{"id":"91da5989-bac4-47fb-8dba-ebe3281a8083","arxiv_id":"2506.18176","paper_version":1,"verdict":"UNVERDICTED","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"This review synthesizes clinical and computational evidence that perceptual multistability could serve as a transdiagnostic, translational window into psychiatric and neurodevelopmental disorders.","lead":"Perceptual multistability, when one ambiguous image or sound flips between interpretations, differs measurably across psychiatric conditions, and this review argues it should become a shared experimental window into those disorders. It connects two computational psychiatry traditions, Bayesian inference and reinforcement learning, and outlines how combining them could link behavior to brain circuits.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The transdiagnostic stability claim pools paradigms whose shared variance is weak, and SCZ switch rates move in opposite directions across tasks, so the clinical bridge may rest on a paradigm-specific effect.","rationale":"The reader's weakest assumption matches mine: cross-paradigm comparability. This is load-bearing because the Abstract's conclusion that multistability is a promising non-invasive clinical tool rests mainly on the transdiagnostic stability pattern. That pattern is assembled from studies that rarely use more than one paradigm per participant, so cross-paradigm correlations within individuals are only weakly established in healthy participants (ref [33]) and almost absent in clinical groups. The paper explicitly flags SCZ discrepancies and faster switching on continuous SFM [84], a direct reversal of the BR result. If these effects are paradigm-specific, multistability could still be useful for studying separate mechanisms, but the stronger claim—one tool bridging computational psychiatry subfields and species—would be premature. The Bayesian–RL integration is an untested proposal, with no equations or simulations, so I do not treat that as a fatal flaw; a review is allowed to set an agenda. However, the agenda's empirical foundation should be internally coherent. I credit the authors for including the contradictory references; the issue is that the synthesis weights BR evidence as green while the disconfirming results are not integrated into the transdiagnostic conclusion. A multilevel analysis allowing group-by-paradigm interactions would settle whether the conflict is noise or signal. The reader's UNVERDICTED verdict remains appropriate because this is a roadmap; my concern justifies a stronger caveat inside the roadmap rather than a change of verdict.","tokens_in":29688,"tokens_out":4461,"duration_ms":55159,"concrete_test":"Take the individual switch-rate and dominance-duration data from the clinical studies cited in Percept stability (or collect one new dataset administering BR, Necker/Schröder ambiguous figures, and continuous and intermittent SFM to the same SCZ, BD, MDD, at-risk, and control participants), and fit a multilevel latent-variable model with a common 'perceptual stability' factor and paradigm-specific loadings. Compare this model against one with paradigm-specific group effects, such as SCZ slower in BR but faster in SFM, using WAIC or cross-validation, and estimate the BR–SFM within-person correlation after controlling for diagnosis. If the paradigm-specific model fits better or the shared factor explains little variance, the transdiagnostic 'increased stability' claim is unsupported; if a single factor reproduces the group differences, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central clinical pillar—that increased perceptual stability is transdiagnostic in BD, MDD, and SCZ (Percept stability; Table 1)—requires that binocular rivalry (BR), ambiguous figures, and structure-from-motion (SFM) measure a shared latent computation. The manuscript itself contains both pieces of evidence against this requirement. Box 1 cites Cao et al. (ref [33]) showing only limited within-individual correlation between BR and SFM switch rates, i.e., the shared variance among paradigms is weak. The Percept stability section then documents opposite effects in SCZ: slower BR switching [72,79–81], but decreased stability for the Rubin vase [55], no difference for the Schröder staircase [54], reduced stability for intermittent SFM [78,82,83], and faster continuous-SFM switching along the psychosis spectrum [84]. If task-specific mechanisms dominate, the 'increased stability' pattern is largely a BR-specific finding, and the proposed Bayesian–RL integration would explain paradigm-specific computations rather than one common window. The acknowledgement of discrepancies does not repair this, because Table 1 rates the BR evidence green and relegates the contradictions to italics, allowing the transdiagnostic conclusion to be built on the least reliable shared core. If switch rate or dominance duration is not a stable, task-general trait, the translational and cross-species claims in the Abstract and Figure 1d lose their empirical base. This is a correctness risk in the review's internal synthesis, not a disagreement with external consensus: the cited literature itself contains the disconfirming pattern.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This perspective/review paper argues that perceptual multistability offers a transdiagnostic window into psychiatric and neurodevelopmental disorders. It organizes clinical findings into three facets—percept formation, percept stability, and percept failures—and then reviews two computational psychiatry frameworks, Bayesian inference (sampling, predictive coding, circular inference) and reinforcement learning (perceptual switches as internal actions in a POMDP). The central claim is that multistability can bridge these frameworks, link human and animal studies, and connect behavior to neural circuits, making it a promising non-invasive translational tool. The paper includes boxes with formal descriptions of predictive coding, circular inference, and POMDPs, a summary table of evidence, and an 'Outstanding questions' list.","tokens_in":29921,"tokens_out":5637,"duration_ms":70373,"significance":"If the central claim holds, multistability would be a valuable transdiagnostic assay that connects altered perception to specific computational deficits and to circuit-level mechanisms across species. The paper's strengths are its broad coverage of paradigms, species, and modalities; its explicit acknowledgment of contradictory findings; its self-contained mathematical boxes; and its concrete translational suggestions (animal models, pharmacology, genetics, online data collection). The synthesis is programmatic rather than demonstrative, but the clinical and computational integration it proposes is timely and could guide future meta-analyses and model-based studies. The main risk is that the transdiagnostic conclusion rests on pooling paradigms whose shared variance is weak, as the manuscript itself notes.","major_comments":[{"comment":"The transdiagnostic 'increased perceptual stability' claim is load-bearing for the Abstract and Figure 1, but it is built on pooling binocular rivalry (BR), ambiguous figures, and structure-from-motion (SFM) as one construct. The manuscript itself provides the evidence against this pooling: Box 1 states that BR and SFM switch rates show only limited within-individual correlation [33], and the Percept stability section reports opposite directions in schizophrenia across paradigms (slower BR [72,79–81], decreased stability for the Rubin vase [55], no difference for the Schröder staircase [54], reduced stability for intermittent SFM [78,82,83], and faster continuous-SFM switching [84]). Table 1 codes BR as green and relegates the contradictions to italics, so the 'compelling evidence' conclusion is effectively supported by the paradigm with the least established shared core. A meta-analysis with paradigm as a moderator, or within-subject cross-paradigm correlations in clinical groups, is needed before the transdiagnostic stability conclusion can stand.","section":"Percept stability; Box 1; Table 1"},{"comment":"The Table 1 legend defines red as 'low level of evidence: contradictory results across studies,' yet the SCZ, BD, and MDD rows are green while containing italicized contradictory references (notably [68] under SCZ and MDD). By the legend's own definition, contradictory results should downgrade the row to red, or the legend should explain why a contradiction does not meet that threshold. As published, the evidence grading is internally inconsistent and overstates support for the transdiagnostic claim.","section":"Table 1, color-code legend"},{"comment":"The proposed integration of Bayesian inference and reinforcement learning is only sketched as two broad options—replacing a POMDP's belief update with hierarchical inference, or adopting active inference—without a concrete model, parameter mapping, or falsifiable prediction linking reward, priors, and dominance durations. Since the Abstract's bridging claim is a central contribution, at least one explicit computational example (e.g., a hierarchical-belief POMDP or an active-inference formulation) with a stated predicted dissociation would substantiate the synthesis rather than leaving it as an agenda.","section":"Toward a unified synthesis"}],"minor_comments":[{"comment":"The section heading 'Toward a unified synthetis' contains a typo; it should read 'synthesis'.","section":"Toward a unified synthetis (heading)"},{"comment":"The phrase 'two distincts domains' should be 'two distinct domains.'","section":"Concluding remarks"},{"comment":"Spelling is inconsistent: Box 1 uses 'Schroeder stairs' while the Glossary and text use 'Schröder staircase'; please standardize.","section":"Box 1"},{"comment":"Reference [68] is cited as a 'but see' for both MDD and SCZ, but the reference is a bipolar-disorder study; the intended contradictory evidence for those diagnostic groups should be re-checked and replaced with appropriate citations if needed.","section":"Percept stability"},{"comment":"Many entries lack volume, page, or year information (e.g., [2], [6], [7], [18], [23], [30], [36], [49], [51], [67], [68]), which hinders verification of the claims in Table 1 and the text; the bibliography should be completed.","section":"Reference list"},{"comment":"The message-passing equations do not define the summation set or the relationship between M_ji and M_ij; adding one sentence to clarify indices would make the model self-contained.","section":"Box 3"}],"recommendation":"major_revision","confidential_remarks":"The paper is a broad perspective with a heavy but not disqualifying self-citation pattern. The main editorial risk is overclaiming: the transdiagnostic stability conclusion needs to be either restricted to binocular rivalry or supported by a paradigm-moderated meta-analysis. The manuscript is likely suitable for a review/perspective venue after these revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nWhat you should know: this is a narrative review and research roadmap, not a primary report. There is no new data, no new equations, no model fits. Its value is in the synthesis and in the explicit acknowledgment of contradictions in the clinical literature. The authors are careful—they flag 'but see' references, color-code Table 1 by evidence level, and admit that SCZ findings go in opposite directions across paradigms. That is more honest than most reviews in this area.\n\nWhat is actually new is modest: the proposal to bridge Bayesian and reinforcement-learning accounts of multistability by replacing the flat POMDP belief update with hierarchical inference, plus the active-inference reinterpretation of reward-modulated rivalry. These are sketched in a few paragraphs, with no worked examples. As a research agenda it is plausible; as a contribution it is thin.\n\nThe main soft spot is the transdiagnostic claim. The paper says there is compelling evidence for increased perceptual stability in binocular rivalry across psychiatric conditions, and takes this to support a transdiagnostic perspective. But the paper itself cites evidence that binocular rivalry and structure-from-motion switch rates correlate only weakly (Box 1, ref 33), and that in SCZ, binocular rivalry shows slower switching while other paradigms show reduced stability or null results. The authors do not hide this—they put the contradictions in italics in Table 1 and in the text. Yet the abstract and Figure 1d still present multistability as a general window on psychiatric disorders. That framing rests on a binocular-rivalry-specific effect. If these paradigms are not measuring the same latent computation, the transdiagnostic story is largely a rivalry story. This is a limitation in the internal synthesis, not an external attack; the cited literature itself contains the disconfirming pattern. Given the paper's embrace of a transdiagnostic window, this tension should have been addressed head-on, ideally with a quantitative check of whether the effects actually generalize across paradigms.\n\nThe theoretical scaffolding leans heavily on the authors' own prior work—circular inference, the bistable circular-inference model, and the RL account of multistability. That is not a flaw in a review, but it means the proposed Bayesian-RL bridge is built from one school's models and remains a verbal agenda.\n\nBottom line: this is a fair, well-written review that will help people entering multistability or computational psychiatry, and clinicians wanting a compact map of the evidence. It is not a quantitative synthesis, and the transdiagnostic language overstates cross-paradigm consistency. It deserves a serious referee and likely publication after revision, with more careful framing of the generality claim. I would not cite it as a source for transdiagnostic conclusions without checking the primary papers, but I would point newcomers to it as a starting point.","headline":"A useful, honest narrative review that overreaches slightly in its transdiagnostic framing; the binocular-rivalry evidence is solid but the cross-paradigm consistency is not.","tokens_in":30483,"tokens_out":2415,"would_cite":true,"duration_ms":27881,"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 perceptual multistability can serve as a non-invasive, transdiagnostic window into psychiatric disorders, bridging Bayesian inference and reinforcement learning and linking behavior to neural circuits.","keywords":["perceptual multistability","binocular rivalry","ambiguous figures","structure-from-motion","computational psychiatry","Bayesian inference","reinforcement learning","switch rate"],"falsifier":"Measure binocular rivalry and structure-from-motion switch rates in the same individuals across a transdiagnostic sample spanning bipolar disorder, depression, schizophrenia, anxiety, and controls; if cross-paradigm correlations are near zero or the direction of group differences reverses with paradigm, the shared-mechanism premise that carries the transdiagnostic synthesis would be falsified.","tokens_in":29476,"feed_emoji":"👁️","tokens_out":7687,"duration_ms":80958,"temperature":0.7,"pith_summary":"This review argues that perceptual multistability—the spontaneous alternation between two or more interpretations of an unchanging ambiguous image—is a particularly informative window into psychiatric and neurodevelopmental disorders. The authors organize existing findings into three facets: how percepts are formed, how stable they are, and how often they fail, and show that each facet is altered in clinical populations. They then propose that these facets map onto the two main computational approaches in psychiatry: Bayesian inference, which treats perception as unconscious inference, and reinforcement learning, which treats perceptual switches as internal actions. If the synthesis is right, multistability tasks could become a non-invasive, transdiagnostic assay that ties altered perception to specific computational deficits and to neural circuits that can be studied across species. This is a synthesis paper, so its contribution is a unified framework and a roadmap, not new experimental data.","feed_headline":"Perceptual switches could become a transdiagnostic psychiatric assay","feed_subtitle":"A review links binocular rivalry and ambiguous figures to Bayesian and reinforcement-learning accounts of mental illness.","key_machinery":"The central object is the multistable percept itself: an ambiguous stimulus that supports more than one interpretation, studied through binocular rivalry, ambiguous figures, and structure-from-motion. Its behavioral readouts are switch rate and dominance duration, which the paper treats as trait-like indices of perceptual stability. The computational machinery has three parts: hierarchical Bayesian inference, including predictive coding and circular inference, where message-passing uses a correction factor $a$ to remove redundant information and $a<1$ produces overcounted evidence; reinforcement learning formulated as a partially observable Markov decision process (POMDP) in which a perceptual switch is an internal action chosen to maximize long-run utility; and the authors' integrative proposal to combine them by enriching the POMDP's belief update or by adopting active inference. These components carry the argument because they let clinical differences in percept formation, stability, and failure be restated as differences in prior weighting, precision, information-loop strength, reward sensitivity, and decision urgency.","core_discovery":"On the paper's own terms, the central discovery is that a single phenomenon—perceptual multistability—exposes three separable dimensions of perceptual dysfunction that recur across diagnoses: altered integration of priors, emotions, and sensory cues during percept formation; altered perceptual stability, most consistently increased stability in binocular rivalry for bipolar disorder, major depressive disorder, and schizophrenia, with decreased stability in anxiety disorders; and increased 'mixed' percept failures in autism and ADHD. The authors' claim is that these dimensions align with distinct computational accounts, so that the same behavioral task can bridge Bayesian and reinforcement-learning frameworks and link behavior to circuit-level mechanisms such as excitatory-inhibitory balance and dopaminergic modulation. They argue that multistability is therefore a practical tool for translational research, connecting human experiments to animal models and to transdiagnostic computational psychiatry.","pith_inferences":["Beyond the paper's claims, a direct test of its shared-mechanism premise would be to compute cross-paradigm correlations of switch rates within the same clinical subjects; the paper itself notes the rivalry–structure-from-motion correlation is limited, so low correlations would force the transdiagnostic story to separate into paradigm-specific stories.","Beyond the paper's claims, the reinforcement-learning account would predict that altering average reward rate—through monetary incentives, motivational state, or dopaminergic medication—should shift switch rates in clinical groups in the direction of opportunity-cost theory, a prediction testable in existing rivalry paradigms.","Beyond the paper's claims, if perceptual switches are internal actions, multistability could serve as a stripped-down model for other internally generated cognitive sequences such as mental simulation in planning, connecting this framework to cognitive-map research in psychiatry.","Beyond the paper's claims, a matrix design crossing the three facets (formation, stability, failure) with three paradigms could dissociate psychiatric symptom dimensions better than any single task, since the facets may map onto distinct computational parameters."],"forward_implications":["Switch rate and dominance duration could serve as transdiagnostic trait markers, detectable in first-degree relatives and along the psychosis spectrum before full diagnosis.","Because multistability is observed across species, the same behavioral readouts can be used in animal models, where causal tools can test whether altered stability arises from specific circuits or neuromodulators.","Reward manipulations during rivalry should modulate switch rates if switches are internal actions, giving clinical studies a handle on motivational and dopaminergic contributions to perception.","Bayesian and reinforcement-learning accounts can be unified, so that phenomena one account explains poorly—such as increased stability in depression or schizophrenia—can be reinterpreted as interactions between belief updating and value-based action selection.","Mixed percepts offer a separate axis tied to excitatory-inhibitory balance, potentially distinguishing neurodevelopmental conditions from mood and psychotic disorders."],"supporting_citations":[{"why":"Establishes the original finding that binocular rivalry switch rate is slow in bipolar disorder, the seed of the transdiagnostic stability claim.","marker":"[66]"},{"why":"Replicates slow binocular rivalry in bipolar disorder and is cited as a contradictory reference, marking the stability finding's reproducibility and its limits.","marker":"[68]"},{"why":"Reports slower and less variable rivalry rates across bipolar disorder, OCD, major depression, and schizophrenia, the paper's core transdiagnostic evidence.","marker":"[72]"},{"why":"Shows faster bistable switching along the psychosis spectrum, the key boundary condition that paradigm choice can reverse the stability pattern.","marker":"[84]"},{"why":"Introduces the circular inference model of schizophrenia, the computational mechanism the paper uses to link altered priors and excitatory-inhibitory imbalance to symptoms.","marker":"[123]"},{"why":"Provides the dynamical circular inference theory of bistable perception, turning overcounted top-down or bottom-up messages into a bistable attractor.","marker":"[130]"},{"why":"Frames multistability as internal foraging with perceptual value, the decision-theoretic perspective that treats switches as actions rather than passive reweighting.","marker":"[5]"},{"why":"Proposes the explicit decision-theoretic and POMDP model of switches as internal actions, the reinforcement-learning foundation the paper wants to bridge with Bayesian accounts.","marker":"[148]"},{"why":"Documents limited correlation between binocular rivalry and structure-from-motion switch rates, the evidence that sets the paper's weakest shared-mechanism assumption.","marker":"[33]"},{"why":"Shows comparable gamma-like alternation-rate distributions across bistable paradigms, the empirical basis for treating paradigms as sharing temporal dynamics.","marker":"[19]"}],"fun_headline_variants":["Perceptual multistability: a transdiagnostic window into mental illness","The same perceptual illusion bridges Bayesian and reinforcement psychiatry","A noninvasive perceptual test could unite animal and human psychiatric research","Perceptual switches reveal shared cognitive faults across psychiatric diagnoses"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that binocular rivalry, ambiguous figures, and structure-from-motion share enough of a common timing mechanism, and that switch rate and dominance duration are stable trait-like measures of it, so that findings from different paradigms and species can be pooled into a single transdiagnostic pattern.","fun_headline_variants_meta":{"raw":{"variants":["Perceptual multistability: a transdiagnostic window into mental illness","The same perceptual illusion bridges Bayesian and reinforcement psychiatry","A noninvasive perceptual test could unite animal and human psychiatric research","Perceptual switches reveal shared cognitive faults across psychiatric diagnoses"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000995,"raw_usage":{"total_tokens":4161,"prompt_tokens":835,"completion_tokens":3326,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":451,"completion_tokens_details":{"reasoning_tokens":3258}},"tokens_in":451,"tokens_out":3326,"duration_ms":29840,"temperature":1.0,"reasoning_tokens":3258,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:23:11.402395+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure binocular rivalry and structure-from-motion switch rates in the same individuals across a transdiagnostic sample spanning bipolar disorder, depression, schizophrenia, anxiety, and controls; if cross-paradigm correlations are near zero or the direction of group differences reverses with paradigm, the shared-mechanism premise that carries the transdiagnostic synthesis would be falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Replicates slow binocular rivalry in bipolar disorder and is cited as a contradictory reference, marking the stability finding's reproducibility and its limits."},{"cited_title":"Multistability, perceptual value, and internal foraging","cited_arxiv_id":null,"evidence_quote":"Frames multistability as internal foraging with perceptual value, the decision-theoretic perspective that treats switches as actions rather than passive reweighting."},{"cited_title":"A decision-theoretic model of multistability: Perceptual switches as internal actions","cited_arxiv_id":null,"evidence_quote":"Proposes the explicit decision-theoretic and POMDP model of switches as internal actions, the reinforcement-learning foundation the paper wants to bridge with Bayesian accounts."},{"cited_title":"Brascamp, Raymond van Ee, Wiebe R","cited_arxiv_id":null,"evidence_quote":"Shows comparable gamma-like alternation-rate distributions across bistable paradigms, the empirical basis for treating paradigms as sharing temporal dynamics."}],"review_version":1}