Pith. sign in

REVIEW 3 major objections 4 minor 20 references

A flexible Bayesian non-parametric mixture model reveals multiple dependencies of swap errors in visual working memory

T0 review · 3 major / 4 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Swap errors in a direction-cued, location-reported working-memory task depend on distance in the report dimension, implying that some misbindings happen at encoding rather than retrieval.

desk verdict A flexible Bayesian non-parametric swap-error model that is well-built and honestly validated, but its central claim of report-dimension dependence rests on a self-referential null that deserves a cautious referee. read the letter →

arxiv 2505.01178 v1 pith:XJUVSU4L submitted 2025-05-02 q-bio.NC cs.LG

Puria Radmard , Paul M. Bays , Máté Lengyel This is my paper · ORCID
classification q-bio.NCcs.LG
keywords visualworkingmemoryswaperrorsBayesiannonparametricmixturemodelGaussianprocessfeaturebindingdelayedestimationcomparison
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper aims to determine where swap errors in visual working memory originate—whether they come from failures to encode the right feature bindings, from noise in storage, or from errors at retrieval when the cue is shown. To separate these, it introduces a Bayesian non-parametric mixture model, BNS, that lets the probability of swapping to each distractor depend freely on its distance to the target in both the probed and the reported feature dimensions, with a Gaussian-process prior over that dependence. Fitting BNS to several published delayed-estimation datasets reproduces the well-known rise in swap errors with cue similarity. In one direction-cued, location-reported dataset, however, BNS finds an additional non-monotonic modulation by distance in the reported location dimension. Because the cue is unknown when the array is encoded, the paper concludes that these particular swaps reflect misbinding at stimulus presentation, challenging the prevailing retrieval-only explanation.

What carries the argument

The load-bearing object is the swap function f, a Gaussian-process function on the displacement vector between each distractor and the cued item, x = (x_p, x_r), whose evaluations at distractor displacements become the logits of the mixture components in a circular response model, with the function equipped with a Weinland (periodic) kernel. Choosing which coordinates enter x gives model classes M ∈ {none, probe, report, both}, so comparing models amounts to asking whether swapping depends on cue-feature proximity, report-feature proximity, neither, or both. A sparse variational Gaussian-process approximation (SVGP) makes the intractable posterior over f tractable, and the model's interpretable structure lets the inferred f be read directly as evidence about the mechanism: probe-only dependence points to retrieval failure, report-dimension dependence in a location-report task points to encoding failure.

What would settle it

A targeted experiment would vary the spatial separation between items in the direction-cued, location-report task while holding the distribution of direction distances fixed; if swap rates do not rise when locations are closer, the inferred report-dimension dependence is not a genuine behavioural effect. Alternatively, refitting the Figure 7 null distribution with synthetic data generated from a probe-only model that matches the real dataset's trial counts, set sizes, and stimulus layouts, and showing that ΔBIC* against the both model no longer falls outside the null, would falsify the encoding-error conclusion.

Watch

Extended reading notes

Core claim

The paper's central discovery is a report-dimension dependence in swaps for a random-dot-motion direction-cued, location-reported task, in which the probability of swapping to a distractor increases when that distractor is spatially close to the target, with a sharp non-monotonic structure in the direction dimension as well. The authors show that a model conditioning swaps only on cue-feature distance cannot reproduce the real data's statistical structure, while a model conditioning on both probe and report distances can; the report-dimension modulation survives a null-distribution test built on BIC differences between real and synthetic data. They interpret this as evidence that some swap errors in that dataset are caused by an encoding error at stimulus presentation time—a misbinding of cue and report features before the cue is known—rather than by a failure to retrieve the correctly bound item.

Load-bearing premise

The load-bearing premise is that the probe-only model, when fitted to real data and used to generate synthetic data, faithfully reproduces the real data's statistical structure whenever swap errors truly depend only on the probe dimension; if that generative null model is misspecified, the observed report-dimension effect could arise without any true report-dimension dependence.

Editorial extensions

If this is right

  • Swap variability in the direction-cued RDK dataset is not exhausted by probe-feature similarity, so models built only on retrieval errors will leave systematic structure unexplained.
  • The inferred report-dimension dependence implies at least some swap errors originate before the cue is presented, during encoding of feature bindings.
  • The non-monotonic probe dependence for RDK directions is compatible with directions being stored as orientation-like representations; BNS discovered this without needing a predefined metric on the direction circle.
  • The BIC-based null-distribution test provides a reusable template for deciding when an added dependency in swap behaviour is real rather than a complexity artifact.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct testable extension would vary spatial separation between items in the direction-cued, location-report task while holding direction-distance distributions fixed; if swap rates do not rise with spatial proximity, the inferred encoding-stage effect is not genuine.
  • Applied to other two-feature delayed-estimation tasks, BNS may uncover report-dimension dependencies that one-dimensional mixture models would misattribute to cueing errors, potentially reassigning some known swap-error effects to encoding.
  • Because the swap function is pooled across subjects, the report-dimension effect could be driven by a subset of participants or by low-coherence trials; subject-level or hierarchical fits with more data would test this.
  • The validity of the Figure 7 null distribution rests on the probe-only model being a faithful generative description whenever no true report-dimension effect exists; checking the null model's adequacy directly is needed before the encoding-error conclusion is fully secured.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper introduces BNS, a Bayesian non-parametric mixture model of swap errors in visual working memory. The response distribution is a mixture over remembered items, with component weights determined by a Gaussian-process swap function of the circular displacement between distractor and cue in the probe and/or report feature dimensions. Sparse variational GP inference is used, and model comparison is based on a per-subject BIC approximation. Fitted to several existing datasets, the model recovers probe-similarity dependence in orientation-cued and Schneegans & Bays data, and reveals a non-monotonic probe dependence together with a report-location dependence in an RDK direction-cued, location-report dataset. The authors interpret the report-dimension dependence as evidence for an encoding-stage error mechanism, challenging retrieval-based accounts of swap errors.

Significance. If the report-dimension dependence were firmly established, the paper would make an important empirical contribution by challenging the dominance of retrieval-based explanations of swap errors, and the BNS model itself would be a useful flexible descriptive tool. The clear generative specification and the synthetic-data recovery experiments in Figures 2 and 3 are genuine strengths. However, the central validation test for the report-dimension effect is fragile, and the current evidence is not sufficient to support the strong mechanistic conclusion drawn in the Discussion.

major comments (3)
  1. [Results validation and model recovery; Figure 7 and Eq. (6)] The null distribution for ΔBIC* is generated by fitting the probe-only model to the real data, sampling synthetic datasets from that fitted model, and refitting candidate models. This parametric bootstrap is only valid if the probe-only model captures all statistical structure of the real data except the report-dimension dependence. The paper itself shows in Figure 2C that the both model retains spurious report-dimension modulation when fitted to probe-only synthetic data at realistic trial counts, so the null distribution is expected to contain such artifacts. The test therefore demonstrates only that real data differ from synthetic probe-only data in some respect that the both model captures; it does not isolate the report dimension as the cause. A control using null data generated from a both model with the report-dimension component set to zero, or a posterior predictive check on the report-dimension marginal, is needed to support the claim.
  2. [Discussion ('this implies swap errors in this dataset were made due to an error in encoding at stimulus presentation…] This conclusion overstates what the model can show. Earlier in the paper, in the section 'Relation to mechanisms underlying swap errors', the authors correctly note that 'we have not ruled out the possibility of a report dimension dependence also arising from retrieval error'. The modeling result establishes a statistical dependence in the report dimension, not a processing stage. The conclusion should be phrased as consistency with an encoding contribution rather than as a direct implication.
  3. [Model comparison; Eq. (4)] The disaggregated BIC omits the dimensionality of the variational parameters ψ and is described in the text as 'largely heuristic'. This is load-bearing because model recovery in Figure 3 shows that a true 'both' model is not recovered at realistic trial counts, which motivates the alternative test in Figure 7. The authors should assess the sensitivity of the Figure 7 conclusions to the BIC penalty, for example by also reporting cross-validated log-likelihood or a penalty based on the effective number of parameters.
minor comments (4)
  1. [Equation (3) and surrounding text] The denominator contains 'ee ˜πn', which appears to be a typo for e^{\tilde{\pi}_n}; similarly, 'over likelihood function' should read 'overall likelihood function'.
  2. [Figure 2 caption] The caption contains the typo 'miniumum' and the text has 'indiviaul'; these should be corrected.
  3. [Abstract and terminology] The term 'Bayesian non-parametric' is used although the GP prior has kernel hyperparameters; clarifying that 'non-parametric' refers to the swap function f, not to the absence of all parameters, would help readers.
  4. [Results, Mixed dependence] The sentence 'Figure 5C shows a similar degree as modulation for M = report as for probe' is grammatically unclear and should be reworded.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: BNS is a descriptive model fit to behavioural data, and the report-dimension claim is supported by synthetic-data validation rather than by definitional equivalence or self-citation.

full rationale

The paper is an empirical, model-based analysis rather than a derivation from first principles, so most circularity patterns do not apply. The central claim—report-dimension swap dependence in the RDK dataset—comes from fitting the 'both' model to real behavioural data and comparing it with one-dimensional alternatives via BIC. This is a standard descriptive model comparison: fitted parameters are not re-labelled as predictions, and the procedure is validated on synthetic data with known ground truth in Figures 2 and 3. The Figure 7 validation uses a parametric bootstrap null generated from the probe-only model fitted to real data, which is self-referential in the sense that the null inherits any misspecification of the fitted probe-only model; however, this is a concern about statistical validity of the null distribution, not circularity by construction. The null model does not contain report-dimension swap dependence, so observing a significant ΔBIC* is not logically equivalent to the conclusion. The paper's admission that the 2D model retains residual report-dimension modulation on probe-only synthetic data is a caveat about interpretability of the real-data fit, not a definitional reduction of the claim to its input. Citations to prior work by the same authors are used for baseline models, datasets, and neural-comparison models; none is invoked as an authority that forces the central conclusion, and no uniqueness theorem or ansatz is imported through a self-citation chain. The central finding therefore has independent empirical content and does not reduce to its inputs by construction.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The model introduces a latent swap function f with a GP prior, plus learned kernel hyperparameters and subject-level emission parameters. No new physical or neural entities are postulated. The main additional assumptions are the mixture emission structure and the shared-across-subjects swap function, both acknowledged as limitations or inherited from prior models.

free parameters (5)
  • Kernel amplitude σω^2 = estimated per dataset
    Controls the overall magnitude of swap-function modulation in the GP kernel (Eqs. 1-2). Learned from data.
  • Kernel roughness τ (or τp, τr) = estimated per dataset
    Controls the smoothness and range of the swap function in probe and report dimensions. Learned from data.
  • Kernel noise σ0^2 = estimated per dataset
    Variance of the independent-noise term in the GP kernel (Eq. 1). Learned from data.
  • Uniform guess logit π0 = estimated per dataset
    Logit of the uniform response component, learned when the uniform component is included in the model.
  • Per-subject emission parameters (κ or α, γ) = estimated per subject
    Concentration (von Mises) or stability and scale (wrapped stable) of the response error distribution. Subject-specific.
assumptions (5)
  • domain assumption Responses are generated as a mixture of item-centered emission distributions plus an optional uniform component.
    This is the standard structure of VWM mixture models (Bays et al., 2009; Bays, 2016), assumed rather than derived. It constrains all conclusions about component weights.
  • ad hoc to paper The swap function f is a draw from a zero-mean Gaussian process with the specified kernel (Weinland kernel for circular inputs).
    The GP prior is the model's core assumption; the Weinland kernel (Eq. 2) is chosen for circular inputs without mechanistic justification beyond smoothness.
  • ad hoc to paper All subjects in a dataset share the same swap function f.
    The authors state this pooling is forced by data limitations and can cause underfitting; it is acknowledged as a limitation in the Discussion.
  • ad hoc to paper The disaggregated BIC (Eq. 4) is a valid model comparison metric for subject-level comparisons, despite omitting variational parameter dimensionality.
    The authors call it 'largely heuristic' and rely on synthetic calibration to correct for the resulting bias.
  • domain assumption Retrieval-stage noise cannot produce a swap dependence on report-feature distance, so that dependence, if real, indicates an encoding error.
    The authors explicitly state 'we have not ruled out the possibility of a report dimension dependence also arising from retrieval error,' making this interpretive step an unresolved assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A flexible Bayesian non-parametric mixture model reveals multiple dependencies of swap errors in visual working memory." pith.science (2026). https://pith.science/paper/XJUVSU4L

@misc{pith2026250501178,
  author       = {Pith},
  title        = {Pith review of: A flexible Bayesian non-parametric mixture model reveals multiple dependencies of swap errors in visual working memory},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XJUVSU4L}},
  note         = {Machine review of arXiv:2505.01178}
}
read the original abstract

Human behavioural data in psychophysics has been used to elucidate the underlying mechanisms of many cognitive processes, such as attention, sensorimotor integration, and perceptual decision making. Visual working memory has particularly benefited from this approach: analyses of VWM errors have proven crucial for understanding VWM capacity and coding schemes, in turn constraining neural models of both. One poorly understood class of VWM errors are swap errors, whereby participants recall an uncued item from memory. Swap errors could arise from erroneous memory encoding, noisy storage, or errors at retrieval time - previous research has mostly implicated the latter two. However, these studies made strong a priori assumptions on the detailed mechanisms and/or parametric form of errors contributed by these sources. Here, we pursue a data-driven approach instead, introducing a Bayesian non-parametric mixture model of swap errors (BNS) which provides a flexible descriptive model of swapping behaviour, such that swaps are allowed to depend on both the probed and reported features of every stimulus item. We fit BNS to the trial-by-trial behaviour of human participants and show that it recapitulates the strong dependence of swaps on cue similarity in multiple datasets. Critically, BNS reveals that this dependence coexists with a non-monotonic modulation in the report feature dimension for a random dot motion direction-cued, location-reported dataset. The form of the modulation inferred by BNS opens new questions about the importance of memory encoding in causing swap errors in VWM, a distinct source to the previously suggested binding and cueing errors. Our analyses, combining qualitative comparisons of the highly interpretable BNS parameter structure with rigorous quantitative model comparison and recovery methods, show that previous interpretations of swap errors may have been incomplete.

Figures

Figures reproduced from arXiv: 2505.01178 by the authors.

Figure 1
Figure 1. Graphical model of BNS, described in detail in the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Variational approximations of the swap function posterior given synthetic data [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. ∆BICs(M ;M ′ ) (average per-trial difference; see main text) for all combinations of model classes. The swap function amplitude refers to modulation of swap dependence in the generating model parameters. Some examples of different amplitudes are shown in Figure 2A. The data generating model M is denoted by the dots on the abscissa. In most cases, each model is fit with data from 10 ‘synthetic subjects’, each perform… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: ∆BIC∗ s (see main text) on various datasets. In all cases, the baseline model M ′ is a none (flat) swap function with a von Mises emission and a uniform component (‘vM. + unif’), and ∆BIC is compared for other swap function forms and for the wrapped stable (‘ws’) emiss…
Figure 5
Figure 5. Figure 5: Variational posterior mean functions µq = Eq[ f ] for one dimensional models in the A: Schneegans & Bays (2016) separated by role of each feature, and B orientation￾cued and C direction-cued McMaster et al. (2022) datasets, separated by stimulus strengths (elongation a…
Figure 6
Figure 6. Figure 6: Mean of f when fit to McMaster et al. (2022) direction-cued RDK dataset (A) and orientation cued ellipse dataset (B) - location is recalled in both cases. Consult [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Please consult Figure 3 for interpreting the [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 11 canonical work pages

  1. [1]

    J., & Johnston, W

    Alleman, M., Panichello, M., Buschman, T. J., & Johnston, W. J. (2024, August). The neural basis of swap er- rors in working memory. Proceedings of the National Academy of Sciences , 121(33). Retrieved from http:// dx.doi.org/10.1073/pnas.2401032121 doi: 10.1073/ pnas.2401032121

    work page doi:10.1073/pnas.2401032121 2024

  2. [2]

    Bays, Catalao, R. F. G., & Husain, M. (2009, September). The precision of visual working memory is set by allocation of a shared resource. Journal of Vision , 9(10), 7–7. Retrieved from https://doi.org/10.1167/9.10.7 doi: 10.1167/ 9.10.7

    work page doi:10.1167/9.10.7 2009

  3. [3]

    Y ., & Husain, M

    Bays, Wu, E. Y ., & Husain, M. (2011, May). Storage and binding of object features in visual working mem- ory. Neuropsychologia, 49(6), 1622–1631. Retrieved from https://doi.org/10.1016/j.neuropsychologia .2010.12.023 doi: 10.1016/j.neuropsychologia.2010.12 .023

    doi:10.1016/j.neuropsychologia 2011

  4. [4]

    Bays, P. M. (2016, January). Evaluating and excluding swap errors in analogue tests of working memory. Scien- tific Reports , 6(1). Retrieved from http://dx.doi.org/ 10.1038/srep19203 doi: 10.1038/srep19203

    work page doi:10.1038/srep19203 2016

  5. [5]

    J., Ardalan, A., Tsodyks, M., & Qian, N

    Cueva, C. J., Ardalan, A., Tsodyks, M., & Qian, N. (2021). Re- current neural network models for working memory of con- tinuous variables: activity manifolds, connectivity patterns, and dynamic codes. ArXiv, abs/2111.01275. Retrieved from https://api.semanticscholar.org/CorpusID: 240419711

    arXiv 2021

  6. [6]

    M., & Ferber, S

    Emrich, S. M., & Ferber, S. (2012, April). Competition in- creases binding errors in visual working memory. Journal of Vision , 12(4), 12–12. Retrieved from http://dx.doi .org/10.1167/12.4.12 doi: 10.1167/12.4.12

    work page doi:10.1167/12.4.12 2012

  7. [7]

    Gorgoraptis, N., Catalao, R. F. G., Bays, P. M., & Husain, M. (2011, June). Dynamic updating of working memory re- sources for visual objects. Journal of Neuroscience, 31(23), 8502–8511. Retrieved from https://doi.org/10.1523/ jneurosci.0208-11.2011 doi: 10.1523/jneurosci.0208 -11.2011

    work page doi:10.1523/jneurosci.0208 2011

  8. [9]

    Leibfried, F ., Dutordoir, V., John, S., & Durrande, N. (2020). A tutorial on sparse gaussian processes and variational in- ference. arXiv. Retrieved from https://arxiv.org/abs/ 2012.13962 doi: 10.48550/ARXIV.2012.13962

    arXiv doi:10.48550/arxiv.2012.13962 2020

Show all 20 references
  1. [10]

    J., & Vogel, E

    Luck, S. J., & Vogel, E. K. (1997, November). The capac- ity of visual working memory for features and conjunctions. Nature, 390(6657), 279–281. Retrieved from http:// dx.doi.org/10.1038/36846 doi: 10.1038/36846

    1997 doi

  2. [11]

    J., Husain, M., & Bays, P

    Ma, W. J., Husain, M., & Bays, P. M. (2014, February). Chang- ing concepts of working memory. Nature Neuroscience , 17(3), 347–356. Retrieved from http://dx.doi.org/ 10.1038/nn.3655 doi: 10.1038/nn.3655

    2014 doi

  3. [12]

    (2015, January)

    Matthey, L., Bays, P ., & Dayan, P . (2015, January). A probabilistic palimpsest model of visual short-term mem- ory. PLOS Computational Biology , 11(1), e1004003. Retrieved from http://dx.doi.org/10.1371/journal .pcbi.1004003 doi: 10.1371/journal.pcbi.1004003

    2015 doi

  4. [13]

    M., Tomi ´c, I., Schneegans, S., & Bays, P

    McMaster, J. M., Tomi ´c, I., Schneegans, S., & Bays, P . (2022, September). Swap errors in visual working mem- ory are fully explained by cue-feature variability. Cogni- tive Psychology , 137, 101493. Retrieved from http:// dx.doi.org/10.1016/j.cogpsych.2022.101493 doi: 10.10...

    2022

  5. [14]

    Oberauer, K., & Lin, H.-Y . (2017). An interference model of visual working memory. Psychological Review , 124(1), 21–59. Retrieved from http://dx.doi.org/10.1037/ rev0000044 doi: 10.1037/rev0000044

    2017 doi

  6. [15]

    (2008, January)

    Pewsey, A. (2008, January). The wrapped stable fam- ily of distributions as a flexible model for circular data. Computational Statistics & Data Analysis , 52(3), 1516–1523. Retrieved from http://dx.doi.org/10 .1016/j.csda.2007.04.017 doi: 10.1016/j.csda.2007 .04.017

    2008 doi

  7. [16]

    E., & Williams, C

    Rasmussen, C. E., & Williams, C. K. I. (2005). Gaussian processes for machine learning . The MIT Press. Re- trieved from http://dx.doi.org/10.7551/mitpress/ 3206.001.0001 doi: 10.7551/mitpress/3206.001.0001

    2005 doi

  8. [17]

    (2017, March)

    Schneegans, S., & Bays, P . (2017, March). Neural ar- chitecture for feature binding in visual working memory. The Journal of Neuroscience , 37(14), 3913–3925. Re- trieved from http://dx.doi.org/10.1523/JNEUROSCI .3493-16.2017 doi: 10.1523/jneurosci.3493-16.2017

    2017 doi

  9. [18]

    Schneegans, S., & Bays, P. M. (2016, October). No fixed item limit in visuospatial working memory. Cortex, 83, 181–193. Retrieved from http://dx.doi.org/10.1016/j.cortex .2016.07.021 doi: 10.1016/j.cortex.2016.07.021

    2016 doi

  10. [19]

    (2023, November)

    Lengyel, M. (2023, November). Optimal information load- ing into working memory explains dynamic coding in the prefrontal cortex. Proceedings of the National Academy of Sciences , 120(48). Retrieved from http://dx.doi .org/10.1073/pnas.2307991120 doi: 10.1073/pnas .2307991120

    2023 doi

  11. [20]

    (2014, March)

    Swan, G., & Wyble, B. (2014, March). The binding pool: A model of shared neural resources for distinct items in vi- sual working memory. Attention, Perception, & Psy- chophysics, 76(7), 2136–2157. Retrieved from https:// doi.org/10.3758/s13414-014-0633-3 doi: 10.3758/ s134...

    2014 doi

  12. [21]

    J., & Y ang, G

    Xie, Y ., Duan, Y ., Cheng, A., Jiang, P ., Cueva, C. J., & Y ang, G. R. (2023, March). Natural constraints explain working memory capacity limitations in sensory-cognitive models. Cold Spring Harbor Laboratory. Retrieved from http://dx .doi.org/10.1101/2023.03.30.534982 doi: ...

    2023 doi

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.