Pith. sign in

REVIEW 2 major objections 2 minor 42 references

AdaFusion: Prompt-Guided Inference with Adaptive Fusion of Pathology Foundation Models

T0 review · 2 major / 2 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read In the delayed Vicsek model at high agent speed, time delay acts as a control parameter that broadens the phase-separated noise window.

desk verdict The file under review is actually a delayed-Vicsek simulation paper, not the AdaFusion abstract; the physics is interesting and mostly sound, but the main novel phase boundary lacks error bars and the package is mismatched. read the letter →

arxiv 2508.05084 v2 pith:R62AD64M submitted 2025-08-07 cs.CV

Yuxiang Xiao , Yang Hu , Bin Li , Tianyang Zhang , Zexi Li , Huazhu Fu , Jens Rittscher , Kaixiang Yang This is my paper · ORCID
classification cs.CV
keywords Vicsekmodeltimedelayactivematterphaseseparationorder-disordertransitioncollectivemotiontravelingbandsdelay-inducedcontrol
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 numerically studies the Vicsek model with delayed alignment interactions at a relatively high agent speed and a large fixed system size. It claims that delay, while preserving the three standard phases (ordered, liquid–gas coexistence, disordered), qualitatively changes their stability: the upper critical noise for the transition to the disordered state rises with delay, the lower critical noise for the ordered–coexistence transition responds non-monotonically, and as a result the noise window in which phase separation occurs widens. Delay also accelerates band formation, increases the number of bands, and generates swirling arcs whose radius grows with delay. The broader implication is that time delay can serve as a control parameter for tuning the dynamic phase behavior of active matter systems.

What carries the argument

The delayed Vicsek model itself, where each agent aligns to the average orientation of neighbors it perceived $\tau$ time steps earlier, with random noise. The analysis uses standard order parameters (average polarization, polarization variance, Binder cumulant) plus the height $C_1$ of the first positive peak of the directional density autocorrelation function, computed from a coarse 13-bin histogram, to locate the phase boundaries. The key mechanism is the reduction of effective interaction time at long delays, which weakens collision-induced flocking and produces large-radius swirling arcs that seed quicker band formation.

What would settle it

Recompute the phase boundaries $\eta_{o|s}$ and $\eta_{s|d}$ using a finer directional density histogram or many more snapshots (or a different order parameter such as local density correlation) and check whether the non-monotonic dependence of $\eta_{o|s}$ and the monotonic rise of $\eta_{s|d}$ persist; alternatively, run simulations at larger system sizes to see whether the $C_1$ jumps sharpen or move.

Watch

Extended reading notes

Core claim

In the delayed Vicsek model with speed $v_0=0.5$, reduced delay times $\bar\tau$ from 0 to 2.5, and system sizes $N=65536$ and $131072$ at densities $\rho=1$ and $2$, the authors find that the order–disorder transition remains discontinuous and bistable for all delays. The critical noise $\eta_{o|s}$ for the ordered-to-coexistence transition first increases then decreases with delay, while $\eta_{s|d}$ for the coexistence-to-disorder transition increases monotonically and appears to saturate, broadening the coexistence noise interval $\eta_{s|d}-\eta_{o|s}$. In this phase-separated state, the number of traveling bands increases and the stripe-formation time decreases with delay, an accelerat

Load-bearing premise

The phase boundaries that support the claims are read off from jumps in the peak height $C_1$ of a one-dimensional density autocorrelation built from only thirteen bins and averaged over ten snapshots per noise value, so any shift in where the jump occurs would change the reported trends.

Editorial extensions

If this is right

  • If delay broadens the coexistence window, natural or synthetic swarms with longer perception–reaction lags may spend more time in phase-separated, banded configurations.
  • The saturating $\eta_{s|d}$ suggests that beyond a moderate delay, further delay no longer stabilizes the ordered state, giving a finite operating window for delay-based control.
  • The delay-induced shortening of band formation time with growing swirl radius offers a mechanistic route to accelerate pattern formation in active matter.
  • The near-independence of polarization relaxation time from delay separates time scales: delay changes phase boundaries and band dynamics without slowing global ordering kinetics.
  • At high speeds, delay can reversibly switch the system between ordered and phase-separated states, offering a control knob complementary to noise.

Reading between the lines

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

  • If the non-monotonic $\eta_{o|s}$ reflects the susceptibility trade-off proposed for slow agents, then at intermediate delays the system may be most responsive to external perturbations—a testable prediction for experiments with tunable feedback delay.
  • The observed saturation of order-parameter features with delay hints that the effective delay-controlled parameter might be $v_0 \tau$; if so, different speed–delay combinations with the same product could produce similar phase behavior.
  • The observed bistability between parallel and cross-sea band configurations suggests the delayed VM may host multiple coexisting banded attractors; classifying their stability boundaries would settle whether these are transient or genuinely multistable states.
  • One could test the swirl-radius mechanism experimentally by measuring the typical curvature of transient density arcs in light-controlled bacterial or robotic swarms with imposed feedback delay.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 2 minor

Summary. The manuscript presents a numerical study of the delayed Vicsek model at high speed (v0=0.5) and large system sizes (N=65,536 and 131,072), mapping the phase diagram in the (noise η, reduced delay τ̃) plane. Using polarization, its variance, the Binder cumulant, and a directional density autocorrelation peak height C1, the authors identify ordered, liquid-gas coexistence, and disordered phases. They report that the upper boundary η_{s|d} increases monotonically with delay, while the lower boundary η_{o|s} is non-monotonic, leading to a broadening of the coexistence window. They also report that the number of traveling bands increases and the band-formation time decreases with delay, attributed to the emergence of large swirls. The paper claims time delay acts as an effective control parameter for phase behavior.

Significance. The results are potentially significant for active matter physics, as they extend the delayed Vicsek model to a high-speed regime and propose delay as a control parameter. The study is based on large-scale simulations and uses multiple order parameters, including a Binder cumulant and susceptibility, which are standard tools. However, the key lower-phase boundary η_{o|s}, which underpins the non-monotonic claim and the broadening of the coexistence window, is determined from a coarse autocorrelation peak C1 with no error bars. The visual stripe counting is also subjective. These measurement weaknesses currently limit the strength of the conclusions.

major comments (2)
  1. [Sec. 5 and Appendix B (Eqs. B.3–B.6, Figs. 8–9)] The lower phase boundary η_{o|s}, central to the paper's main claims, is extracted from the height C1 of the first peak of the directional density autocorrelation, computed with only 13 histogram bins (Appendix B) and averaged over ten snapshots. No error bars or confidence intervals are given. The normalization in Eq. (B.6) divides by the maximum absolute value of the correlation, which itself fluctuates. This makes the location of the 'sharp increase' in C1 sensitive to bin width and sampling noise. A modest systematic shift in the C1 threshold could alter the reported non-monotonic η_{o|s} trend and even change the sign of the window broadening. The authors should re-derive η_{o|s} with finer binning, more snapshots, bootstrap estimates, and demonstrate robustness of the non-monotonic dependence.
  2. [Fig. 2f and Sec. 5] The claim that the maximum number of stripes increases with delay is based on visual inspection of snapshots. The paper itself acknowledges this is subjective and notes the ambiguity of 'cross-sea' states, where stripes move in different directions. A more objective measure (e.g., automated stripe detection or Fourier analysis) is needed to support this claim. Additionally, the stability of the cross-sea configurations is unresolved, which may affect the counting protocol. Without a reproducible criterion, this part of the phase characterization is not verifiable.
minor comments (2)
  1. [Abstract] The abstract supplied at the top of the submission describes a pathology foundation model framework (AdaFusion) and does not correspond to the manuscript text, which concerns the delayed Vicsek model. This mismatch must be corrected before any further consideration.
  2. [Sec. 4, Fig. 4] The relaxation time t_R is defined as the first time the polarization reaches half of its maximum value. For a bistable system, this first-passage quantity is not the standard relaxation time and may be dominated by a single fluctuation. The conclusion that 'delay has no significant effect' would be stronger with a more conventional autocorrelation-time estimate and statistical tests.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity; minor self-citation to Ref. [29] is explanatory only.

full rationale

The central claims—the η_o|s and η_s|d phase boundaries, their delay dependence, and band-formation dynamics—are obtained from direct numerical simulations of the delayed Vicsek model using standard order parameters: polarization (Eq. 3), Binder cumulant (Eq. 5), susceptibility (Eq. 6), and the directional-autocorrelation peak height C1 (Appendix B). These quantities are measured from the simulated trajectories and are not fitted parameters; the exponential fit in Eq. (7) is a descriptive characterization of the correlation-length decay, not a prediction built from the phase boundaries. The only self-citations are to Ref. [29] (same research group), used for the susceptibility interpretation, the saturation of order-parameter features, and the effective parameter v0τ for long delays. These cited results are used as interpretive context and to rationalize observations after the fact; they are not inputs from which the numerical phase diagram is derived. No ansatz is smuggled in via self-citation, and no uniqueness theorem is imported. The paper does contain robustness limitations—the C1 estimator uses a 13-bin histogram and ten snapshots (Appendix B), stripe counting is visual (Sec. 5), and cross-sea state stability is unresolved—but these are statistical/reliability concerns, not evidence that a result is definitionally identical to its input. Hence no circular step; score 2 reflects minor non-load-bearing self-citation.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claims rest on the delayed VM update rule, the chosen high-speed large-size simulation protocol, the initial condition for negative times, and the statistical indicators used to assign phases. The free parameters listed are the numerical outputs (phase boundaries, fitted decay times, visually counted stripe numbers) that the broadened coexistence window and accelerated band formation claims depend on, not inputs fitted to force the conclusions.

free parameters (3)
  • reduced stripe-formation time t_RS = values from exponential fits in Fig. 6, e.g., decreasing until tau_bar=3/2 then slightly increasing
    Extracted by fitting Eq. (7) to simulated radial correlation length; the claim that delay accelerates band formation depends directly on these fitted decay times.
  • phase-boundary noise values eta_o|s and eta_s|d = eta_s|d approx. 0.478, 0.613, 0.666, 0.677 for tau_bar=0, 1/2, 3/2, 5/2 at rho=2
    Numerical estimates from C1 peak height and order-parameter discontinuities; the central claim about broadening of the phase-separated window rests on their differences.
  • maximum stripe number per delay = visual counts from snapshots, e.g., increasing with delay
    Obtained by visual inspection of snapshots (Sec. 5) and explicitly subject to error; used to support the claim that band number grows with delay.
assumptions (4)
  • domain assumption Delayed alignment rule Eq. (2): agents align with neighbors perceived at time t-tau*Delta t
    The model definition itself; the study is a numerical exploration of this dynamics, not a derivation from microscopic physics.
  • ad hoc to paper High-speed parameter regime v0=0.5 (v0*Delta t/R=1/2) with N=65,536 or 131,072 and L=256
    Chosen to observe phase separation at large system size; results may not transfer to the low-speed regime studied before.
  • ad hoc to paper Initial condition: uniform positions, random orientations, same state for all negative times
    Required to iterate the delayed Eq. (2); steady-state claims rely on long simulation times (10^6 steps) rather than a convergence proof.
  • standard math Binder cumulant, polarization variance, and C1 peak height are valid indicators of the three phases
    Standard order parameters from the Vicsek-model literature (Refs [32,33,34,35,39]); used to delineate ordered, phase-separated, and disordered regimes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AdaFusion: Prompt-Guided Inference with Adaptive Fusion of Pathology Foundation Models." pith.science (2026). https://pith.science/paper/R62AD64M

@misc{pith2026250805084,
  author       = {Pith},
  title        = {Pith review of: AdaFusion: Prompt-Guided Inference with Adaptive Fusion of Pathology Foundation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R62AD64M}},
  note         = {Machine review of arXiv:2508.05084}
}
read the original abstract

Pathology foundation models (PFMs) have demonstrated strong representational capabilities through self-supervised pre-training on large-scale, unannotated histopathology image datasets. However, their diverse yet opaque pretraining contexts, shaped by both data-related and structural/training factors, introduce latent biases that hinder generalisability and transparency in downstream applications. In this paper, we propose AdaFusion, a novel prompt-guided inference framework that, to our knowledge, is among the very first to dynamically integrate complementary knowledge from multiple PFMs. Our method compresses and aligns tile-level features from diverse models and employs a lightweight attention mechanism to adaptively fuse them based on tissue phenotype context. We evaluate AdaFusion on three real-world benchmarks spanning treatment response prediction, tumour grading, and spatial gene expression inference. Our approach consistently surpasses individual PFMs across both classification and regression tasks, while offering interpretable insights into each model's biosemantic specialisation. These results highlight AdaFusion's ability to bridge heterogeneous PFMs, achieving both enhanced performance and interpretability of model-specific inductive biases.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 1 canonical work pages

  1. [29]

    Holubec V, Geiss D, Loos S A, Kroy K and Cichos F 2021 Physical Review Letters 127 258001 publisher: American Physical Society URL https://link.aps.org/doi/10.1103/ PhysRevLett.127.258001

    2021

  2. [1]

    Vicsek T and Zafeiris A 2012 Physics Reports 517 71–140 ISSN 0370-1573 collective motion URL https://www.sciencedirect.com/science/article/pii/S0370157312000968

    2012

  3. [2]

    Ramaswamy S 2010 Annual Review of Condensed Matter Physics 1 323–345 URL https: //doi.org/10.1146/annurev-conmatphys-070909-104101

    doi:10.1146/annurev-conmatphys-070909-104101 2010

  4. [3]

    Bechinger C, Di Leonardo R, L¨ owen H, Reichhardt C, Volpe G and Volpe G 2016 Reviews of Modern Physics 88 045006 URL https://link.aps.org/doi/10.1103/RevModPhys.88. 045006

    doi:10.1103/revmodphys.88 2016

  5. [4]

    Zhang H P, Be’er A, Florin E L and Swinney H L 2010 Proceedings of the National Academy of Sciences 107 13626–13630

    2010

  6. [5]

    Ben-Jacob E, Schochet O, Tenenbaum A, Cohen I, Czirok A and Vicsek T 1994 Nature 368 46–49

    1994

  7. [6]

    Cavagna A, Giardina I and Grigera T S 2018 Physics Reports 728 1–62

    2018

  8. [7]

    Shahhoseini Z, Sarvi M and Saberi M 2018 Physica A: Statistical Mechanics and its Applications 491 101–111

    2018

Show all 42 references
  1. [8]

    Helbing D, Farkas I J, Molnar P and Vicsek T 2002 Pedestrian and evacuation dynamics 21 21–58

    2002

  2. [9]

    Mijalkov M, McDaniel A, Wehr J and Volpe G 2016 Physical Review X 6 011008

    2016

  3. [10]

    V´ as´ arhelyi G, Vir´ agh C, Somorjai G, Nepusz T, Eiben A E and Vicsek T 2018Science Robotics 3 eaat3536 publisher: American Association for the Advancement of Science URL https: //www.science.org/doi/abs/10.1126/scirobotics.aat3536

    doi

  4. [11]

    Khadka U, Holubec V, Yang H and Cichos F 2018 Nature communications 9 1–9

    2018

  5. [12]

    Fruchart M, Hanai R, Littlewood P B and Vitelli V 2021 Nature 592 363–369 ISSN 1476-4687 URL https://doi.org/10.1038/s41586-021-03375-9

    2021 doi

  6. [13]

    org/10.1088/1367-2630/abcc1e

    Loos S A M and Klapp S H L 2020 New Journal of Physics 22 123051 URL https://dx.doi. org/10.1088/1367-2630/abcc1e

    2020 doi

  7. [14]

    Sch¨ oll E, H¨ ovel P, Flunkert V and Dahlem M A 2009 Time-delayed feedback control: From simple models to lasers and neural systems Complex time-delay systems (Springer) pp 85–150

    2009

  8. [15]

    Loos S 2021 Stochastic Systems with Time Delay: Probabilistic and Thermodynamic Descriptions of Non-Markovian Processes Far From Equilibrium (Springer International Publishing) ISBN 9783030807726 URL https://books.google.cz/books?id=svV6zwEACAAJ �������������� ���������� �� ��...

    2021

  9. [16]

    Davis L 2003 Physica A: Statistical Mechanics and its Applications 319 557–567 ISSN 0378-4371 URL https://www.sciencedirect.com/science/article/pii/S0378437102014577

    2003

  10. [17]

    Forgoston E and Schwartz I B 2008 Phys. Rev. E 77(3) 035203 URL https://link.aps.org/ doi/10.1103/PhysRevE.77.035203

    2008 doi

  11. [18]

    Piwowarczyk R, Selin M, Ihle T and Volpe G 2019 Physical Review E 100 012607 publisher: American Physical Society URL https://link.aps.org/doi/10.1103/PhysRevE.100.012607

    2019 doi

  12. [19]

    Geiss D, Kroy K and Holubec V 2019 New Journal of Physics 21 093014 ISSN 1367-2630 publisher: IOP Publishing URL https://dx.doi.org/10.1088/1367-2630/ab3d76

    2019 doi

  13. [20]

    Loos S A M and Klapp S H L 2019 Scientific reports 9 1–11

    2019

  14. [21]

    Mui˜ nos-Landin S, Fischer A, Holubec V and Cichos F 2021 Science Robotics 6

    2021

  15. [22]

    Wang X, Chen P C, Kroy K, Holubec V and Cichos F 2023 Nature Communications 14 56 ISSN 2041-1723 number: 1 Publisher: Nature Publishing Group URL https://www.nature.com/ articles/s41467-022-35427-7

    2023

  16. [23]

    Fr¨ anzl M, Mui˜ nos-Landin S, Holubec V and Cichos F 2021ACS Nano 15 3434–3440 ISSN 1936- 0851 URL https://doi.org/10.1021/acsnano.0c10598

    1936 doi

  17. [24]

    org/doi/10.1103/PhysRevE.91.042720

    Romanczuk P and Salbreux G 2015 Physical Review E 91(4) 042720 URL https://link.aps. org/doi/10.1103/PhysRevE.91.042720

    2015 doi

  18. [25]

    Diz-Mu˜ noz A, Romanczuk P, Yu W, Bergert M, Ivanovitch K, Salbreux G, Heisenberg C P and Paluch E K 2016 BMC Biology 14 74 ISSN 1741-7007 URL https://doi.org/10.1186/ s12915-016-0294-x

    2016

  19. [26]

    Tarama S, Egelhaaf S U and L¨ owen H 2019 Phys. Rev. E 100(2) 022609 URL https://link. aps.org/doi/10.1103/PhysRevE.100.022609

    2019 doi

  20. [27]

    Kopp R A and Klapp S H L 2023 EPL 143 17002 URL https://doi.org/10.1209/0295-5075/ acdf19

    2023 doi

  21. [28]

    Erban R, Haˇ skovec J and Sun Y 2016SIAM Journal on Applied Mathematics 76 1535–1557 URL https://doi.org/10.1137/15M1030467

    doi

  22. [30]

    Geiß D, Kroy K and Holubec V 2022 Physical Review E 106 054612 ISSN 2470-0045, 2470-0053 arXiv:2205.12069 [cond-mat] URL http://arxiv.org/abs/2205.12069

    2022 arXiv

  23. [31]

    Pakpour F and Vicsek T 2024 Physica A: Statistical Mechanics and its Applications 634 129453 ISSN 0378-4371 URL https://www.sciencedirect.com/science/article/pii/ S0378437123010087

    2024

  24. [32]

    Solon A P and Tailleur J 2013 Phys. Rev. Lett. 111(7) 078101 URL https://link.aps.org/doi/ 10.1103/PhysRevLett.111.078101

    2013 doi

  25. [33]

    Chat´ e H 2020 Annual Review of Condensed Matter Physics 11 189–212 ISSN 1947-5462 URL https://www.annualreviews.org/content/journals/10.1146/ annurev-conmatphys-031119-050752

    2020

  26. [34]

    Chat´ e H, Ginelli F, Gr´ egoire G and Raynaud F 2008 Physical Review E 77 046113 publisher: American Physical Society URL https://link.aps.org/doi/10.1103/PhysRevE.77.046113

    2008 doi

  27. [35]

    Vicsek T, Czir´ ok A, Ben-Jacob E, Cohen I and Shochet O 1995 Physical Review Letters 75 1226–1229 publisher: American Physical Society URL https://link.aps.org/doi/10.1103/ PhysRevLett.75.1226

    1995

  28. [36]

    Chat´ e H, Ginelli F, Gr´ egoire G, Peruani F and Raynaud F 2008The European Physical Journal B 64 451–456 ISSN 1434-6036 URL https://doi.org/10.1140/epjb/e2008-00275-9

    doi

  29. [37]

    Nagy M, Daruka I and Vicsek T 2007 Physica A: Statistical Mechanics and its Applications 373 445–454 ISSN 0378-4371 URL https://www.sciencedirect.com/science/article/pii/ S0378437106006510

    2007

  30. [38]

    Giraldo-Barreto J and Holubec V 2025 Active matter flocking via predictive alignment ( Preprint 2504.07778) URL https://arxiv.org/abs/2504.07778 �������������� ���������� �� ����� ������ ����� 20

    2025 arXiv

  31. [39]

    Solon A P, Chat´ e H and Tailleur J 2015 Physical Review Letters 114 068101 ISSN 0031-9007, 1079-7114 URL https://link.aps.org/doi/10.1103/PhysRevLett.114.068101

    2015 doi

  32. [40]

    K¨ ursten R and Ihle T 2020 Phys. Rev. Lett. 125(18) 188003 URL https://link.aps.org/doi/ 10.1103/PhysRevLett.125.188003

    2020 doi

  33. [41]

    B¨ auerle T, L¨ offler R C and Bechinger C 2020Nature Communications 11 2547 ISSN 2041-1723 number: 1 Publisher: Nature Publishing Group URL https://www.nature.com/articles/ s41467-020-16161-4

  34. [42]

    Ginelli F 2016 The European Physical Journal Special Topics 225 2099–2117 ISSN 1951-6355, 1951-6401 URL http://link.springer.com/10.1140/epjst/e2016-60066-8

    2016 doi

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.