The functional advantages listed (mentalizing for belief understanding, proactivity and preparation, active perception, and learning) are presented as benefits that an adaptive ToM would confer, but the paper does not provide evidence that the proposed integration of simulation and teleological models, or any existing architecture, actually delivers these benefits. The examples of Milliez et al. and Görür et al. illustrate current limitations rather than demonstrating the proposed pathway. The claim that these advantages will transfer from human infants to robots is therefore asserted rather than derived.","section":"Sec. 2.1"}],"minor_comments":[{"comment":"The term \"adaptive ToM\" is used throughout, but it is never formally defined beyond \"adaptive attribution of mental states.\" A precise definition would help distinguish it from standard ToM implementations.","section":"Sec. 2.1"},{"comment":"The sentence \"Against the simulation theory as a base for mentalizing is also some evidence of its inability to support action understanding in novel situations\" is awkwardly phrased and should be reworded for clarity.","section":"Sec. 3.2"},{"comment":"The two debates (innate versus learned, bottom-up versus top-down) are introduced but not connected explicitly to the proposed integration; making this connection would strengthen the future-directions discussion.","section":"Sec. 4"},{"comment":"Reference 18 is a self-citation to the authors' prior conference paper; the text should indicate what new material the present contribution adds beyond that paper.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"This is a review/position manuscript with no novel experimental or computational contribution. The main risk is that the integration proposal in Sec. 3.3 is a research agenda rather than a testable claim. The journal may wish to consider whether such perspective articles fit the scope; if accepted, it should be clearly labeled as a position paper and the integration proposal framed as a call for future work rather than a demonstrated solution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis is a position paper, not a results paper, and it reads honestly as one. The useful part is a clear review of simulation and teleological ToM theories and a reasoned case that robots could benefit from adaptive mentalizing. The authors organize the functional advantages well—belief tracking, proactivity, active perception, learning—and they summarize Baker et al. and the simulation literature accurately. They also don't oversell what current robotic implementations actually do.\n\nThe central proposal, in Section 3.3, is that integrating simulation and teleological models will yield better ToM. That is plausible but underspecified. The stress-test note is right: the paper gives no representational interface, no algorithm for how a mirroring-derived trajectory becomes a prior for inverse planning, and no account of how teleological reasoning supplies proposals for simulation. Without that, the integration is a hope, not a testable architecture. The paper also notes Baker's model is computationally demanding, yet the adaptive robotics goal demands online operation; no approximation is suggested. And it isn't shown that the two approaches' failure modes are complementary rather than overlapping.\n\nThese are soft spots, but they are soft in proportion. The paper frames this as a research direction, not a proven result, and its claims are appropriately hedged. The novelty is incremental—combining these accounts already exists in the psychology literature (Keysers and Gazzola, 2007)—but the application to robotic architecture design is a reasonable extension.\n\nThere are no red flags in the references or citation practice. The one self-citation (reference 18) points to their own earlier conference paper for further detail, not to inflate a novel claim.\n\nIf I were handling this, I would send it to peer review only if the venue explicitly publishes position or review pieces. For a robotics venue expecting empirical or formal contributions, it would be a poor fit. But as a perspective piece, it deserves a serious referee rather than a desk reject: the writing is clear, the review is balanced, and the integration proposal identifies a concrete gap worth probing. My recommendation: let it go to review with the request that the authors specify at least a minimal computational interface, or bound the kinds of integration they mean.\n\nBest.","headline":"A clear, honest position paper that makes a plausible but underspecified case for integrating simulation and teleological ToM in robots; worth publishing as a perspective, not as a demonstrated result.","tokens_in":8723,"tokens_out":2039,"would_cite":false,"duration_ms":20501,"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":"This paper argues that social robots need an adaptive Theory of Mind built by integrating teleological reasoning and simulation, drawing on infant development.","keywords":["theory of mind","social robotics","developmental robotics","mentalizing","teleological reasoning","simulation theory","human-robot interaction","belief tracking"],"falsifier":"Build three robotic or simulated agents—one using only teleological inference, one using only simulation, and one using the proposed integration—and run them on a false-belief task with unfamiliar agents; if the integrated agent does not consistently beat the better single-model agent in predicting action or belief, the paper's central proposal is not supported.","tokens_in":7797,"feed_emoji":"🤖","tokens_out":5261,"duration_ms":50469,"temperature":0.7,"pith_summary":"This position paper argues that social robots will keep falling short in everyday, unstructured settings until they can infer hidden mental states—beliefs, desires, intentions—rather than merely recognize actions. The authors build this case by reviewing research on infants' adaptive Theory of Mind and identifying four functional advantages it would give robots: tracking false beliefs, acting proactively, perceiving actively, and learning efficiently. They then argue that the two dominant computational accounts of mentalizing, teleological reasoning and simulation, are not rivals to choose between but complementary processes that should be integrated in future robot architectures. If they are right, the path to flexible social robots runs through developmental psychology, and the same architectures could in turn serve as testbeds for theories of how human mentalizing works.","feed_headline":"Merge two mind-reading models to give robots a social sense","feed_subtitle":"A robotics review argues teleological reasoning and simulation must work together for robots to infer beliefs and act in the wild.","key_machinery":"The load-bearing object of the paper is the proposed integration loop between two theories of how minds are read. Teleological reasoning (from the rationality principle) infers goals from observable actions against situational constraints; simulation theory instead proposes that we understand others by re-using our own mental states and motor representations. The paper's key move is to treat these not as competing accounts but as complementary components: mirroring or simulation can generate the action trajectories leading to a goal, while teleological inference can choose among candidate goals by rationality, and the same loop can run in reverse. This two-way complementarity is the machinery that is supposed to deliver adaptive mentalizing in robots, with the four functional advantages—belief tracking, proactivity, active perception, and learning—as its downstream effects.","core_discovery":"On the paper's own terms, the central claim is that an adaptive Theory of Mind—the capacity to attribute and reason about others' mental states—can and should be built into social robots, and that the way to do it is to combine the teleological and simulation accounts rather than treat them as competing. Teleological reasoning treats action as goal-directed and rational, letting an observer infer intentions from situational constraints; simulation lets an observer use their own mental states as a model of another's. The authors propose a complementary architecture in which simulation supplies the concrete action sequences or trajectory options and teleological inference selects among them using rationality, with the reverse flow also possible: teleological reasoning proposes candidate goals, and simulation chooses between them by internally enacting the resulting experience. This integration, they argue, is what would unlock mentalizing for belief understanding, proactivity and preparation, active perception, and learning, and thereby improve human-robot interaction in unpredictable environments such as disaster response and construction sites.","pith_inferences":["An implication left implicit in the paper is that the two models would need to be engaged under different conditions—teleological inference when the context is novel or top-down control is needed, simulation when familiar bottom-up cues dominate—so a workable architecture would also need a meta-level mechanism to choose between them.","If the integration claim is right, a concrete testable prediction follows: in false-belief tasks with unfamiliar agents, an integrated model should outperform either a pure simulation or a pure teleological model, especially when the observed agent's behavior is not perfectly rational.","The authors' emphasis on infant abilities hints at a developmental curriculum for robots: start with teleological priors, then add simulation-based learning from interaction, rather than attempting adult-level Theory of Mind in one step."],"forward_implications":["Robots equipped with an integrated adaptive ToM could understand that a person digging through rubble is searching for survivors, not just recognize the digging motion.","Such robots could anticipate a human partner's needs before an action is completed, enabling proactive assistance in collaborative tasks.","Reliance on massive human-recorded datasets would shrink, because mental-state inference supplies context that bottom-up action recognition lacks.","The same architectures would provide a physical testbed for developmental psychology, letting theories about infant mentalizing be evaluated in an embodied agent."],"supporting_citations":[{"why":"Supplies the rationality principle that grounds teleological reasoning about goals from actions.","marker":"[28]"},{"why":"Supplies the simulation theory account of mind-reading through re-use of one's own mental states.","marker":"[34]"},{"why":"Supplies a neural-evidence precedent for integrating simulation and mentalistic accounts rather than treating them as opposed.","marker":"[37]"},{"why":"Supplies a Bayesian computational ToM model based on the teleological principle, used as the main example of the teleological approach's strengths and computational cost.","marker":"[32]"},{"why":"Supplies an implemented robot that tracks beliefs and takes perspectives but relies on hard-coded hypotheses, illustrating current limitations.","marker":"[20]"},{"why":"Supplies a neural network that predicts agents in false-belief situations while assuming access to all states and actions, an assumption the paper says embodied robots cannot usually meet.","marker":"[8]"},{"why":"Earlier companion work by the authors that details the four functional advantages of adaptive ToM for robotics.","marker":"[18]"},{"why":"Provides the working definition of Theory of Mind and the framing of simulation theory as a basis for mentalizing.","marker":"[12]"},{"why":"Provides evidence of predictive motor activation in infants during action observation, used to support the simulation account.","marker":"[35]"},{"why":"Provides evidence that infants attribute goals even to biomechanically impossible actions, used to support the teleological account.","marker":"[27]"}],"fun_headline_variants":["Robots need both simulation and teleology for true mindreading","Combine two ToM models to make robots socially aware","Adaptive Theory of Mind: merge teleology and simulation in robots","For robot social sense, fuse teleological and simulation reasoning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument depends on the assumption that combining teleological reasoning and simulation inside a robot architecture will recreate the adaptive social abilities of human infants, even though the paper does not supply a concrete computational mechanism or demonstration of that transfer.","fun_headline_variants_meta":{"raw":{"variants":["Robots need both simulation and teleology for true mindreading","Combine two ToM models to make robots socially aware","Adaptive Theory of Mind: merge teleology and simulation in robots","For robot social sense, fuse teleological and simulation reasoning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000403,"raw_usage":{"total_tokens":2101,"prompt_tokens":945,"completion_tokens":1156,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":1087}},"tokens_in":561,"tokens_out":1156,"duration_ms":7965,"temperature":1.0,"reasoning_tokens":1087,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:57:17.651296+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build three robotic or simulated agents—one using only teleological inference, one using only simulation, and one using the proposed integration—and run them on a false-belief task with unfamiliar agents; if the integrated agent does not consistently beat the better single-model agent in predicting action or belief, the paper's central proposal is not supported.","supporting_citations":[{"cited_title":"Trends in Cognitive Sciences 7, 287-292 (2003)","cited_arxiv_id":null,"evidence_quote":"Supplies the rationality principle that grounds teleological reasoning about goals from actions."},{"cited_title":"Trends in Cognitive Sciences 2, 493-501 (1998)","cited_arxiv_id":null,"evidence_quote":"Supplies the simulation theory account of mind-reading through re-use of one's own mental states."},{"cited_title":"Trends in Cognitive Sciences 11, 194-196 (2007)","cited_arxiv_id":null,"evidence_quote":"Supplies a neural-evidence precedent for integrating simulation and mentalistic accounts rather than treating them as opposed."},{"cited_title":"Nature Human Behaviour 1:0064 (2017)","cited_arxiv_id":null,"evidence_quote":"Supplies a Bayesian computational ToM model based on the teleological principle, used as the main example of the teleological approach's strengths and computational cost."},{"cited_title":"The 23rd 10 IEEE International Symposium on Robot and Human Interactive Communication, pp","cited_arxiv_id":null,"evidence_quote":"Supplies an implemented robot that tracks beliefs and takes perspectives but relies on hard-coded hypotheses, illustrating current limitations."},{"cited_title":"The 11th Computer Science and Electronic Engineering Conference","cited_arxiv_id":null,"evidence_quote":"Earlier companion work by the authors that details the four functional advantages of adaptive ToM for robotics."},{"cited_title":"in The Oxford Handbook of Philosophy of Cognitive Sci- ence, (Oxford: Oxford University Press) (2012)","cited_arxiv_id":null,"evidence_quote":"Provides the working definition of Theory of Mind and the framing of simulation theory as a basis for mentalizing."},{"cited_title":"Biology letters 5, 769-772 (2009)","cited_arxiv_id":null,"evidence_quote":"Provides evidence of predictive motor activation in infants during action observation, used to support the simulation account."},{"cited_title":"Cognition 107, 1059-1069 (2008)","cited_arxiv_id":null,"evidence_quote":"Provides evidence that infants attribute goals even to biomechanically impossible actions, used to support the teleological account."}],"review_version":1}