acceptodds
Under review as a conference paper at ICLR 2027

Reading the Velocity Field: Closed-Form Analysis of a Cross-Modal Flow Adapter

Abstract

Flow-matching adapters specialise a vision–language model from a few labelled images: they regress a velocity field carrying image features toward the text embedding of their class, then integrate it at test time. We show the population optimum of this objective is available in closed form. Inverting the interpolant removes the image feature from the regression target, and in the noise-free setting these methods use, what remains is geometry: the field's value is a convex combination of the text anchors, the Euler iterate has an exact form, and the classifier logits split into a zero-shot term and a posterior smoothed by the anchor Gram matrix . That split is useful. Reading the posterior off a trained field and integrating the field are the same estimator applying and , so neither dominates and the better choice depends on how well conditioned is. Interpolating them recovers both as limits, and the weight is not free: since the residual covariance in anchor coordinates is , the forward map itself, the interpolating estimator is Bayes-optimal with . This yields a decoder with no tuned hyperparameters needing one forward pass and no ODE solve. On the five difficult benchmarks of jiang2026exploring it modestly but consistently improves on integration with almost tuning free. A matched-form control attributes the gain to the field rather than the estimator: applied to raw features the same decoder provides at par performance over zero-shot. Our results highlight that principled closed-form analysis can simplify inference while modestly improving few-shot adaptation.

Primary area:
learning theory
Keywords:
Reading the Velocity Field, Closed-Form Analysis, Cross-Modal Flow Adapter

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ubpROpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper is based on a new few-shot learning paradigm, FMA, and proposes a decoder-based inference method to solve FMA's inference hyperparameter-dependent problem. With the same training velocity, the decoder-based method can achieve stable improvement across datasets and shot settings. Finally, this manuscript also analyzes the limitations of this method, i.e., the decoder recovers information that weaker upstream adapters leave unused.

Critical strengths
  1. This paper aims at a key challenge in flow-based few-shot learning methods, that is, the inference/transport between the text and visual modalities. The proposed method is clean and useful, showing stable improvement.
  2. The experimental results are rigorous, especially by providing higher reproducibility results for FMA.
  3. The whole analysis is systematic and reuses the FMA's trained velocity field to ablate the decoder-based inference contribution. The FMA proposes a new flow-based few-shot learning method; the training is reasonable and uses a heuristic inference setting. However, in this manuscript, the inference becomes a closed-form stable transport, making this paper and FMA contribute to a new paradigm together.
  4. The limitation and contribution claim are rigorous and reasonable.
  5. The ablation and theoretical justification are sufficient to support the main claim.
Critical weaknesses

Lacks discussion and citation of the related (following FMA) flow-matching-based adaptation works.

Questions for authors
  1. Can this inference method be extended to other interpolation fields, such as in spherical or hyperbolic space?
  2. Can this method be general enough for other transport tasks, such as retrieval or even visual generation tasks?
  3. Why can this kind of flow-based adaptation outperform one-step adaptation methods like CoOp or CLIP-LoRA?
4 comments

Thank you for the positive review.

Related flow-matching work. Section 2 now discusses the follow-ups to FMA: hyperbolic and radial-angular flows for few-shot adaptation (Li et al., Chen et al.), and FlowCIR, FlowComposer and GeneFlow for other tasks. They change how or where the field is trained; our analysis is about how a trained field should be read, so it applies to them as well.

Q1 (spherical or hyperbolic). Partly, and the analysis says which part. Our closed form rests on one step: given the class anchor and the current state, the source feature is determined, so it drops out of the regression target. This still holds on the sphere, because the state lies on the great circle through the image feature and its anchor, and in hyperbolic space, where it lies on the corresponding geodesic. So the optimal field is still a posterior-weighted mixture of one velocity per anchor, and the network still only has to estimate class probabilities.

What changes is the velocity toward each anchor. In the linear case it is linear in the anchor, which is what makes integration equal to "zero-shot logits plus posterior times the Gram matrix". On the sphere or in hyperbolic space it follows the logarithmic map, which is not linear in the anchor, so this decomposition, and with it the exact form of our decoder, holds only to first order in the angle between the state and the anchors. CLIP features occupy a narrow cone, so we expect the approximation to be good, but we have not tested it. This is relevant in practice because two follow-ups to FMA already train in these geometries with fixed class anchors (Li et al., Chen et al.), so their fields should have the same posterior structure. Extending the decoder to them is the natural next step, and we mention it in the conclusion.

Q2 (retrieval or generation). Retrieval against a fixed gallery fits our setting, with gallery items as anchors. Generation does not, because the targets are not a finite set.

Q3 (vs. CoOp or CLIP-LoRA). FMA is stacked on these methods rather than replacing them. It adds a learned class posterior on top of their zero-shot logits, which is why the gain shrinks when the upstream method is already strong (CoOp, CLIP-LoRA at 16 shots) and is largest at 1 shot (Table 3).

I still have one question: why this kind of flow matching based adaptation can further improve CLIP-based adaptation even with the same training data.

Thank you for the follow-up. Methods such as CoOp or CLIP-LoRA reduce the gap between image features and class text embeddings, but they do not close it: on the same training data some misalignment between the two modalities remains.

The flow head is trained on exactly this remaining gap, moving each image feature toward its class text embedding, so it adds to the upstream method rather than repeating it.

This is the motivation of the original FMA, and our analysis is consistent with it: the class scores after integration are the upstream method's own logits plus a learned correction toward the class embeddings (Corollary 4). Accordingly the gain is largest when more of the gap is left (CLIP-Adapter 62.46 to 70.51, Table 8) and smallest when the upstream method has already aligned the modalities well (CoOp, and CLIP-LoRA at 16 shots).

OK, I decided to maintain my score.

Reviewer_j4pNOpenReview ↗
Rating
3
Confidence
1
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies flow-matching adapters with fixed text anchors. It analyzes the geometry and logit decomposition of the ideal noise-free field, interprets direct coefficient read-out and anchor scoring through the Gram matrix, and proposes a regularized single-pass decoder. On five difficult datasets, the reported average accuracy improves from 69.27% to 71.52%. The paper also explores parameter estimation without validation grid search. The mechanistic question is valuable, but several connections between the idealized analysis and the learned networks remain insufficiently established.

Critical strengths
  1. The paper asks a valuable question about whether flow adaptation benefits from the learned mapping or from repeated integration. Replacing the decoder on the same trained field helps isolate the effect of inference. The single-pass method improves over integration on all five datasets in Table 1 while requiring fewer field evaluations, establishing practical relevance.

  2. The experimental controls make the reported improvements more credible. The paper distinguishes published and reproduced results and includes raw-feature, shuffled-anchor, prototype-substitution, and cross-dataset selection controls.

Critical weaknesses

Approximately straight trajectories do not yet establish equivalence between one initial field evaluation and validation-selected multi-step integration. An exactly constant-velocity straight flow admits a one-step solution [1], but directional consistency or a small average straightness measure does not automatically control the velocity error at t=0. Early-stopped integration is also different from endpoint evaluation. Please report output discrepancies and prediction disagreement, connect them to the measured straightness, and compare �0��0�, integration, and the proposed decoder directly. The static anchor-regression result of 72.13% versus 71.25% for full-time training also warrants confirmation under the same training pipeline.

References [1] Liu, X., Gong, C., & Liu, Q. Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow. ICLR 2023.

Questions for authors

Which theoretical conclusions remain exact when network outputs leave the anchor affine hull and coefficients are not simplex-valued?

2 comments

Thank you for the careful review.

One step vs. integration. We agree that straightness alone does not show equivalence, and the paper no longer argues that it does. Section 6 presents straightness only as an explanation, and the decoder is compared with integration directly in every table. Appendix C adds a bound on the gap between one step and integration, and Table 4 reports the measurements you asked for on the 15 group-A fields:

  • One Euler step of the integration length changes the prediction on 5.6% of test images (2.3% to 5.0% on four datasets, 14.3% on Aircraft) and is 0.35 points more accurate on average.
  • The measured discrepancy (0.058) is about half the bound (0.102) and below it on every test image.
  • The discrepancy is not ordered by straightness across datasets (Spearman −0.41), because it scales with the step count as well, so we do not claim that straighter fields give smaller one-step error.
  • Our decoder disagrees with integration on 19.5% of images and is 2.1 points more accurate, so it is a different read-out rather than an approximation of integration.

Anchor regression (72.13 vs. 71.25). Both are group-I runs with the same training loop, which Appendix C now states.

Q1 (outputs off the anchor hull). See the new Remark 5: the flow maps, the logit decomposition and the two readings hold for any network; hull membership and the probability interpretation need the optimal field; Bayes optimality needs the Gaussian model.

For convenience, here are the new results and statements in full, so they can be read without the PDF.

1. One step vs. integration: what straightness controls

Write vmv_m for the velocity at Euler step mm and vˉ=1M∑mvm\bar v=\frac1M\sum_m v_m. After MM steps, zM=x+tMvˉz_M=x+t_M\bar v; one step of the same length gives z^M=x+tMv0\hat z_M=x+t_M v_0. With the discrete straightness Sh=1M∑m∥vm−vˉ∥2/∥vˉ∥2\mathcal S_h=\frac1M\sum_m\|v_m-\bar v\|^2/\|\bar v\|^2, and since ∥v0−vˉ∥2≤∑m∥vm−vˉ∥2\|v_0-\bar v\|^2\le\sum_m\|v_m-\bar v\|^2,

∥zM−z^M∥∥zM−x∥=∥v0−vˉ∥∥vˉ∥≤M Sh.\frac{\|z_M-\hat z_M\|}{\|z_M-x\|}=\frac{\|v_0-\bar v\|}{\|\bar v\|}\le\sqrt{M\,\mathcal S_h}.

So straightness controls the one-step error only up to a factor M\sqrt M, as you pointed out. The paper no longer uses straightness to justify the decoder; it is compared with integration directly in every table.

2. Measurements on the trained fields

Group A, 16 shots, CLIP ViT-B/16, three seeds per dataset (15 fields). INT: integration with the validation-selected MM; ONE: one step of the same length; DEC: our decoder. Accuracy and disagreement with INT in %.

datasetMMacc INTacc ONEacc DECdis. ONEdis. DECdiscrepancyboundS\mathcal S
Aircraft6.740.1340.7743.9714.2544.110.0780.1430.0052
EuroSAT6.383.9284.4285.923.468.740.0570.1110.0025
DTD6.769.9870.4171.754.9820.760.0560.1070.0051
SUN3973.773.6173.6874.852.3211.310.0410.0610.0123
Cars4.379.4479.5881.123.1012.560.0560.0860.0087
mean5.569.4269.7771.525.6219.500.0580.1020.0068
  • One step changes the prediction on 5.6% of test images and is never less accurate than integration (+0.35 on average).
  • The measured discrepancy is about half the bound, and below it on every test image.
  • Across the 15 fields the discrepancy is not ordered by S\mathcal S (Pearson −0.28, Spearman −0.41), because the bound also scales with MM: SUN397 has the largest S\mathcal S but the smallest MM and the smallest discrepancy. We therefore do not claim that straighter fields give smaller one-step error.
  • The decoder disagrees with integration on 19.5% of images and is 2.1 points more accurate: it is a different read-out, not an approximation of integration.

3. Static anchor regression (72.13 vs. 71.25)

Both are group-I runs, trained with the same loop, seeds and decoder; only the time sampling differs (t≡0t\equiv0 vs. t∼U[0,1]t\sim\mathcal U[0,1]). The t≡0t\equiv0 network has straightness 0.116 against 0.007 for the flow-trained one, so the flow objective, not the architecture, makes the fields straight.

4. What stays exact off the anchor hull (Remark 5)

Write any network output as c^=C⊤q+ξ\hat c=\mathbf C^\top q+\xi, with q=G−1Cc^q=\mathbf G^{-1}\mathbf C\hat c the least-squares anchor coefficients (any real vector) and ξ\xi orthogonal to all anchors.

  • Exact for any network: the continuous and Euler flow maps, zm=(1−tm)x+tmamz_m=(1-t_m)x+t_m a_m; the logit decomposition ⟨zm,ck⟩=(1−tm)⟨x,ck⟩+tm(Gqˉm)k\langle z_m,c_k\rangle=(1-t_m)\langle x,c_k\rangle+t_m(\mathbf G\bar q_m)_k, since ξ\xi is orthogonal to every anchor and drops out; the two readings G−1Cc^=q\mathbf G^{-1}\mathbf C\hat c=q and Cc^=Gq\mathbf C\hat c=\mathbf G q; and the limits of the ridge decoder.
  • Need the optimal noise-free field: membership of c^\hat c and ama_m in the convex hull, qq being a probability vector, convexity of each Euler step, and rank confinement.
  • Need a model assumption: Bayes optimality of the ridge decoder assumes Gaussian residuals.

Empirically the trained fields respect the affine structure but not the convex one (negative coefficient mass grows with KK), which is the first case.

Reviewer_yMSgOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper works on the few-shot image classification task. It derives a closed-form analysis of cross-modal flow matching and introduces a hyperparameter-free, single-pass decoder that improves few-shot image classification accuracy while eliminating iterative ODE integration. The authors conduct extensive experiments to verify their idea.

Critical strengths
  • The paper’s strongest contribution is its empirical and analytical investigation of how existing flow-matching adapters behave. The proposed decoder is a reasonable practical outcome of this investigation.

  • The experiments cover multiple backbones, baselines, and ablations.

Critical weaknesses

Unclear Research Motivation, Poor Organization, and Insufficient Related Work

Overall the quality of this paper is relatively low. The writing style and organization of this paper looks very AI generated. For example, it’s really hard to tell what is the problem that is studied from the abstract.

  • The introduction provides background on flow-matching adapters and presents several theoretical observations about their learned velocity fields. However, it does not clearly establish the underlying research problem or identify a concrete limitation of existing approaches that motivates the proposed analysis. The paper primarily argues that the population-optimal velocity field admits a closed-form characterization, but it remains unclear why obtaining this characterization is important for few-shot adaptation or what fundamental limitation it addresses.

  • Additionally, the related work is very shallow. The authors only present the “big” work instead of a deep analysis of the related literature. It's very hard to comprehend what literature exists for this problem and how this work differs from that. There are only 13 references in the whole paper, despite 60 pages of text.

Inadequate Description and Justification of Experimental Baselines

  • Although the paper reports comparisons with several existing few-shot adaptation methods, the experimental section provides limited explanation of the selected baselines and their relevance to the proposed approach. Most methods are introduced through references and numerical results, without sufficiently discussing their underlying differences or why they constitute appropriate comparisons.

  • Furthermore, the distinction between standalone few-shot adaptation methods, alternative decoding strategies, and methods used as upstream adapters is not clearly established. A more structured explanation would improve the clarity of the experimental evaluation and help readers better understand the significance of the reported improvements.

Inconsistencies in Experimental Implementation and Reporting

  • Inconsistent Training Procedures: Appendix D states that Table 2 uses the released training procedure, whereas Appendix H states that Group C, corresponding to the CLIP-LoRA experiments in Table 11, uses cached-feature training. However, Tables 2 and 11 report identical results for the overlapping experiments, up to the reported precision. The authors should clarify which training procedure was used to generate these results and resolve the apparent inconsistency between the two sections.

  • Incomplete Dataset Coverage: Table 7 states that the results cover five difficult datasets. However, EuroSAT is missing from the 1-shot CLIP setting, meaning that the reported average is computed over only four datasets. The authors should explain this omission and clarify the evaluation protocol, particularly since the other shot settings include all five datasets.

Minor Weaknesses

  • The abstract contains typos and garbled phrasing (“modelstly,” “with almost tuning free,” “provides at par performance over zero-shot”)

  • Misleading result highlighting. In Table #2, the authors bold the entire row with the highest average accuracy, although their method is not the best on every dataset. In the 16-shot setting, it outperforms reproduced FMA on only 2 of 5 datasets, with a marginal average improvement of 0.1 percentage points. I suggest highlighting the best result in each column instead.

  • Is it standard practice to include more than 20 pages of code in the appendix? What is the purpose of this?

Questions for authors

See Weaknesses

4 comments

Thank you for the detailed review. We took each point seriously and revised the paper accordingly. Below we go through your three scores (presentation, informativeness, soundness) and then the specific weaknesses. All numbers refer to the revised PDF.

Presentation: the problem is now stated first

You wrote that it was hard to tell from the abstract what problem is studied. We agree, and the paper is now organised around it.

The problem. Flow-matching adapters such as FMA classify by following a learned velocity field for a number of Euler steps that is tuned per dataset, because accuracy rises and then falls along the path. FMA's own analysis predicts that integration lands exactly on the correct class embedding, which would mean perfect accuracy (Appendix F). So three basic questions were open: what function the network learns, what integration computes and why stopping early helps, and whether integration is the right way to use the trained field at all. Without the first answer, one cannot tell whether FMA's gains come from the learned map or from the integration around it.

What changed in the paper.

  • The abstract is rewritten in plain language, without mathematics, and states this problem in its first sentences.
  • The introduction poses the three questions explicitly, and Sections 4 to 7 answer them in order.
  • Figure 1 (new, page 2) shows FMA's inference, the closed-form field and our decoder side by side; Figure 2 (new, page 6) shows why neither existing way of reading the field is enough.
  • Secondary derivations moved to Appendices A and H, so the main text follows one argument.
  • The 20 pages of code listings are removed (they remain in the supplementary archive); the PDF is 27 pages instead of 57.
  • The related work is rewritten (55 references instead of 13; details in part 2).

Informativeness:

You wrote that you learned little from the paper. We think the submitted version hid its findings behind the derivations, so here they are listed diretly:

  1. Integration has a closed form, for any trained network. After M Euler steps, the class scores are exactly the zero-shot logits plus the class posterior blurred by the similarity between class names (Corollary 4, Remark 5). The tuned step count only sets the weight between these two terms. This explains why it must be tuned and why accuracy peaks before the end of the path, which FMA's analysis could not.
  2. Neither existing reading is right in general, and we can predict which is better. Reading the posterior directly removes the blur but amplifies noise. Which one wins is predicted by how similar the class names are: across the five datasets the advantage of direct reading falls with the conditioning of the class-name Gram matrix (r = −0.89, Figure 2c). Direct reading wins on EuroSAT (+3.0) and loses on Aircraft (−25.6).
  3. The optimal reading needs no tuning. The Bayes estimator between the two readings is a ridge solve, and its regulariser is not a free parameter: it equals a noise-to-prior ratio that the field's own outputs estimate (Theorem 6). With nothing tuned it reaches 71.46, within 0.06 of the validation grid (Table 2).

Why this matters in practice. The result replaces a tuned multi-step ODE solve with one network evaluation and a small linear solve, and it improves accuracy on the same trained fields: 69.27 to 71.52 at 16 shots, better on all five datasets, and the best entry on every dataset at 1 shot on CLIP-LoRA features (+2.2 over the reported FMA, Table 3).

We believe that major concern was on the presentation rather than the actual method, and revision can change the pre-rebuttal assessment. We follow concerns of soundness in following extended comment.

AuthorsOpenReview ↗

Soundness: the inconsistencies are resolved, and the claims are checked

We understand the soundness score came mainly from the reporting inconsistencies you found. Both were real and both are fixed; we also list what supports the claims.

Tables 2 and 11 report identical numbers. They are the same runs: Table 3 (formerly 2) is the dataset average of the cells in Table 13 (formerly 11). The appendix text that described them as different training procedures was wrong. Appendix E now lists the training loop of every result group. In doing so we also found and now state that the 16-shot headline setting pools three released-loop seeds with one cached-loop replicate; using the released-loop seeds only, integration averages 69.42 and the decoder 71.52, so no conclusion changes.

EuroSAT missing at 1 shot. These runs had failed because of a batching bug: with 10 support images, fewer than the batch size, no training step was taken. We fixed it (Appendix E) and ran them, and Tables 8 and 17 now include all five datasets (1-shot CLIP average: 51.91 for integration, 54.83 for the decoder). EuroSAT has only 10 validation images at 1 shot, so the selected numbers there are unreliable; at the test oracle the two readings differ by only +0.8, and we say so in the table note rather than claim the larger gap.

What supports the claims.

  • Every formal statement is checked numerically by scripts in the supplement, to between 10^-15 and 10^-17 for the identities (Table 5).
  • The gain comes from the field, not the decoder. The identical decoder on raw CLIP features scores 49.13 against 49.11 for zero-shot, and against randomly permuted anchors it falls to chance (Table 2).
  • No tuning on test. Every hyperparameter, including the baseline's step count (which the released FMA code selects on the test set), is chosen on the validation split.
  • Statistics. The decoder is better in 20 of 20 dataset and seed cells (Wilcoxon p < 0.001). With its two parameters chosen on the other four datasets, the gain is +1.41 (95% CI [+0.23, +2.74]).
  • A conservative baseline. Our FMA reproduction matches the published numbers within one reported standard deviation on most datasets and is higher on the others.
  • New measurements. One step versus full integration on the trained fields (Appendix C, Table 4).

Baselines (W2)

The new "Compared methods" paragraph (Section 8) separates three groups and says why each is there:

  1. Decoders of the same trained field (integration, direct read-out, ours): the main comparison, since network, features and training are identical.
  2. Training-free heads (prototypes, ridge, plug-in Gaussian): they test whether the gain needs a learned field.
  3. Upstream adapters (CoOp, CLIP-Adapter, CLIP-LoRA) used as feature extractors, as in FMA, and the published methods of Table 3 for context.

Table 1 is organised by these groups.

Related work (W1)

Section 2 now covers prompt learning (CoOp, CoCoOp, ProGrad, KgCoOp, PromptSRC, PLOT, MaPLe), adapters and caches (CLIP-Adapter, TaskRes, Tip-Adapter, APE, SuS-X), CLIP-LoRA, closed-form and linear-probe baselines (GDA, LP++, CLAP, cross-modal adaptation), the follow-ups to FMA (hyperbolic and radial-angular flows, FlowCIR, FlowComposer, GeneFlow), flow matching theory, and regularised inversion. It also states how we differ: these methods propose ways to fit an adapter, while we characterise what the transport adapter has learned and how to read it.

Minor points

  • Typos in the abstract. Fixed; the abstract is rewritten.
  • Highlighting. Bold now marks the best entry per column or row. Your reading of the 16-shot block is correct: over our own reproduction the decoder wins 2 of 5 datasets, by +0.1; Section 9 lists this as a limitation. Table 3 now shows the decoder's margin explicitly at every shot count, and every dataset at 1 shot, where the gain is largest.
  • Code in the appendix. Removed; Appendix J lists the scripts.

We believe these changes address the presentation, informativeness and soundness concerns, and we would be grateful if you would reconsider your scores in light of the revision. We are happy to answer any further question during the discussion.

Reviewer_yMSgOpenReview ↗

Response

(1) My main concern was about the presentation and writing of the paper, which was clearly rushed, full of AI jargon, without any coherence or structure that you would expect in an ICLR or top-level conference submission. It was hard to focus on the technical details with such poor readability.

(2) A simple side-by-side comparison of your current version with the original submission shows that almost all the sections have gone through major changes, including the abstract, introduction, and methodology. While the new version has improved readability, it is again clear evidence of how rushed the original submission was. It's also not fair to submit such a poor-quality version and hope to fix everything in the rebuttal. The current version requires a complete end-to-end review, which is not typical during the rebuttal process.

(3) I am maintaining my original score. I leave the rest of the decision to the ACs and SACs.

Thank you for reading the revision and for the follow-up.

  • We accept that the original submission fell short in presentation, and we appreciate the clear feedback.
  • The revision was made in direct response to your review, to improve readability. It does not change the paper's content.
  • The theorems, the decoder, the experimental protocol and all reported results are the same as in the submission. The only additions are those requested by reviewers (Table 4, the EuroSAT 1-shot runs, Remark 5). No conclusion has changed.
  • The analysis, method and experiments are entirely our own work; the revision changed how they are presented, not what they are.

We respect your decision and thank you again for your time and comments.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response

We thank all three reviewers. We have uploaded a revised PDF; the theorems and the main results are unchanged.

Key results.

  • On the same trained fields, our one-pass decoder improves on FMA's integration on all five difficult datasets (69.27 to 71.52 at 16 shots), and by +1.41 when its two parameters are chosen on the other four datasets (95% CI [+0.23, +2.74]).
  • The gain comes from the learned field, not from the decoder: applied to raw CLIP features, the same decoder scores 49.13 against 49.11 for zero-shot (Table 2).
  • The gain is largest when data is scarce: on CLIP-LoRA features at 1 shot it is the best entry on every dataset, +2.2 on average over the reported FMA (Table 3).
  • With no tuned parameter the decoder reaches 71.46, within 0.06 of the validation grid (Table 2).

Main changes.

  • Motivation first. The introduction now opens with the problem (FMA's step count is tuned per dataset, and its analysis predicts perfect accuracy) and three questions that Sections 4 to 7 answer in turn.
  • Figures. New Figure 1 (overview, page 2) and Figure 2 (the two readings of a trained field, page 6).
  • Experiments. A "Compared methods" paragraph in Section 8; Table 1 grouped by role; bold marks the best entry per row or column; Table 3 shows every dataset at 1 shot plus rows with the decoder's margin.
  • Related work. Section 2 rewritten, including the follow-ups to FMA (55 references instead of 13).
  • New experiments. One step vs. integration on the trained fields (Appendix C, Table 4), and the missing EuroSAT 1-shot runs (Tables 8 and 17).
  • Reporting. Appendix E now lists the training loop of each group and how seeds and splits are used. The abstract is rewritten in plain language and the code listings are removed (27 pages instead of 57).
AuthorsOpenReview ↗

Revised supplementary material

We have uploaded an updated supplementary archive alongside the revised PDF. It adds the scripts behind the new results, namely the one-step vs. integration comparison (Appendix C, Table 4) and a script that lists the training loop of every run (Appendix E), together with their output reports. It also includes the fix for the batching issue that affected the EuroSAT 1-shot runs. Every number in the revised tables can now be reproduced from the released code. We hope this makes it easy to check the revision. We would be glad to answer any remaining questions while the discussion period is open.

AuthorsOpenReview ↗

Summary for the Area Chair

Dear Area Chair,

Thank you for handling our paper. Our aim is simple: FMA classifies by integrating a learned velocity field, and we wanted to understand what that field has actually learned and how it is best read. The reviewers saw value in this. j4pN found the question of whether the gain comes from the learned map or from the integration a valuable one, and ubpR found the method clean and the experiments rigorous. The main criticism, as yMSg put it in the follow-up, was the presentation, and we agree it needed work.

Below we summarise each concern and our response. INT means FMA integration and DEC our one-pass decoder. Table numbers refer to the revised PDF.

Concern (reviewer)What we didResult
Problem not clear from the abstract (yMSg)Rewrote the abstract and introduction, added Figs. 1 and 2The paper now starts with the problem: FMA's step count must be tuned per dataset, and FMA's analysis predicts perfect accuracy (App. F)
Related work too short, 13 references (yMSg)Rewrote Section 255 references now, and we explain how our work differs
Baselines not explained (yMSg)Added a "Compared methods" paragraph and grouped Table 1Three groups: other ways to read the same trained field, training-free heads, and upstream adapters
Tables 2 and 11 show the same numbers but different training (yMSg)Checked the training loop of every runSame runs. The appendix text was wrong, and we corrected it
EuroSAT missing at 1 shot (yMSg)Fixed a batching bug and ran 3 seedsAverage over five datasets: INT 51.91, DEC 54.83. EuroSAT has only 10 validation images at 1 shot; at the test oracle the gap is 0.8, which we note in the table
Typos in the abstract (yMSg)Rewrote itFixed
Bold used for whole rows (yMSg)Bold now marks the best entry per column or rowOn CLIP-LoRA at 16 shots, DEC beats our FMA reproduction on 2 of 5 datasets, by 0.1 on average. We list this as a limitation
Code in the appendix (yMSg)Moved it to the supplementaryThe PDF went from 57 to 27 pages
Revision too large (yMSg, follow-up)Kept the theorems, method and protocol the sameAll reported numbers are the same, except the 1-shot average, which now includes EuroSAT. New results were added only where reviewers asked (Table 4, Remark 5, EuroSAT 1-shot)
Straight paths do not prove one step equals integration (j4pN)Added a bound and measured one step, integration and DEC on 15 trained fields (Table 4)One step changes 5.6% of predictions, and the error (0.058) is about half the bound (0.102). Straighter fields did not give smaller errors (Spearman −0.41), so we dropped that claim. DEC differs from INT on 19.5% of images and is 2.1 points better, so we compare the two directly
Anchor regression, 72.13 vs. 71.25 (j4pN)Checked both runsSame training loop, seeds and decoder; only how t is sampled differs
What still holds off the anchor hull (j4pN)Added Remark 5The flow maps and the logit decomposition hold for any network. Hull membership needs the optimal field, and Bayes optimality needs the Gaussian model
FMA follow-ups missing (ubpR)Added them to Section 2They change how the field is trained; our analysis is about how it is read, so it applies to them too
Spherical or hyperbolic fields (ubpR)Replied in the threadThe decomposition holds approximately there; we have not tested it
Retrieval or generation (ubpR)Replied in the threadRetrieval with a fixed gallery fits; generation does not, since there is no finite set of targets
Why flow helps on top of CoOp or CLIP-LoRA (ubpR)Replied in the threadThose methods leave some gap between image and text features, and the flow head learns that gap. So it helps most when the gap is large (CLIP-Adapter 62.46 to 70.51) and least for CoOp and CLIP-LoRA at 16 shots

To sum up, the two inconsistencies yMSg found were real and are now fixed, and j4pN's request is answered with new measurements in Table 4. The theorems, the method and the main results are the same as in the submission; what changed is mostly how clearly they are explained. We thank the reviewers again for their detailed feedback, which helped us make the paper clearer. We hope these results address the concerns raised.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.