{"id":"441e5279-9c5c-41b2-b67b-f3f5904768ef","arxiv_id":"2608.04616","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A new framework says that insistence on sameness in autism is the result of minimizing the entropy distance between perceived stimuli and stored memories by constraining the world to known patterns.","lead":"This paper proposes that insistence on sameness in autism is a way to minimize surprise and uncertainty, measured by an entropy distance between the environment and memory. It argues that when learning is limited, a person keeps the environment fixed to match known memories, and it turns this idea into therapy design guidelines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The §2.1 dichotomy ignores that A2–A3 describe a memory capable of updating, so the paper never shows why a person with autism must reduce D_H by constraining R rather than by learning; the explanatory conclusion is assumed, not derived.","rationale":"The reader's verdict is REJECT, and the stress-test supports that rejection, so no verdict adjustment is recommended. The reader's weakest assumption concentrated on A2–A3 and Remark 2.1: if abstract semantics are available or memory is not a nearest-neighbor store, the argument fails. The stress-test identifies a sharper problem that survives even if those assumptions are granted exactly as stated. Because A2–A3 explicitly allow storing sequences and retrieving their elements, they allow the memory to be updated by new experience. The paper's asymmetry—individuals with autism may only constrain R, not learn about R—is not entailed by any stated assumption; it is imported into Problem 3.1 when M is held fixed while R is chosen. This is a genuine gap in the central explanatory claim, and it increases rather than decreases the correctness risk. There is no independent support in the form of data, simulation, or formal verification that could compensate. The proposed simulation is a minimal, feasible check: it instantiates the paper's own assumptions and asks whether the learning route is available in the model. If the simulation shows learning works, the paper would need a substantially revised model; if it shows learning fails because M cannot update, then the missing assumption would be exposed and could be stated explicitly. The concern is about the argument, not the author, and no theatrical or ad hominem characterization is intended.","tokens_in":10336,"tokens_out":3243,"duration_ms":43379,"concrete_test":"Simulate the processing loop of Fig. 1 under A1–A4 with a stationary finite-state environment and a memory that stores and updates nearest-neighbor sequence items. Compare two policies over time: (i) let R vary and update M from exposure; (ii) constrain R to the currently remembered sequence. Compute D_H(R,M) under each policy. If policy (i) drives D_H to zero, the §3.1 dichotomy fails: the model itself admits the learning route, so constraining R is not forced. To preserve the paper's conclusion, one would need to add an explicit plasticity constraint or learning cost and show that it, not A2–A3, selects insistence on sameness.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central inference in §2.1 is that D_H(R,M) can be minimized either by learning about R (storing that knowledge in M) or by constraining R to the already known M. Section 3.1 attributes insistence on sameness to the second route, motivated by cognitive restriction to discrimination, memorization and prediction of tangible properties (Remark 2.1, Definition 4.1). But a system satisfying A2–A3—memory stores and retrieves sequences, and each memory item is a nearest-neighbor classification—is not a system that cannot learn; it is a system that can store new sequences and update its model through exposure. Remark 2.2 states only that inference is limited to replication of memorized stimuli, which is sufficient for learning the raw distribution of R. No additional assumption (a cost or bound on memory updates, a plasticity deficit, or an irreducibility of H(R|M)) is introduced to rule out the learning route. Problem 3.1 accordingly treats M as fixed ('Given an initial state of memory M') and optimizes over R alone, which presupposes the conclusion rather than deriving it. The paper itself labels the account a framework with validation deferred, and B1–B5 are post hoc translations of clinical descriptions into the model's vocabulary. The load-bearing claim therefore rests on an unstated premise: that M is effectively unchangeable for this population.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an information-theoretic framework to explain insistence on sameness in autism. It defines an entropy distance D_H(R,M)=H(R|M)+H(M|R) between an environmental random variable R and a memory variable M, interprets the two conditional entropies as surprise and uncertainty, and argues that an individual can minimize D_H either by learning about R or by constraining R to the already-known M. Under assumptions A1-A4 and Remark 2.1, which restrict the individual to nearest-neighbor memorization of tangible properties without abstract semantics, the paper concludes that insistence on sameness is a manifestation of the second strategy. It then derives therapeutic guidelines G1-G5, formulates learning therapy as an optimization problem (Problem 3.1), proposes a Turing test-like validation (V2), and offers a new definition of autism (Definition 4.1). The framework is explicitly presented as a first step requiring validation.","tokens_in":10552,"tokens_out":4844,"duration_ms":50489,"significance":"The mathematical identities in Section 2.1 are correct, and the paper is transparent about its assumptions and limitations. If the framework were supported by evidence, it could provide a formal basis for therapy design and a new definition of autism. However, the central claim, that insistence on sameness is a consequence of constraining R to memory, is an analogy rather than a derivation, and the paper provides no empirical data. The assumptions in A2-A3 and Remark 2.1 are strong empirical claims that are stated rather than justified, and the proposed validation V2 cannot test these assumptions independently. These issues undermine the explanatory contribution as it stands.","major_comments":[{"comment":"The central dichotomy in Section 2.1, that D_H can be minimized either by learning about R or by constraining R to M, is a property of the metric, but the conclusion in Section 3.1 that insistence on sameness is the latter strategy requires an independent reason why the learning route is unavailable. Assumptions A2-A3 and Remark 2.2 describe a memory capable of storing new sequences and retrieving them, so nothing in the model rules out learning the raw distribution of R. Problem 3.1 treats M as fixed and optimizes over R only, which presupposes the conclusion rather than deriving it. The paper needs an explicit assumption about a cost or bound on memory updates, or an empirical argument, to support the claimed asymmetry.","section":"Sec. 2.1, Sec. 3.1, Remark 2.2, Problem 3.1"},{"comment":"The assertion that individuals with autism have no access to abstract semantics and are restricted to discrimination, memorization, and prediction of tangible properties is a strong empirical claim. It is introduced as an assumption in Remark 2.1 and then restated as Definition 4.1, but no evidence or independent argument is provided. The nearest-neighbor memory model in A3 is itself a modeling choice, so the definition is circular with the assumptions and cannot serve as independent support for the framework. This is load-bearing because Definition 4.1 is presented as a conclusion of the paper.","section":"Remark 2.1, Definition 4.1"},{"comment":"The behaviors B1-B5 are translated into the model's vocabulary using surprise and uncertainty, which are the same quantities that define the metric. For example, B3 says that forcing deterministic scenarios reduces surprise and uncertainty, which follows directly from Eq. (4), but it does not explain why the individual adopts this strategy rather than some other. The explanations in B1-B5 are to a large extent restatements of the model's definitions in clinical terms, and their explanatory value is limited without independent evidence linking the abstract information-theoretic quantities to the observed behaviors.","section":"Sec. 3.1, B1-B5"},{"comment":"The proposed Turing test-like validation method cannot test the core assumptions A2-A3 and Remark 2.1, because the avatar's behavior is generated by the framework itself. Observing that the simulated behavior resembles clinical descriptions of insistence on sameness only shows internal consistency of the model, not that the assumptions hold for real individuals with autism. Thus the paper's claim that the framework can be validated without experiments involving individuals with autism is not supported. A validation that does not confront the model with independent data cannot break the circularity between the assumptions and the conclusions.","section":"Sec. 4.1, V2"}],"minor_comments":[{"comment":"The notation in Eqs. (2) and (3) is confusing: the conditioning event in D_H(R,M|M=m_{n-1}) and D_H(R,M|M=r_n) does not match the terms inside, and the claim that H(M=m_{n-1}|R)=0 in Eq. (2) is not obvious as written. Please clarify the conditioning notation and the derivation of these zero terms.","section":"Sec. 2.5, Eqs. (2)-(3)"},{"comment":"In B5, 'version to learning' appears to be a typo for 'aversion to learning.'","section":"Sec. 3.1, B5"},{"comment":"Assumption A3 refers to a nearest-neighbor algorithm but does not specify the distance function. Since the framework's predictions may depend on this choice, please state whether the results are invariant to the distance metric or specify what class of distance functions is assumed.","section":"Sec. 2.3, A3"},{"comment":"The phrasing 'should a priori be assumed as unrelated' is awkward; 'should be assumed a priori to be unrelated' would be clearer.","section":"Sec. 3.2, G2"},{"comment":"In the Conclusion, 'the insistent on sameness' should be 'the insistence on sameness.'","section":"Conclusion"}],"recommendation":"reject","confidential_remarks":"The paper is ambitious and clearly written, but the central inference is not logically supported. The dichotomy between learning and constraining R is used to explain insistence on sameness, yet the model's own assumptions allow learning, and no independent reason is given for why individuals with autism must take the constraining route. The proposed validation is not capable of testing the core assumptions. These are not merely presentation issues; they concern the main claim. The paper would need substantial new empirical or theoretical work to address them, which is beyond a routine revision. It might be more suitable for a venue that explicitly welcomes speculative or hypothesis-generating work, or it could be reframed as a conjecture with precise falsifiable predictions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nYou should know this paper for what it is: a clear, well-written framework that formalizes the idea that insistence on sameness reduces an entropy distance between stimuli and memory. The math is standard and correct; the problem is that the central clinical conclusion is assumed rather than derived. If you're looking for a testable hypothesis, it's a good one. If you're looking for evidence, it's not there yet.\n\nWhat is genuinely new: the specific formulation using D_H(R,M)=H(R|M)+H(M|R) for this phenomenon, the definition of autism as an impairment restricted to tangible properties (Definition 4.1), and the therapy guidelines cast as an optimization problem (Problem 3.1). The paper also does a decent job of organizing existing ideas—surprise, uncertainty, thresholds, aberrant precision—under one notation, and it is honest about its status as a framework with validation deferred. The Turing test-like validation proposal (V2) is creative, though it would only test internal consistency, not real-world truth.\n\nThe load-bearing soft spot is in Sections 2.1 and 3.1. The paper says one can reduce D_H either by learning about R (updating M) or by constraining R to M. It then attributes insistence on sameness to the second route, citing cognitive restrictions. But Assumptions A2–A3 describe a memory that stores sequences and classifies by nearest neighbor. That system can still update M and learn the distribution of R. Nothing in the assumptions rules out the learning route—no cost for memory updates, no plasticity deficit, no bound on H(R|M). Problem 3.1 treats M as a fixed initial state and optimizes over R alone, which presupposes the conclusion. The stress-test note is right: the paper never shows why an autistic individual must constrain R rather than learn. The behavioral descriptions B1–B5 are post hoc restatements in the model's vocabulary, and the guidelines G1–G5 follow from the assumptions rather than independently confirming them.\n\nThe paper is not lazy. It is transparent about what would need to be tested, and the author explicitly labels the account as a framework. The flaws are in the gap between the formalism and the clinical claim, not in the exposition.\n\nWho should read it: researchers in computational models of autism, assistive robotics, and therapists interested in formalizing care routines. I would not cite it as an established explanation, but I would send it to a serious referee. The hypothesis is testable, and peer review could push the author to add an explicit assumption about memory plasticity or present the fixed-M optimization as a conditional hypothesis rather than a conclusion. Recommendation: engage with it, but keep your eyes open. It deserves review, not desk rejection.","headline":"A clear, honest framework whose central explanation is an unstated assumption; worth reading as a hypothesis, not as a derivation.","tokens_in":11121,"tokens_out":4309,"would_cite":false,"duration_ms":45713,"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 insistence on sameness in autism is an optimal response to cognitive limits: with memory restricted to raw, tangible stimuli, the only way to reduce surprise and uncertainty is to keep the environment identical to…","keywords":["insistence on sameness","autism","entropy distance","uncertainty minimization","conditional entropy","mutual information","therapeutic guidelines","autism definition"],"falsifier":"A concrete observation that would settle it: demonstrate that an individual with autism reliably uses abstract or semantic information—for example, transferring learning across contexts through symbolic labels, understanding delayed reward, or generalizing categories—and the premise that only tangible raw sequences are available fails. Alternatively, run the proposed Turing-test-like validation: build a digital twin governed by nearest-neighbor memory and no semantics; if its simulated behavior does not exhibit insistence on sameness under entropy minimization, the analogy fails.","tokens_in":10049,"feed_emoji":"🧠","tokens_out":5998,"duration_ms":62403,"temperature":0.7,"pith_summary":"The paper proposes that insistence on sameness in autism is not a primary symptom but a rational strategy for minimizing surprise and uncertainty when learning is severely restricted. It models the world as random stimulus sequences $R$ and memory as $M$, measuring the gap between them with the entropy metric $D_H(R,M)=H(R|M)+H(M|R)$, where the two terms quantify surprise and uncertainty. Because the framework assumes memory stores only raw, tangible stimulus sequences and no abstract or semantic meanings, the individual cannot reduce the gap by learning; the only available move is to constrain $R$ to the already-known $M$. If true, this reframes rigid, repetitive behavior as an adaptive response and supports a new definition of autism as an impairment restricted to discrimination, memorization, and prediction of tangible properties. This matters because it turns therapy into a constrained optimization problem: maximize the mutual information between environment and memory while keeping surprise and anxiety below thresholds.","feed_headline":"Sameness in autism may be an entropy-minimizing strategy","feed_subtitle":"A new model treats rigid routines as the only way to cut surprise when abstract learning is closed off.","key_machinery":"The machinery is the entropy metric $D_H(R,M)=H(R|M)+H(M|R)$ together with a stimuli processing loop and Assumptions A1–A4. The metric's two conditional entropies measure surprise during perception and uncertainty during prediction; the loop updates memory by nearest-neighbor classification $m_n=\\rho(r_n)$; and Assumption A3 together with Remark 2.1 make memory a nearest-neighbor store of raw sequences with no semantic content. The identity $D_H(R,M)=H(R,M)-I(R;M)$ connects minimization of the metric to maximization of mutual information, which becomes the objective in the therapy-as-optimization formulation.","core_discovery":"The central discovery is the identification of insistence on sameness with the constrained minimization of $D_H(R,M)=H(R|M)+H(M|R)$: to lower the entropy distance between environment and memory, an agent can learn about $R$ and store that knowledge in $M$, or it can restrict $R$ to the already-known $M$. For a non-verbal low-functioning individual with autism, whose memory is assumed to be a nearest-neighbor store of raw stimulus sequences without access to abstract or semantic properties, the learning route is effectively closed, so insistence on sameness—pedantry, routine, ritual, rigidity, vigilance, and aversion to new stimuli—is the remaining way to keep surprise and uncertainty below thresholds $T_{so}$ and $T_a$. Consequently, the paper proposes Definition 4.1: autism is an impairment in which cognitive functions are restricted to discrimination, memorization, and prediction of tangible properties of the environment.","pith_inferences":["If this account is correct, insistence on sameness should intensify when novelty or cognitive load increases; graded-exposure studies in familiar settings could test that prediction directly.","The proposed avatar-based validation risks circularity if the avatar's behavior is generated by the same assumptions being validated; a stronger test would have the avatar predict behaviors not explicitly programmed, such as a preference for branches located early in a sequence.","The framework suggests that interventions teaching abstract concepts such as delayed reward or social roles will not transfer for low-functioning individuals unless anchored to tangible sequences—an empirical claim that could be compared against existing intervention outcomes.","If some autistic individuals demonstrate abstract transfer despite restricted memory, the theory would need to relax Remark 2.1 rather than abandon the entropy-minimization core."],"forward_implications":["Insistence on sameness—pedantry, routine, ritual, rigidity, vigilance, and aversion to new stimuli—is recast as the optimal available strategy for keeping surprise and uncertainty below thresholds $T_{so}$ and $T_a$ when learning is restricted.","Learning therapy becomes an optimization problem: find environments $R$ that maximize mutual information $I(R;M)$ subject to $H(R|M=m_{n-1}) < T_{so}$ and $H(M|R=r_n) < T_a$.","Therapeutic guidelines follow: use only tangible objects, assume pictures and words are unrelated to what they represent unless taught, treat memorized sequences as wholes, keep branches in sequences minimal, and use distinguishing artifacts to reduce choice uncertainty.","Self-stimulation activities may be low-entropy known stimuli that keep surprise and uncertainty inside the comfort zone during sensory deprivation.","A new definition follows: autism is an impairment in which cognitive functions are restricted to discrimination, memorization, and prediction of tangible properties of the environment."],"supporting_citations":[{"why":"Supplies the information-theoretic definitions and identities, including the entropy metric and the relation $D_H(R,M)=H(R,M)-I(R;M)$, that the entire argument builds on.","marker":"MacKay (2003)"},{"why":"Provides the free-energy principle and the generic claim that adaptive agents minimize long-term average surprise, which the paper's Conjecture 4.1 extends to autism.","marker":"Friston (2010)"},{"why":"Defines autism spectrum disorder, severity levels, and the diagnostic criteria including insistence on sameness that the paper reinterprets.","marker":"American Psychiatric Association, DSM-5 Task Force (2013)"},{"why":"Links the free-energy principle to aberrant precision in autism, supplying the precision-related vocabulary used in the behavioral analogies B1–B5.","marker":"Lawson et al. (2014)"},{"why":"Provides empirical evidence connecting insistence on sameness with effortful control and anxiety, which anchors the behavioral analogies and therapy guidelines.","marker":"Uljarević et al. (2017)"},{"why":"Proposes narrowing the definition of autism, a suggestion the paper quotes in support of its own new Definition 4.1.","marker":"Mottron and Bzdok (2020)"},{"why":"Supplies the discussion of low endogenous neural noise and its effects on autism traits, used in Remarks 2.4 and 3.1 and in the self-stimulation conjecture.","marker":"Davis and Plaisted-Grant (2015)"},{"why":"Provides the clinical compendium describing low-functioning individuals, meltdowns, and examples of self-stimulatory behavior that the framework draws on.","marker":"Volkmar (2021)"}],"fun_headline_variants":["Autism sameness may cut surprise when learning is blocked","Sameness as entropy reduction: a new autism model","Why rigid routines persist: an entropy-based theory","Insistence on sameness as uncertainty minimization","Autism: sameness reduces surprise, says entropy model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that an autistic person's memory stores raw stimulus sequences by nearest-neighbor matching and has no access to abstract or semantic properties, so constraining the environment to match memory is the only way to reduce surprise and uncertainty.","fun_headline_variants_meta":{"raw":{"variants":["Autism sameness may cut surprise when learning is blocked","Sameness as entropy reduction: a new autism model","Why rigid routines persist: an entropy-based theory","Insistence on sameness as uncertainty minimization","Autism: sameness reduces surprise, says entropy model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000527,"raw_usage":{"total_tokens":2606,"prompt_tokens":1068,"completion_tokens":1538,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":684,"completion_tokens_details":{"reasoning_tokens":1463}},"tokens_in":684,"tokens_out":1538,"duration_ms":13330,"temperature":1.0,"reasoning_tokens":1463,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:35:14.385958+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete observation that would settle it: demonstrate that an individual with autism reliably uses abstract or semantic information—for example, transferring learning across contexts through symbolic labels, understanding delayed reward, or generalizing categories—and the premise that only tangible raw sequences are available fails. Alternatively, run the proposed Turing-test-like validation: build a digital twin governed by nearest-neighbor memory and no semantics; if its simulated behavior does not exhibit insistence on sameness under entropy minimization, the analogy fails.","supporting_citations":[],"review_version":1}