acceptodds
Under review as a conference paper at ICLR 2027

CALIBRE: Fold-Calibrated Readouts for Frozen Pathology Encoders

Abstract

Frozen pathology encoders are evaluated on spatial gene expression prediction, but fixed readouts can obscure their value. HEST-Benchmark's ridge penalty leaves 256.0 of 256 effective degrees of freedom. We introduce CALIBRE, which adapts regularisation by selecting each encoder's penalty and projection dimension on training slides, averages seven encoders and applies an existing within-slide smoother. Calibration improves all seven pre-registered backbones on nine tasks. CALIBRE scores 0.4664 and the smoothed uncalibrated average 0.4575, versus 0.4480 from STFlow's published task scores. On 19 held-out breast cancer patients, calibration raises the primary encoder from 0.0811 to 0.1033 (p=0.0039). CALIBRE scores 0.1213, versus 0.0765 for STFlow, a flow-matching model with training-selected epochs (p=0.00064), or 0.0580 with a tuned learning rate (p=0.00042). On a 22-patient kidney cohort from another institution and platform, registered before download, calibration improves the primary encoder for every patient and CALIBRE exceeds STFlow, 0.2144 versus 0.1795 (p=0.000009). Predictors sharing encoder and smoother separate only against STFlow weakened by tuning. HER2ST's 0.2644 versus 0.2622 remains unresolved, and averaging is resolved on one of four cohorts. Training-selected frozen readouts change frozen-versus-trained comparisons.

Primary area:
datasets and benchmarks
Keywords:
pathology foundation models, frozen representations, linear readout, ridge regularization, model ensembling, spatial gene expression, benchmark protocol

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_eFuVOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper evaluates frozen pathology encoders for spatial gene-expression prediction. It shows that HEST’s fixed PCA-Ridge readout is poorly calibrated, with minimal effective regularization. The proposed CALIBRE pipeline tunes PCA dimension and Ridge penalty via cross-validation, optionally adding encoder ensembling and spatial smoothing. Across HEST and two external cohorts, calibrated readouts consistently improve frozen encoders and often match or outperform STFlow, highlighting the importance of training-selected readouts in fair model comparisons.

Critical strengths
  1. The paper identifies a concrete issue in the HEST protocol: the prescribed Ridge penalty appears nearly inactive. The effective-degrees-of-freedom analysis provides a clear explanation for why readout calibration may improve performance.

  2. The single-encoder experiments are carefully controlled: encoders, outer splits, targets, and metrics are fixed, while readout hyperparameters are selected within training folds. Results are also tested on external cohorts.

  3. The paper usefully distinguishes standardized fixed-probe evaluation from per-encoder, training-tuned readouts, a distinction that is important for interpreting frozen-encoder benchmarks.

Critical weaknesses
  1. The title and framing are somewhat misleading. “Fold-Calibrated Readouts for Frozen Pathology Encoders” suggests a broadly applicable technical design for adapting pathology encoders across diverse downstream tasks. In practice, the paper is primarily a correction to a specific evaluation choice in HEST: PCA-plus-Ridge prediction of spatial gene expression from frozen embeddings. The paper would benefit from stating this narrower contribution more explicitly.

  2. The central evidence is narrow in scope. The paper convincingly shows that fixed readout parameters can affect results under the HEST-style frozen-encoder + PCA + Ridge protocol for spatial gene-expression regression. However, it remains unclear whether this issue generalizes to other pathology tasks, benchmarks, or probe families, such as classification, survival prediction, nonlinear heads, or different pooling strategies.

  3. The role of the seven-encoder ensemble is unclear given the paper’s evaluation framing. Ensembling is reasonable for constructing a strong predictive system, but it is not an evaluation of an individual frozen encoder. Its inclusion blurs the distinction between auditing the HEST readout protocol and proposing a multi-model prediction pipeline, and it complicates comparisons against single-model trained baselines such as STFlow.

Questions for authors

See Weaknesses.

3 comments
AuthorsOpenReview ↗

Author Response to Reviewer eFuV

Thank you for the review. The paper concerns PCA-ridge readouts for spatial gene-expression regression from frozen encoders, and the revised abstract, Section 6 (Limitations), and Section 7 (Conclusion) state this narrower scope explicitly. Those sections also report that a seven-encoder STFlow ensemble leaves the kidney margin unresolved. Our contribution is diagnosing an almost inactive benchmark penalty and measuring how correcting that readout changes frozen-encoder and baseline comparisons.

The evidence for that correction. HEST's prescribed ridge penalty (100/(dg) = 0.0078 at d=256, g=50) mirrors a sample-averaged classification-probe constant applied to squared error summed over 1,293 to 67,651 training spots (Appendix L.1), retaining 256.0 of 256 effective degrees of freedom at the reported precision. Correcting that readout raises the pre-registered H-Optimus-1 nine-task mean from its published 0.4229 to 0.4465 (Table A.2) with embeddings, outer splits, targets, and metrics fixed, without averaging or smoothing. The corresponding unsmoothed single-encoder changes are 0.0811 to 0.1033 in the fold-internal ST-Net analysis (Table 2) and 0.1880 to 0.1991 on kidney (Appendix T). On reconstructed HESCAPE folds grouped by released sample index, cross-validating the inherited HEST head gains +0.0174 on UNI and +0.0151 on GigaPath (Appendix C). Probe selection on VCBench's released random split raises scGPT by +0.0668 without changing the leader or model ranking (Appendix J). Neither check establishes patient- or site-held-out transfer. These results do not extend to classification, survival prediction, nonlinear heads, or other pooling strategies, and we do not claim a general adaptation method.

The seven-encoder recipe. This recipe evaluates a combined prediction pipeline rather than an individual representation, and the revision presents it explicitly as a system baseline. Table 1 and Appendix O separate encoder count, calibration, and access to smoothing. That separation changes the conclusion. With encoder and smoother matched, the ST-Net and kidney comparisons against released-rate STFlow remain unresolved after Holm correction (Table 2 and Appendix T), separating on ST-Net only when inner learning-rate tuning weakens STFlow (Appendix S). On HER2ST, even the full recipe differs from densely training-selected STFlow by only +0.0022 over eight patients, with a 95% paired-patient t interval of [-0.0370, 0.0415] (Appendix P). The full-system results therefore do not show that calibration alone generally outperforms trained architectures. In new Appendix Y, a post-hoc kidney experiment removes the encoder mismatch by giving STFlow all seven encoders, with one model per encoder (six new encoder arms, with stored H-Optimus-1 runs reused). With encoders and smoothing matched, the full recipe leads by only +0.0019 (95% paired bootstrap interval [-0.0098, 0.0142], 12/22 patients, p=0.92), whereas the seven-model STFlow ensemble exceeds smoothed single-encoder STFlow by +0.0180 on H-Optimus-1, one of its members, and by +0.0283 on UNI, which it does not contain. On this cohort, most of the margin over single-encoder STFlow comes from averaging encoders rather than from the readout.

Technical framing. Inner cross-validation, ridge regression, averaging, and spatial smoothing are established procedures. On HEST, a strong fixed setting recovers most of the calibration gain, which further limits any claim that elaborate per-task tuning is essential. We retain the submitted title during discussion. The final version will add "for Spatial Gene-Expression Prediction" to the title to make this scope explicit.

Thanks for your response. The rebuttal clarifies that the contribution is primarily an audit and recalibration of a specific HEST-style PCA–Ridge protocol for spatial gene-expression prediction, rather than a general method for frozen pathology encoders.

While this analysis is useful, it reinforces my main concern: the scope is too narrow for ICLR, especially given the lack of evidence beyond this benchmark setting. I will therefore maintain my original score.

Thank you for following up and noting that the revised framing and component separation are clear. While our main claim is properly scoped to linear spatial-expression probing (in the 15 November revised PDF via the PDF icon beside the title, changes in blue), the paper's empirical evidence and evaluation implications extend well beyond a single HEST configuration in three concrete ways:

1. A general probe-scaling diagnosis, plus distinct tests of per-fold calibration and fixed-setting transfer.

  • Diagnosing effective-regularization collapse (Appendix L.1): HEST's prescribed ridge penalty (100/(dg) = 0.0078 at d=256, g=50) mirrors a sample-averaged classification-probe constant applied to squared error summed over 1,293 to 67,651 training spots. That spot-count scaling alone accounts for 3.1 to 4.8 of the 6.1 orders-of-magnitude penalty gap and leaves 256.0 of 256 effective degrees of freedom (virtually zero shrinkage), exposing a general linear-probing pitfall whenever training-set sizes span orders of magnitude.
  • Broad backbone impact on HEST (Section 4, Appendix L): With embeddings, splits, targets, and metrics held fixed across the benchmark slides, per-fold inner-CV calibration improves all 7/7 pre-registered pathology backbones, and a single calibrated cell applied unchanged improves all 15/15 tested backbones and 132/135 backbone-task cells (+0.0368 mean gain, positive on all nine cancer tasks).

2. Multi-cohort and multi-benchmark evidence beyond HEST.

  • External patient-held-out cohorts across institutions and platforms: On ST-Net, unsmoothed per-fold calibration lifts the primary encoder (H-Optimus-1) from 0.0811 to 0.1033 across 19 held-out breast cancer patients with training-fold gene selection (p=0.0039p=0.0039, Table 2). Crucially, in Appendix Q's cohort-wide-panel check, applying the fixed HEST constant recovers only 19% of per-fold calibration's gain (+0.0045 vs. +0.0235, modal optimal penalties 106–10810^6\text{–}10^8), showing directly why training-fold calibration is needed across cohorts rather than one universal constant. On an external 22-patient kidney cohort registered before download from a different institution, organ, and platform, per-fold calibration rises from 0.1880 to 0.1991, improving all 22/22 held-out patients (Appendix T; on 8-patient HER2ST with matched P-ES, 0.2408 to 0.2697 has p=0.11p=0.11, Appendix P).
  • Controlled probe sensitivity on HESCAPE and VCBench: On reconstructed HESCAPE five-fold sample-grouped splits, cross-validating a HEST-style head improves both tested encoders across all 5/5 folds (+0.0174 on UNI, p=0.0014p=0.0014; +0.0151 on Prov-GigaPath, p=0.0008p=0.0008, Appendix C, Table C.1). Outside histology, on VCBench's released random 80/20 split for predicting protein-derived targets from RNA embeddings, probe calibration lifts frozen scGPT from 0.0638 to 0.1306 (+0.0668) in a four-encoder comparison while preserving the top ranking (Appendix J).

3. Changing how frozen foundation representations are judged against task-trained generative models.

Published HEST-style evaluations (compounded by test-fold epoch selection in released STFlow baselines) created the impression that frozen pathology foundation models substantially lag end-to-end generative architectures. Once training-fold selection, spatial smoothing, and all seven encoders are matched in our kidney follow-up (Appendix Y, p. 52), our calibrated frozen-encoder system and a seven-encoder STFlow ensemble differ by only +0.0019 (95% CI [−0.0098, 0.0142], p=0.92p=0.92), with most of the multi-encoder margin over single-encoder STFlow driven by encoder averaging (while shared-smoother single-encoder gaps remain unresolved after Holm correction). Showing across 15 backbones, external held-out patient cohorts, and two additional benchmarks that readout miscalibration distorts foundation-model comparisons addresses a core representation-evaluation problem for ICLR, and we hope this supports reconsidering the score.

Reviewer_Q1CjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CALIBRE to improve spatial gene expression prediction using frozen pathology encoders. CALIBRE selects readout parameters through cross-validation, combines predictions from seven encoders and applies spatial smoothing. Experiments across multiple datasets show improved performance and competitive results against trained models.

Critical strengths
  1. The paper identifies an important issue in HEST-Benchmark and shows that the choice of readout can significantly affect the evaluation of pathology foundation models.

  2. The experiments are extensive, covering multiple pathology encoders, different datasets and held-out patient cohorts. The authors also provide detailed ablation studies and statistical analyses.

  3. The results show that frozen pathology encoders can achieve strong performance with properly selected readouts without the need to fine-tune the encoders. This provides useful insights for future foundation model evaluation.

Critical weaknesses
  1. In Sections 4 and 5, a single fixed readout setting improves all 15 tested encoders and 132 out of 135 encoder–task combinations. The authors also show that simply changing the ridge penalty can recover most of these gains. This makes me wonder whether tuning the readout separately for each encoder and task is really necessary, especially since a much simpler fixed setting already works well.

  2. The benefits of using seven encoders are unclear. On HER2ST, a single calibrated encoder performs better than the full ensemble (0.2697 vs. 0.2644). Given the much higher computational cost, is the seven-encoder ensemble really necessary?

  3. Minor: Readout calibration is presented as a key contribution of this paper, while combining pathology foundation models (Neidlinger et al., 2026) and spatial smoothing (Rolf et al., 2020; Sun et al., 2024) have already been explored in prior work. However, the additional gains from calibration are not statistically significant in the matched-component comparisons on HEST and HER2ST (Appendix O).

Questions for authors
  1. Could the authors further explain when CALIBRE is more beneficial than using a single fixed readout setting?

  2. Could the authors make further explanation about the cost and effectiveness trade-off? Have the authors explored using fewer encoders or selecting only the best-performing encoders for each task?

  3. Could the authors provide more evidence that calibration brings consistent additional improvements or why it is necessary beyond ensembling and smoothing?

4 comments
AuthorsOpenReview ↗

Author Response to Reviewer Q1Cj

Thank you for these questions. A fixed readout recovers most of the gain on HEST and kidney, while per-fold selection gains more on ST-Net. The single encoder has a higher HER2ST point estimate, and calibration's gain inside the ensemble depends on the cohort.

Fixed readout versus per-fold tuning. On HEST, the fixed (dimension 512, penalty 10^4) setting improves 132/135 encoder-task cells and recovers a mean 98% of the per-task calibration gain (Appendix L). Penalty-only replacement at the published dimension recovers a mean 96% of the fixed-cell gain, so separate tuning is unnecessary for most of the HEST gain. The penalty optimum spans roughly one decade (Figure L.1), and shrinkage depends on the feature spectrum and training-set size. In Appendix Q's cohort-wide-panel ST-Net analysis, modal selected penalties are 10^6–10^8 rather than 10^4, and the HEST setting gains +0.0045 versus +0.0235 for per-fold selection on H-Optimus-1, recovering only 19%. On kidney, where inner selection chooses (512, 10^4) in 131/154 encoder folds, the fixed setting recovers 96–121% across seven encoders (Appendix T).

A fixed constant still requires prior data from a compatible regime, and the HEST setting used benchmark-level information. Task-excluded selection in Appendix L retains +0.0240 of the +0.0287 all-task gain across seven encoders. We claim no advance over standard inner cross-validation.

Fewer encoders and cost. On HER2ST, the pre-specified calibrated H-Optimus-1 with P-ES scores 0.2697 versus 0.2644 for seven encoders (Appendix O). In the unsmoothed HEST prefix study (Appendix K), one, two, five, and seven calibrated encoders score 0.4465, 0.4569, 0.4667, and 0.4641 in fixed published-rank order. The five-encoder peak is exploratory.

New Appendix Z evaluates two training-side selectors specified before recovering the original inner scores. Within each outer fold, S1 selects the highest-scoring encoder and S3 averages the top three, using only inner-training scores and fixed rank-based ties. All 63 encoder-task reruns reproduce their stored predictions and available inner-score anchors exactly. On the nine unsmoothed HEST tasks at seed 1, S1 scores 0.4460 and S3 0.4600, versus 0.4465 for fixed H-Optimus-1, 0.4620 for the fixed first three, and 0.4641 for all seven (Table Z.1). S1 and S3 trail seven by 0.0181 and 0.0041, and each exceeds seven on only one task. Appendix Z lists every task and S1's fold choices. Neither selector improves the mean over the fixed predictor of the same size. These descriptive results are retrospective on previously analyzed HEST, which lacks patient identifiers, and establish neither patient-held-out transfer nor that seven encoders are always necessary. Selection evaluates all seven candidates on training data, though inference uses fewer.

The full recipe requires seven encoder passes per patch versus one for STFlow (1.8 versus 0.08 hours for 13,620 HER2ST spots, including encoding and a published-readout fit on one RTX A6000, Appendix M). Adaptation logs record 6.11 hours of summed CPU elapsed time across ten tasks, versus 2.03 hours for summed one-seed GPU STFlow medians. The adaptation timings exclude encoding and come from different shared machines, showing no end-to-end speed advantage. Smaller fixed-readout predictors remain useful accuracy-cost alternatives.

Calibration beyond averaging and smoothing. With smoothing available to both arms, calibration inside the ensemble adds +0.0088 on HEST, +0.0134 on HER2ST, and +0.0170 on ST-Net (Appendices O and Q). None of the four registered HEST/HER2ST component contrasts survives Holm correction, while ST-Net's calibration contrast does on the cohort-wide gene panel, whose selection uses all patients' expression as in the benchmark. On kidney, calibration inside the smoothed ensemble adds only +0.0012 (p=0.32), while averaging adds +0.0097 (p=0.00026, Appendix T). In new Appendix Y, giving STFlow all seven encoders and the same smoother on kidney reduces CALIBRE's lead to +0.0019 (p=0.92), while averaging seven STFlow models adds +0.0180 over smoothed single-encoder H-Optimus-1 STFlow (unadjusted p=0.0042).

Separate single-encoder checks with P-ES in both arms raise H-Optimus-1 from 0.4377 to 0.4543 on HEST (p=0.0195), from 0.0943 to 0.1120 on cohort-wide ST-Net (Holm p=0.0433), by +0.0099 across all 22 kidney patients, and from 0.2408 to 0.2697 on HER2ST, where the Wilcoxon test is unresolved (p=0.11, Appendices O, Q, and T). The HEST and HER2ST checks are descriptive additions outside the four-test family. These results do not overturn the unresolved ensemble contrasts or show synergy with smoothing.

Thank you again for your time and thoughtful review. As the discussion deadline approaches, we would be grateful if you could consider our response when your schedule permits. We would be happy to address any remaining questions.

Thank you for the additional experiments and clarifications. However, I remain concerned that its additional contribution of calibration beyond ensembling and smoothing is not consistently significant higher across datasets, making the necessity of the full CALIBRE pipeline unclear. I will therefore maintain my original rating.

Thank you for reviewing the additional experiments and for sharing your assessment. We appreciate your candor and respect your decision.

For clarity on the record, we do not view the full seven-encoder pipeline as strictly necessary in every deployment setting. Rather, separating calibration, ensembling, and spatial smoothing in the revision highlights two distinct regimes. For a single frozen encoder, where a user avoids seven-fold encoding cost, readout calibration alone consistently improves the primary encoder across HEST, ST-Net, and all 22 held-out kidney patients (and remains positive with matched spatial smoothing across those cohorts, while the 8-patient HER2ST check remains unresolved). Once seven encoders are averaged and spatially smoothed, ensembling and smoothing already absorb much of the variance of an uncalibrated linear probe, so calibration's incremental margin inside the full ensemble becomes smaller and cohort-dependent (surviving Holm correction on cohort-wide ST-Net while remaining unresolved on HEST, HER2ST, and kidney). We therefore frame CALIBRE's main contribution as diagnosing and fixing single-encoder readout miscalibration, with smaller calibrated predictors offering practical cost-accuracy trade-offs and the full seven-encoder recipe serving as a transparent system baseline alongside matched STFlow ensembles (Appendix Y). Thank you again for your constructive feedback throughout the review.

Reviewer_an9cOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

It shows that HEST's fixed PCA-256 plus ridge readout applies almost no shrinkage. They propose CALIBRE that selects each encoder’s PCA dimension and ridge penalty by inner cross-validation on training slides. It then averages seven calibrated encoders with equal weight and applies an existing within-slide smoother. This calibration improves all seven pre-registered encoders on HEST. On two held-out cohorts CALIBRE beats STFlow. Comparisons with a shared encoder and the same smoother are mostly unresolved after Holm correction. HER2ST is unresolved.

Critical strengths

It corrects the widely used HEST benchmark. HEST’s readout barely regularises, and STFlow’s published numbers rely on choosing epochs on the test fold. Both findings affect how the field compares frozen and trained predictors. This is the paper’s main value. The evaluation is rigorous. Pre-registration, a cohort registered before its data were downloaded, matched smoother access for both arms, and destructive controls make the results credible .

Critical weaknesses

Novelty is limited and the gains are modest. The ingredients are standard. Most of the lead over STFlow comes from averaging seven encoders and smoothing & not from calibration. Calibration inside the ensemble is not established on the kidney cohort, and 8 of 9 registered extensions failed. The comparison set and scope are narrow. STFlow is the only trained predictor. Several other trained predictors exist (TRIPLEX, BLEEP, Stem, mclSTExp, etc.) . The seven-encoder ensemble costs a lot more encoding compute than the trained model. Its main claim that frozen encoders should be compared through readouts selected on training data, is close to accepted practice in vision probing.

Questions for authors

Why is STFlow the only trained comparator? Do published results allow a comparison with other trained predictors under patient-held-out splits? Have you confirmed the test-fold epoch selection with STFlow’s authors? Does it also affect the numbers they report for other baselines? This determines how much of the “frozen beats trained” story is really about the comparator’s protocol. How much does calibration alone, without averaging or smoothing, close the gap to STFlow at matched input on the held-out cohorts?

1 comment
AuthorsOpenReview ↗

Author Response to Reviewer an9c

Thank you for the review. Calibration improves the frozen baseline at matched input on the held-out cohorts. The full recipe also averages encoders and smooths predictions, and superiority at matched components remains unresolved.

Calibration alone at matched input. Without averaging or smoothing, calibrating UNI, STFlow's input encoder, raises fold-internal ST-Net from 0.0819 to 0.0982 versus training-selected STFlow's 0.0765, and kidney from 0.1833 to 0.1920 versus STFlow's 0.1795 (Table 2, Appendices R and T). The unsmoothed ST-Net difference of +0.0218 is unresolved (p=0.0955), kidney's +0.0125 has an unadjusted secondary p=0.0190, and HER2ST's unsmoothed calibrated UNI differs from densely training-selected STFlow by -0.0007 (p=0.95, Appendix R). With encoder and smoother matched against released-rate STFlow, neither ST-Net (+0.0250 on UNI, +0.0221 on H-Optimus-1, both Holm p=0.19) nor kidney (+0.0124 on UNI, +0.0102 on H-Optimus-1, both Holm p=0.0658) survives Holm correction. On HEST, unsmoothed UNI rises from 0.3857 to 0.4171, exceeding the fixed-schedule STFlow rerun (0.3837) while remaining below the test-epoch-selected rerun (0.4430, Table 1 and Appendix R).

Comparator coverage. We reran STFlow because it reports HEST results and takes frozen UNI features, allowing controlled comparisons on identical inputs (Appendix X.5). Table 1 also lists the published TRIPLEX aggregate, which cannot supply patient-paired comparisons on our held-out ST-Net and kidney splits. We have not rerun TRIPLEX, BLEEP, Stem, or mclSTExp under our protocols, nor verified whether their published evaluations use compatible patient-disjoint splits and fold-internal gene panels, so we make no numerical comparison with them or claim superiority over trained spatial predictors generally.

Test-fold epoch selection. This finding is based on the released evaluation procedure and controlled reruns (Appendix M). We have not received confirmation from STFlow's authors, and we do not know whether iterative baseline numbers reported in their paper are affected (their frozen UNI row uses a 70-tree random forest without epochs, Appendix X.5). Our external comparisons therefore select STFlow epochs on training patients. On HER2ST, dense selection from 56 complete inner-patient curves, fixed before outer scoring, gives STFlow 0.2622 versus CALIBRE's 0.2644 (Appendix P). The +0.0022 difference is positive for 5/8 patients, with a 95% paired-patient t interval of [-0.0370, 0.0415]. Designed after earlier HER2ST outcomes, this diagnostic is not a blind external confirmation and leaves the advantage unresolved.

On the same fold-internal gene panels in ST-Net and kidney, training-selected STFlow scores 0.0765 and 0.1795 (0.0825 and 0.1843 with P-ES) versus the full recipe's 0.1213 and 0.2144 (Table 2 and Appendix T). Those wider margins combine calibration, averaging, and smoothing. In new Appendix Y, a post-hoc kidney experiment tests encoder count directly by giving STFlow all seven encoders, with one model per encoder (six new encoder arms, reusing stored H-Optimus-1 runs) and epochs selected on inner patients only. With all seven encoders available to both, CALIBRE leads by +0.0019 with P-ES (95% paired bootstrap interval [-0.0098, 0.0142], 12/22 patients, two-sided Wilcoxon p=0.92) and by +0.0023 without smoothing (p=0.82). The seven-model STFlow ensemble exceeds smoothed single-encoder STFlow by +0.0180 on H-Optimus-1, one of its members, and by +0.0283 on UNI, which it does not contain (unadjusted p=0.0042 and p=0.00069). On kidney, the margin over single-encoder STFlow therefore comes mostly from averaging encoders, and whether CALIBRE outperforms a trained predictor given the same seven encoders is unresolved. This analysis covers one cohort and one architecture, where the STFlow ensemble trains seven whole-slide models while CALIBRE fits seven ridge readouts.

Contribution and cost. Selecting linear-probe penalties on training data is standard in vision probing (Section 2, Appendix X.2), the pipeline ingredients are established, the kidney ensemble-calibration increment is +0.0012 (p=0.32), and eight of nine registered extensions fail their criteria (Appendix W). The seven-encoder system requires seven forward passes per patch versus one (1.8 versus 0.08 hours for 13,620 HER2ST spots, including encoding and a published-readout fit, Appendix M), and adaptation timings across different CPU and GPU servers exclude encoding and show no end-to-end efficiency advantage. Our contribution is diagnosing negligible effective shrinkage in a widely adopted spatial-expression benchmark and measuring how readout selection changes frozen-encoder and frozen-versus-trained comparisons, with unresolved contrasts and failed extensions retained.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Brief Post-Revision Discussion Summary for the Area Chair

Thank you for overseeing the review of our submission. We briefly summarize the two main clarifications added during the discussion following our 15 November revision:

  1. Evidence beyond a single HEST protocol: In response to Reviewer eFuV, we summarized how HEST's spot-count scaling leaves 256.0 of 256 effective degrees of freedom (Appendix L.1) and consolidated our cross-cohort and cross-benchmark checks: 19 held-out breast cancer patients on ST-Net (Table 2; where transferring HEST's fixed constant recovers only 19% of per-fold calibration's gain in Appendix Q), an external 22-patient kidney cohort registered before download (all 22/22 patients improved, Appendix T; 8-patient HER2ST remains unresolved at p=0.11), 5/5 reconstructed sample-grouped folds on HESCAPE (Appendix C), and RNA-to-protein VCBench (Appendix J).

  2. Single-encoder calibration vs. the seven-encoder system baseline: In response to Reviewer Q1Cj, we clarified that we do not claim the full seven-encoder recipe is required in every setting. For a single frozen encoder (avoiding seven-fold encoding cost), readout calibration alone consistently improves the primary encoder across HEST, ST-Net, and all 22 kidney patients; once seven encoders are averaged and spatially smoothed, ensembling and smoothing already absorb much of the linear-probe variance, making calibration's marginal ensemble gain smaller and cohort-dependent. Matched seven-encoder STFlow on kidney differs from our calibrated frozen-encoder baseline by +0.0019 (p=0.92, Appendix Y).

We remain available if any further clarification would be helpful.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.