{"id":"8bd6c192-98d8-49f5-9413-066bae2372e5","arxiv_id":"2506.19057","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The authors propose that metacognitive uncertainty judgments are posterior-like higher-order representations combining likelihood-like estimates of current noise with prior-like expectations about typical noise.","lead":"This paper proposes that the brain's uncertainty about its own perceptions is built like a Bayesian inference, combining a fresh estimate of current noise with learned expectations about typical noise. It argues that these components leave separate neural traces that can be measured with existing and emerging brain-imaging methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The likelihood-like HOR has no proposed measurement, so the Bayesian decomposition of confidence is underdetermined by the available data.","rationale":"The reader's weakest assumption was that likelihood-like and prior-like HORs exist as separable, measurable neural states; I agree that this is the key empirical gap. However, I would sharpen it: the paper does not merely lack evidence for separability—it explicitly concedes that its principal measurement tool for the likelihood component (TAFKAP/PRINCE) targets FOR uncertainty, not the likelihood-like HOR. Moreover, even if such states exist, the proposed Bayesian combination rule is not identifiable from confidence reports alone unless each component is independently constrained. This makes the central claim underdetermined rather than merely unverified. The paper earns credit for acknowledging some of these limits, especially in the TAFKAP discussion and in the footnote that allows a non-Bayesian comparison process as a fallback. But that fallback weakens the headline claim: the paper's own caveat shows that the posterior-like Bayesian structure is not necessary for its broader thesis about anchoring. Given that this is a perspective/proposal paper rather than a report of new empirical results, my concern does not justify rejection; it does, however, reinforce the reader's conditional verdict. The appropriate next step is a targeted experiment that independently measures or manipulates the likelihood and prior components and tests the Bayesian combination against simpler alternatives.","tokens_in":16578,"tokens_out":3389,"duration_ms":37781,"concrete_test":"Run a perceptual decision experiment with trial-wise stimulus contrast or noise (varying likelihood) and two blocks with different noise statistics (varying prior). On each trial, measure TAFKAP-decoded sensory uncertainty from early visual cortex, confidence reports, and optionally a NERD-style prior estimate per block. Compare three models of confidence via cross-validation: (1) Bayesian posterior readout combining decoded uncertainty and block-specific prior; (2) direct noisy readout of decoded uncertainty with no prior term; (3) a simple weighted sum of the two regressors. If model (1) does not outperform model (2) out-of-sample, the central Bayesian decomposition loses its empirical support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that confidence reports are posterior-like HORs formed by a Bayesian combination of likelihood-like and prior-like HORs (Box 1, Fig. 1). The load-bearing premise is that these components are separable, measurable neural states. That premise is not currently supported, and the paper itself concedes the key gap: it states that TAFKAP and PRINCE measure FOR uncertainty, 'rather than the (likelihood-like) HOR about that uncertainty.' GLMsingle and GSN measure voxel noise, not FOR noise. Thus, no proposed method in the paper actually measures the likelihood-like HOR. The prior-like HOR rests entirely on NERD, which is an analogy between denoising diffusion and DecNef learning and is supported by an unreviewed preprint; it does not demonstrate that the brain stores a separable learned distribution over FOR uncertainty. As a result, the posterior equation p(uncertainty_FOR | estimate_HO) ∝ p(estimate_HO | uncertainty_FOR) p(uncertainty_FOR) has no independently observable likelihood term. Confidence reports alone are compatible with many non-Bayesian or non-hierarchical processes, so the proposed Bayesian decomposition is not identifiable from the measurements the paper offers. The framework may be a useful organizing metaphor, but the central empirical claim—that brains build confidence this way—is currently underdetermined.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes that metacognitive estimates of uncertainty are not direct readouts of first-order uncertainty but rather the product of a hierarchical Bayesian inference process over the brain's own representations. It decomposes higher-order representations (HORs) of uncertainty into three components: likelihood-like HORs (momentary estimates of current FOR uncertainty), prior-like HORs (learned expectations about typical FOR noise), and posterior-like HORs (the integrated result that drives confidence reports). The authors survey existing methods (GLMsingle, GSN, TAFKAP/PRINCE, NERD) and argue that these can be adapted or extended to isolate the hypothesized components. The paper is explicitly a conceptual proposal and includes no new experimental data or quantitative model, but it offers a concrete framework and a research agenda for separating the contributions of current noise estimates and learned noise priors to metacognitive judgments.","tokens_in":16831,"tokens_out":2160,"duration_ms":23584,"significance":"If the proposed decomposition is correct, it would reframe the study of metacognition and confidence as a hierarchical inference problem, aligning uncertainty monitoring with the Bayesian brain framework and connecting metacognition to generative-model and reinforcement-learning approaches. The paper is clearly written and unusually transparent about its own limitations; for example, it explicitly notes that TAFKAP and PRINCE measure FOR uncertainty rather than the likelihood-like HOR about that uncertainty, and that GLMsingle/GSN measure voxel noise rather than FOR noise. That transparency is a strength. However, the framework's central empirical claim—that brains literally maintain separable likelihood-like and prior-like HORs and combine them multiplicatively—is currently supported mainly by analogical arguments and by the authors' own unpublished work (NERD). No method described in the paper independently measures the likelihood-like HOR, and no direct empirical test separating the three components is offered.","major_comments":[{"comment":"The central equation p(uncertainty_FOR | estimate_HO) ∝ p(estimate_HO | uncertainty_FOR) p(uncertainty_FOR) is not identifiable from the measurements the paper describes. The 'likelihood' section concedes that TAFKAP and PRINCE estimate FOR uncertainty 'rather than the (likelihood-like) HOR about that uncertainty,' and that GLMsingle and GSN measure voxel noise, not FOR noise. The paper does not propose any way to observe or estimate the likelihood-like HOR itself. Without an independent observable for this term, confidence reports alone are compatible with many non-Bayesian or non-hierarchical models (e.g., meta-d' or CASANDRE), and the Bayesian decomposition is underdetermined. Please specify a concrete measurement protocol or a set of falsifiable predictions that would distinguish the proposed hierarchical Bayesian process from a direct, non-hierarchical readout of first-order uncertainty.","section":"Box 1; The 'likelihood' section"},{"comment":"The prior-like HOR rests almost entirely on the NERD model, which is supported by an unreviewed preprint (Azimi Asrari & Peters, 2025) and a conference abstract (Azimi Azrari & Peters, 2024). The analogy between denoising diffusion models and DecNef learning is interesting, but the paper does not demonstrate that the brain stores a separable, learnable distribution over FOR uncertainty, nor that NERD's learned noise distribution corresponds to a neural prior-like HOR. The claim that 'the lower-dimensional prior-like uncertainty HORs discovered by NERD could indeed capture individual variation' is presented without details of the analysis, sample size, or statistical results. Please clarify what evidence would confirm or refute the existence of a prior-like HOR and provide details of the NERD results so that readers can assess this load-bearing claim.","section":"The 'prior' section"},{"comment":"The independence and multiplicative combination of the likelihood-like and prior-like components is assumed as an axiom, but no justification is given for why these two HOR types are independent or why they combine as a product. In standard Bayesian perception, the independence assumption is motivated by generative models of the environment; here, the 'estimate_HO' variable is not precisely defined, and it is not clear what physiological or computational constraint would enforce independence between a current noise estimate and a learned noise prior. Please state the conditions under which the multiplicative decomposition would fail (e.g., correlated noise estimates or context-dependent priors) and describe how a failure would be detected empirically.","section":"Box 1; Figure 1"}],"minor_comments":[{"comment":"There is a typo in the sentence 'the relationship between voxelwise noise and FOR noise is as complex a the relationship between voxelwise patterns and the mental structures they represent'; the word 'as' should be 'as the'.","section":"The 'likelihood' section"},{"comment":"The text says GLMsingle quantifies 'how much a given voxel's activity is predicted by a task-relevent variable'; 'relevent' should be 'relevant'.","section":"The 'likelihood' section"},{"comment":"The figure label 'reward calculatuion' contains a typo: 'calculatuion' should be 'calculation'.","section":"Figure 2"},{"comment":"The author name is spelled inconsistently: 'Azimi Asrari' in the text and reference list appears as 'Azimi Azrari' in the 2024 conference abstract reference; please standardize.","section":"References"},{"comment":"The phrase 'inferiortemporal cortex' should be 'inferior temporal cortex'.","section":"The 'prior' section"}],"recommendation":"major_revision","confidential_remarks":"This is a well-written perspective piece, but the central empirical claim currently rests on an unpublished preprint and on an analogy that has not been independently validated. The paper's own admissions about TAFKAP/PRINCE and GLMsingle/GSN mean that the likelihood-like HOR is, by the authors' own description, unmeasured. If the authors can either provide a concrete measurement protocol or explicitly reframe the paper as a purely conceptual proposal without the empirical loading, the contribution would be clearer. I would not reject it, but the load-bearing gaps in evidence and identifiability need to be addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is a well-written perspective paper that splits HORs of uncertainty into three Bayesian components—likelihood-like, prior-like, posterior-like—and pairs each with candidate measurement tools. That packaging is genuinely new, even if the underlying Bayesian account of confidence has antecedents in Fleming & Daw and others. The paper is honest and careful, especially when it concedes that TAFKAP/PRINCE measure FOR uncertainty rather than the likelihood-like HOR, and that GLMsingle/GSN measure voxel noise, not FOR noise. That kind of candor is rare.\n\nWhat it does well: it organizes a messy literature into a testable computational skeleton and connects it to engineering methods that most metacognition researchers won't know well. The figure and Box 1 are clear. The authors also avoid overclaiming—they frame the decomposition as a proposal, not an established result.\n\nThe soft spots are real but proportionate. The stress-test note is correct: no proposed method in the paper actually measures the likelihood-like HOR. The prior-like HOR rests heavily on the NERD model, which is an unreviewed preprint and an analogy to diffusion models. As a result, the posterior equation has no independently observable likelihood term, and confidence reports alone are compatible with many non-Bayesian processes. This does not kill the paper, because the genre is conceptual, but it does mean the empirical payoff is deferred. The authors could strengthen this significantly by specifying a concrete measurement scheme for the likelihood-like HOR or stating what would falsify the decomposition.\n\nFor whom: metacognition and consciousness researchers who want a map of possible computational ingredients and a bridge to neuroimaging/neurofeedback methods. I'd bring it to a reading group but would frame it as a framework-generating piece, not a result. A serious referee should engage with it—the measurement gap is the central thing to push on, and the authors are already halfway to acknowledging it.","headline":"A clear conceptual reorganization of Bayesian confidence with a real measurement gap for the likelihood-like component; worth reading but not as an empirical claim.","tokens_in":17332,"tokens_out":1272,"would_cite":true,"duration_ms":15525,"reading_group":"yes","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 proposes that reported confidence is not a direct readout of first-order uncertainty but the output of a hierarchical Bayesian inference that combines a current noise estimate with learned expectations about typical noise.","keywords":["higher-order representations","metacognition","Bayesian inference","confidence","uncertainty","probabilistic population codes","decoded neurofeedback","neural representations"],"falsifier":"Run a perceptual task in which FOR uncertainty is decoded trial by trial from early visual cortex while the experimenter manipulates the distribution of noise the observer experiences over blocks. If confidence reports are unchanged whenever decoded FOR uncertainty is held constant, even when the experienced noise distribution changes, the claim that prior-like HORs enter confidence is falsified. The same experiment would also fail if neural patterns carrying the learned noise distribution cannot be recombined with patterns carrying current noise to predict confidence.","tokens_in":16373,"feed_emoji":"🧠","tokens_out":6879,"duration_ms":66630,"temperature":0.7,"pith_summary":"The paper proposes that feelings of confidence are not direct reads of how uncertain the current sensory representation is. Instead, the brain is argued to run a second Bayesian inference about its own first-order representation, combining a momentary estimate of how noisy that representation is with a learned expectation of how noisy it usually is in that context. The result is a 'posterior-like' higher-order representation whose readout is reported confidence. The authors argue that isolating the two ingredients is necessary to explain metacognitive calibration, learning without feedback, and debates about higher-order theories of consciousness, and they survey methods that could measure each ingredient separately.","feed_headline":"Confidence is a Bayesian blend of current and expected noise","feed_subtitle":"Feeling confident requires estimating how noisy the signal is now and how noisy it usually is, then combining the two.","key_machinery":"The carrying object is the hierarchical Bayesian decomposition of uncertainty HORs into three named components: likelihood-like HORs, which estimate the momentary reliability of a current first-order representation; prior-like HORs, which encode learned expectations about typical noise along task-relevant dimensions; and posterior-like HORs, which integrate the two to form experienced uncertainty. The formal identity is $p(\\mathrm{uncertainty_{FOR}} | \\mathrm{estimate_{HO}}) \\propto p(\\mathrm{estimate_{HO}} | \\mathrm{uncertainty_{FOR}})\\, p(\\mathrm{uncertainty_{FOR}})$, and the paper uses it to argue that confidence reports are products of two inferential stages rather than direct reads. The machinery also includes the analytical tools proposed for measuring each component: probabilistic population codes and TAFKAP-style decoding for the likelihood-like term, and the NERD diffusion model trained on decoded neurofeedback data for the prior-like term.","core_discovery":"The central claim is that metacognitive confidence is not a direct readout of first-order uncertainty. The brain is proposed to build a second-order Bayesian posterior over its own uncertainty: a current, noisy estimate of how unreliable the first-order representation is (a likelihood-like HOR) is combined with a learned distribution over how unreliable that kind of representation usually is (a prior-like HOR), and the resulting posterior-like HOR is what gets read out as confidence. The paper expresses this as $p(\\mathrm{uncertainty_{FOR}} | \\mathrm{estimate_{HO}}) \\propto p(\\mathrm{estimate_{HO}} | \\mathrm{uncertainty_{FOR}})\\, p(\\mathrm{uncertainty_{FOR}})$. It argues that nearly all existing work studies only the posterior side of this chain, and that isolating the likelihood-like and prior-like components is required to explain dissociations between actual and reported uncertainty and to arbitrate between theories of metacognition and consciousness.","pith_inferences":["If the decomposition is right, the established dissociation between objective and subjective uncertainty in the visual periphery is most naturally a mismatch between the prior-like HOR and the current likelihood-like HOR; this could be tested by retraining the prior while leaving the sensory representation untouched.","The same Bayesian HOR structure could be extended to other properties of first-order representations, such as signal strength, source (external versus internal), or content, giving each higher-order theory of consciousness a measurable set of dimensions.","A direct test would train subjects in two environments with different noise statistics but identical task stimuli, then measure confidence on probe trials where decoded FOR uncertainty is matched: if confidence tracks the training environment, prior-like HORs are causally implicated."],"forward_implications":["Confidence reports should be treated as posterior readouts, not direct measurements of first-order uncertainty, so experiments that use confidence to infer sensory noise need to control for prior-like expectations.","Individual differences in metacognitive calibration can be explained by different learned noise priors rather than by different sensitivity to current uncertainty, which changes how metacognitive training would be designed.","Confidence models that add a single noise term to a decision variable, such as meta-d'-type models, are incomplete by this account because they collapse the likelihood-like and posterior-like stages that the paper separates.","Learning tasks without external feedback, such as decoded neurofeedback, can be reinterpreted as updating prior-like uncertainty distributions, making the NERD model a candidate mechanism for how such learning proceeds."],"supporting_citations":[{"why":"Supplies the empirical precedent that learned expectations about FOR uncertainty as a function of eccentricity can cause dissociations between actual and reported uncertainty.","marker":"Winter & Peters (2022)"},{"why":"Shows that FOR uncertainty decoded from visual cortex predicts behavioral variability, the foundation for reading out likelihood-like HORs.","marker":"van Bergen et al. (2015)"},{"why":"Provides TAFKAP, the improved decoding method proposed for estimating trial-by-trial FOR uncertainty.","marker":"van Bergen & Jehee (2021)"},{"why":"Supplies the probabilistic population code framework in which uncertainty is encoded in population gain, grounding the likelihood-like HOR concept.","marker":"Ma et al. (2006)"},{"why":"Introduces the NERD model for learning noise distributions from neurofeedback data, the paper's proposal for sampling prior-like HORs.","marker":"Azimi Asrari & Peters (2025)"},{"why":"Provides the DecNef reinforcement-learning paradigm from which NERD's prior-like noise distributions are derived.","marker":"Shibata et al. (2011)"},{"why":"Demonstrates the psychophysical decomposition of a posterior percept into likelihood and prior that the paper adapts to the metacognitive domain.","marker":"Stocker & Simoncelli (2006)"},{"why":"Supplies the meta-d' framework that models confidence as a noise-limited readout but, in the paper's view, collapses the likelihood-like and posterior-like components.","marker":"Maniscalco & Lau (2012)"}],"fun_headline_variants":["Confidence: a Bayesian mix of now and usual noise","How brains blend current and expected uncertainty","Metacognition as Bayesian posterior over your own noise","Brain's confidence: combining likelihood and prior of noise","Your confidence is a posterior over your own uncertainty"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument depends on the brain storing two distinguishable kinds of uncertainty information at once: how noisy the current signal feels and how noisy that kind of signal usually is, and on those being separable, measurable neural states. If they are not separable, the Bayesian decomposition is a mathematical metaphor with no neural target.","fun_headline_variants_meta":{"raw":{"variants":["Confidence: a Bayesian mix of now and usual noise","How brains blend current and expected uncertainty","Metacognition as Bayesian posterior over your own noise","Brain's confidence: combining likelihood and prior of noise","Your confidence is a posterior over your own uncertainty"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000172,"raw_usage":{"total_tokens":1250,"prompt_tokens":894,"completion_tokens":356,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":283}},"tokens_in":510,"tokens_out":356,"duration_ms":4044,"temperature":1.0,"reasoning_tokens":283,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:37:16.282148+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a perceptual task in which FOR uncertainty is decoded trial by trial from early visual cortex while the experimenter manipulates the distribution of noise the observer experiences over blocks. If confidence reports are unchanged whenever decoded FOR uncertainty is held constant, even when the experienced noise distribution changes, the claim that prior-like HORs enter confidence is falsified. The same experiment would also fail if neural patterns carrying the learned noise distribution cannot be recombined with patterns carrying current noise to predict confidence.","supporting_citations":[],"review_version":2}