acceptodds
Withdrawn from ICLR 2027

Beyond Text Anchors: Dual-Anchor Flow Matching for Few-Shot and Incremental Adaptation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

11 views

Abstract

Flow-based adaptations improve few-shot recognition with vision–language models by transporting visual features toward their corresponding class textual prototype. However, they implicitly assume that a semantically correct textual prototype is a reliable and discriminative transport endpoint. This assumption may fail when textual class geometry is crowded or misaligned with the visual distinctions required by a downstream task. A support-derived visual prototype offers a natural alternative by capturing task-specific appearance cues, but can itself be unreliable under scarce supervision. This complementary reliability motivates **DualFlow**, a dual-anchor flow-matching framework that retains textual and visual prototypes as two endpoint hypotheses rather than committing solely to either one. DualFlow uses a shared flow backbone with two anchor-specific heads to learn textual and visual-guided velocities, which are adaptively coordinated by a support-derived reliability coefficient shared with the classifier anchors. Beyond conventional few-shot adaptation, this dual-anchor transport-based paradigm is naturally suited to few-shot class-incremental learning (FSCIL): newly arriving classes can be incorporated by appending their textual and visual anchors and adapting only the lightweight velocity fields, while frozen vision–language encoders preserve a shared representation geometry for previously learned classes. Experiments across conventional and incremental few-shot adaptation benchmarks show that DualFlow substantially improves over text-only flow adaptation and complements parameter-efficient VLM adaptation, highlighting endpoint reliability as a critical factor in discriminative feature transport.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Few-Shot Adaptation, Flow Matching, Cross-Modal Learning

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_776hOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DualFlow, a dual-anchor flow-matching framework for few-shot and few-shot class-incremental adaptation of vision-language models. Instead of transporting visual features only toward textual class prototypes, DualFlow retains both textual prototypes and support-derived visual prototypes as complementary endpoints. A shared flow backbone with two endpoint-specific heads learns the corresponding velocity fields, which are combined using a support-derived reliability coefficient. The same formulation is extended to FSCIL by keeping the VLM encoders fixed, appending new textual and visual prototypes as classes arrive, and continually updating only the lightweight flow network. Experiments on 11 few-shot benchmarks and three FSCIL benchmarks show consistent improvements over text-only flow adaptation and competitive performance against existing peft and FSCIL methods.

Critical strengths
  1. The paper clearly identifies that textual prototypes are not always optimal transport endpoints, while visual prototypes can be noisy under limited supervision. Preserving both signals rather than committing to either one is well motivated, and the ablations show that their relative usefulness varies substantially across datasets and shot settings.
  2. DualFlow is evaluated across multiple shot settings, peft backbones, five random support-set seeds, and three FSCIL benchmarks. The component ablations also distinguish adaptive fusion from simply averaging the two anchors or training a single flow toward a fused endpoint.
Critical weaknesses
  1. The theoretical analysis derives an optimal fusion coefficient that depends on the textual risk, visual risk, and their cross-error term. However, the implemented coefficient is simply based on the ratio of estimated downside risks and ignores this cross-correlation term. So the theory establishes the potential benefit of anchor fusion, but does not directly justify the particular estimator used by DualFlow.
  2. Although old prototypes and VLM encoders are frozen, the shared velocity network is continually updated using only the current-session data. Thus, the transformation applied to old-class queries can still change. Moreover, the fusion coefficient is estimated once from the base session and reused for all future classes, despite potentially different text/visual reliability in later sessions. More explicit old-class forgetting analysis would strengthen the continual-learning claim.
Questions for authors

see weaknesses

1 comment
AuthorsOpenReview ↗

Response to Reviewer 776h

We thank the reviewer for recognizing the well-motivated dual-anchor design and comprehensive evaluation and ablation studies. Our detailed responses to the listed concerns are provided below, and the corresponding changes have been incorporated into the revised manuscript.

1.Theoretical Justification of the Practical Fusion Coefficient

We thank the reviewer for this thoughtful comment. Sec. 4.2 establishes that complementary textual and visual anchors can yield a more accurate estimate of the ideal class prototype. The oracle optimal coefficient is

α⋆=clip⁡[0,1](RT−CRT+RV−2C),\alpha^\star=\operatorname{clip}_{[0,1]}\left(\frac{R^T-C}{R^T+R^V-2C}\right),

where RT,RVR^T,R^V denote endpoint estimation risks and CC is their cross-error moment. As derived in Appendix A1.3, when C=0C=0, this reduces to RT/(RT+RV)R^T/(R^T+R^V), motivating the relative-reliability weighting principle.

In practice, ideal class centers are inestimable, making both endpoint risks and their cross-error moment hard to estimate reliably under few-shot setting. In particular, estimating CC requires characterizing the correlation between textual and visual estimation errors, which is challenging with limited support samples. We therefore adopt the simplified relative-risk formulation as a practical approximation, rather than explicitly estimating CC. We use negative classification margins as observable reliability proxies, consistent with prior work using margins to characterize classification uncertainty (Guo et al., ICML 2026, Sample Margin-Aware Recalibration of Temperature Scaling).

Empirically, Tab. A3 validates our adaptive weighting against fixed-weight fusion, improving the 1-shot difficult/easy averages from 47.2/72.1 to 56.6/82.1, and we now move this analysis from Appendix to the main paper in Tab.3 for clearer analysis.

To further validate our coefficient, we compare the support-estimated α\alpha with the best-performing α\alpha selected through dataset-specific grid search under 16-shot CLIP ViT-B/16. The former is computed directly from support samples, whereas the latter is selected post hoc based on test accuracy. We report the results below, more analysis are provided in Appendix A3.4.

CoefficientAircraftSATDTDSUNCarsPetsUCFFlowersCaltechFoodNet
Grid-selected0.60.50.50.40.40.30.40.80.30.10.3
estimated0.46300.41040.46840.39050.37540.42790.43140.80920.32360.01050.3270

Across all 11 datasets, the estimated coefficients differ from the grid-selected values by only 0.055 on average, with differences below 0.1 on 9 of 11 datasets. The test-selected grid-search results serve solely as a diagnostic reference.

These findings demonstrate that our practical estimator achieves performance close to grid-selected fusion weights, while the theoretical analysis establishes the potential benefit of complementary anchor fusion. We have clarified this distinction in Secs. 4.2–4.3.

2.Old-Class Forgetting and Fusion Stability.

We provide a base-class forgetting analysis to directly examine whether updating the shared velocity network affects previously learned classes. Specifically, we keep both the base-class test samples in the base session and candidate classes fixed across sessions, and measure the base-to-final accuracy drop solely on base samples, reporting Δ=Acc0−Accfinal\Delta=\mathrm{Acc}_0-\mathrm{Acc}_{\mathrm{final}}.

MethodminiImageNetCUB200CIFAR100
FMA73.8313.597.23
DualFlow3.393.384.43

DualFlow exhibits substantially smaller accuracy drops across all three benchmarks, demonstrating less old-class forgetting despite continually updating the velocity network with new-class data. We have incorporated this analysis into the main paper in Fig.3, with the detailed protocol and additional analysis provided in Appendix A3.6.

Regarding the fusion coefficient, our design balances stability and adaptivity. The data-rich base session provides a more reliable estimate of relative textual and visual anchor reliability than individual few-shot incremental sessions. Re-estimating the coefficient at each session could introduce calibration fluctuations affecting both velocity composition and classifier anchors for previously learned classes. We therefore retain the base-calibrated coefficient while allowing the velocity fields to adapt to incoming categories. Although this choice promotes consistent fusion geometry, we do not assume that the coefficient remains optimal for all future classes, session-wise recalibration remains a possible extension.

Reviewer_esWAOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes DualFlow, which replaces the single textual endpoint in flow-based few-shot adaptation with a textual prototype and a support-derived visual prototype, learns them through a shared flow trunk with modality-specific heads, and fuses them with a support-calibrated reliability coefficient. It further extends the formulation to few-shot class-incremental learning without replay or distillation. The core observation that endpoint adequacy matters and that textual and visual anchors have dataset-dependent reliability is sensible and practically relevant.

Critical strengths

The paper is clearly written and the motivation is well presented.

Critical weaknesses

The incremental comparison omits a directly relevant training-free two-level modality calibration method for few-shot class-incremental learning, and the flow comparison omits recent work that critiques existing flow-based adaptation and reports strong results across many benchmarks. These omissions weaken the empirical claim of competitiveness.

Adaptive combination of textual and visual prototypes for few-shot vision-language adaptation has already been explored, for example in multimodal prototype guided detection and in recent hyperbolic flow adaptation. The paper should clarify whether the contribution lies in the flow formulation, the reliability coefficient, or the incremental extension, rather than positioning itself mainly against single-endpoint flow methods.

The shared backbone with modality-specific heads, the margin-based reliability estimate, and the standard flow-matching loss are all established ingredients, and the theoretical motivation is a standard bias-variance trade-off for combining two estimators. The paper does not introduce a new mechanism that clearly explains why separate velocity fields must be learned rather than fusing the endpoints before transport.

The incremental claim is weakened by the missing comparison noted above, the name DualFlow has appeared in prior anomaly detection work and may cause confusion, and the ablation table has duplicated row labels and an incomplete column structure that make the reported numbers hard to verify.

Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration

Direct Product Flow Matching: Decoupling Radial and Angular Dynamics for Few-Shot Adaptation

Multi-modal Prototype Guided Few-shot Object Detection

Path-Decoupled Hyperbolic Flow Matching for Few-Shot Adaptation

DualFlow: Dual-Branch Flow for Unsupervised Anomaly Detection and Localization

Questions for authors

See Weaknesses

2 comments
AuthorsOpenReview ↗

Response to Reviewer esWA (1/2)

We thank the reviewer for recognizing that the paper is clearly written and that the motivation is well presented. Our detailed responses to the listed concerns are provided below, and the corresponding changes have been incorporated into the revised manuscript.

Comparison of baselines

We clarify that the relevant training-free method [a] has already been included as BiMC [CVPR 2025] in Table 4, and hyperbolic flow-based method [b] has already been compared in Tab.1 as HFM (Li et al., 2026) in our first version. Regarding the recent NeurIPS 2026 accepted paper [c] (accepted on Sep 24, 2026), we provide a comparison under the CLIP-Adapter backbone, where DualFlow outperforms DP-FM, with more comprehensive results reported in Tab. 1.

MethodDifficultEasy
CLIP-Adapter62.486.6
+DP-FM72.388.5
+DualFlow72.788.7

Related work

We would like to clarify the distinctions from the cited works.

  • HFM [b] does NOT combine textual and visual prototypes as transport endpoints. It relies exclusively on textual prototypes as transport targets, focusing instead on decoupling transport dynamics in hyperbolic space.

  • MP-DETR [d] addresses few-shot object detection and employs a trainable three-layer MLP gating network to fuse textual and visual prototypes for DETR-based detection. In contrast, DualFlow directly estimates anchor reliability from support-set classification margins without learning an additional fusion predictor. More importantly, MP-DETR does not learn modality-specific velocity fields or use multimodal prototypes as transport endpoints.

Our contribution is a reliability-aware dual-anchor transport framework that consistently coordinates text- and visual-guided velocity fields and classifier anchors for both conventional few-shot and incremental few-shot learning.

Methodological Novelty

We clarify that DualFlow is not merely a combination of existing ingredients. Our key contribution lies in addressing transport endpoint reliability in flow-based VLM adaptation, rather than the individual architectural components. Existing methods (FMA, HFM, DP-FM) rely solely on textual prototypes, which can suffer from cross-modal misalignment and crowded textual geometry (Fig. 1a). Our controlled comparison further shows that neither textual nor visual anchors are universally reliable (Fig. 1c), motivating their complementary use.

To address this, we propose a reliability-aware dual-anchor transport framework, theoretically showing when complementary anchors reduce transport target errors and using support-calibrated fusion to consistently coordinate velocity fields and classifier anchors. The same formulation naturally extends to FSCIL without replay, distillation, or incremental-specific architectures, substantially outperforming FMA across all three benchmarks (Tab. 4). To our knowledge, DualFlow is the first flow-based VLM adaptation framework unifying conventional few-shot and incremental few-shot learning through a shared dual-anchor formulation.

AuthorsOpenReview ↗

Response to Reviewer esWA (2/2)

Justification for Separate Velocity Fields

As established in Sec. 4.2 and Eq. (A3), under the ideal linear-interpolation formulation, combining the two velocity targets is mathematically equivalent to transporting toward the fused anchor. However, equivalent ideal targets do not necessarily imply equivalent learned dynamics, since the two approaches involve different interpolations and targets. Single-flow learning combines anchors before training, collapsing distinct transport geometries and supervision signals into a single target. Under limited data and model capacity, this may restrict modality-specific transport learning. In contrast, our shared-trunk, dual-head architecture captures common transport structure while allowing each head to specialize in its respective anchor geometry. The resulting velocities are then adaptively composed, preserving complementary transport supervision before fusion.

Empirically, Tab. A3 (Row 6) directly evaluates a single flow trained toward the adaptively fused endpoint using the same reliability coefficient. We have moved this analysis from the appendix to the main paper in Tab. 3 for clarity. We further repeat this comparison across five random seeds. Difficult dataset average accuracies are reported below as mean ± standard deviation, with more detailed dataset-wise results provided in Appendix A3.2:

Method1-shot4-shot16-shot
Single Head55.60±0.2063.24±0.2972.52±0.34
DualFlow56.33±0.1663.98±0.1473.03±0.27
Gain+0.73+0.74+0.51

DualFlow consistently outperforms the single-flow alternative, supporting the benefit of separate transport learning. Meanwhile, both variants outperform the text-only baseline (Tab. A3), demonstrating the effectiveness of reliability-aware anchor fusion. Together, these results validate the complementary contributions of adaptive anchor fusion and modality-specific transport learning.

Incremental Comparison

We clarify that BiMC[a] is already included in Tab. 4, as detailed in Response 1. Furthermore, Fig. 3 and Appendix A3.6 provide additional FSCIL analyses, demonstrating that DualFlow exhibits less base-class forgetting than competing methods.

Naming Clarification.

The work DualFlow: Dual-Branch Flow for Unsupervised Anomaly Detection and Localization addresses a different task, whereas our DualFlow focuses on textual- and visual-anchor transport for VLM adaptation. Our contribution does not depend on the uniqueness of the acronym.

Ablation Table Clarification.

The repeated labels (1)–(4) in Tab. 3 refer to the same configurations evaluated under different shot settings (1, 4, and 16), rather than duplicated experiments. This follows a common presentation convention in few-shot learning, as adopted in [b] and CLIP-LoRA. Regarding the column structure, Tab. 3 reports all five difficult and six easy datasets, together with their respective averages, no result columns are missing. Tab. A3 additionally provides more ablation results, with their configurations explained in Appendix A3.2.

References

[a] Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration - CVPR 2025

[b] Path-Decoupled Hyperbolic Flow Matching for Few-Shot Adaptation - ICML 2026

[c] Direct Product Flow Matching: Decoupling Radial and Angular Dynamics for Few-Shot Adaptation - NeurIPS 2026

[d] Multi-modal Prototype Guided Few-shot Object Detection - ACM MM 2025

Reviewer_FA1sOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the choice of transport endpoints in flow-based few-shot adaptation of vision-language models. It argues that text prototypes can be insufficiently discriminative, whereas support-derived visual prototypes are informative but noisy. DualFlow trains text-directed and visual-directed velocity fields using a shared backbone and separate heads. A support-estimated reliability coefficient combines both the velocity fields and the classifier anchors. The same framework is extended to few-shot class-incremental learning by freezing the feature encoders, retaining class anchors, and updating the velocity network on new-session data. Evaluation covers eleven conventional few-shot benchmarks and three FSCIL benchmarks, with additional component ablations and five-seed few-shot results in the appendix.

Critical strengths
  1. The paper is well written and easy to follow.
  2. The experimental results demonstrate the effectiveness of the proposed method.
  3. Using both textual and visual prototypes to guide feature alignment before classification is a reasonable design choice.
Critical weaknesses
  1. The use of flow matching to align visual features with class prototypes comes from FMA. DualFlow adds visual anchors and combines the two velocity fields. This seems a fairly incremental extension of the existing framework.

  2. The fusion weight needs stronger validation. The theory uses endpoint estimation errors and their cross-correlation, but the implemented rule uses negative classification margins on the support set. The paper does not establish the connection between these quantities. In the one-shot setting, different augmentations of one image may also underestimate the variation across images of the same class.

  3. Does DualFlow have convergence or stability guarantees for the learned transport? In particular, under what assumptions does combining the two velocity fields preserve the correct class correspondence? The endpoint error analysis in Section 4.2 does not directly answer this question.

  4. The additional computation is unclear. The two heads share a backbone, but the paper does not compare parameter counts, training time, inference latency or memory use with FMA and HFM. These measurements would help assess the practical cost of the gains.

  5. The paper reports component ablations and five-seed results, but provides little analysis of sensitivity to implementation choices. The ODE solver, step size and number of function evaluations are not clearly specified. It would also help to show how the results change with these settings and the number of augmentation views.

  6. The theoretical analysis motivates combining textual and visual anchors, but does not explain why two separate velocity fields are needed. At the conditional-target level, combining the two velocities is equivalent to defining a single velocity toward the fused anchor. Could the authors clarify what the dual-flow design adds beyond this simpler alternative, and whether the theory provides any justification for that choice?

Questions for authors

Please see weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer FA1s (1/3)

We thank the reviewer for recognizing that the paper is well written and easy to follow, appreciating our textual and visual prototypes design, and acknowledging the effectiveness of the proposed method demonstrated by the experimental results. Our detailed responses to the listed concerns are provided below, and the corresponding changes have been incorporated into the revised manuscript.

1. Contribution beyond FMA

We clarify that DualFlow goes beyond adding visual anchors to FMA by addressing transport endpoint reliability, overlooked by existing text-only flow methods (FMA, HFM). As shown in Fig. 1, neither textual nor visual anchors are universally reliable. DualFlow therefore learns separate text- and visual-guided velocity fields, adaptively fused using a support-estimated reliability coefficient shared with classifier prototypes. Tab. 3 shows a 47.2% → 56.6% improvement over fixed-weight fusion on 1-shot difficult datasets, while Tab. 1 demonstrates consistent gains over FMA across backbones and shot settings.

Furthermore, DualFlow extends naturally to few-shot class-incremental learning without replay, distillation, or incremental-specific architectures. By retaining existing class prototypes and adapting lightweight velocity networks, it incorporates new classes within the same framework. Tab. 4 and Fig. A1 demonstrate substantial improvements over FMA across all three FSCIL benchmarks:

MethodminiImageNetCUB200CIFAR100
FMA31.9362.5174.42
DualFlow90.2876.8377.78
Gain+58.35+14.32+3.36

2. Validation of Fusion Weight and One-shot augmentation

Validation of Fusion Weight. Sec. 4.2 establishes that complementary textual and visual anchors can reduce endpoint estimation error, with the oracle optimal visual weight

α⋆=clip⁡[0,1](RT−CRT+RV−2C),\alpha^\star=\operatorname{clip}_{[0,1]}\left(\frac{R^T-C}{R^T+R^V-2C}\right),

where RT,RVR^T,R^V denote endpoint estimation risks and CC is their cross-error moment. When C=0C=0, this reduces to RT/(RT+RV)R^T/(R^T+R^V) (Appendix A1.3), motivating our relative-reliability weighting principle.

In practice, the ideal class centers are unknown, making RT,RVR^T,R^V, and particularly the cross-error moment CC hard to estimate reliably under few-shot setting. We therefore adopt the simplified relative-risk formulation as a practical approximation, using negative classification margins as observable, task-aligned reliability proxies rather than directly estimating these theoretical quantities.

The connection is motivated by the sensitivity of classification margins to prototype errors:

∣mM(z)−mμ(z)∣≤∥ecM∥2+max⁡k≠c∥ekM∥2.|m^M(z)-m^\mu(z)|\leq \lVert e_c^M\rVert_2+\max_{k\ne c}\lVert e_k^M\rVert_2.

where ekM=AkM−μke_k^M=A_k^M-\mu_k and ∥z∥2=1\|z\|_2=1. Thus, endpoint estimation errors can directly perturb classification margins, motivating margin-based unreliability as a practical surrogate consistent with prior work using margins to characterize classification uncertainty (Guo et al., ICML 2026, Sample Margin-Aware Recalibration of Temperature Scaling).

Empirically, we validate our adaptive weighting through two comparisons. First, the ablation (now moved from Tab. A3 to main-text Tab. 3) shows 1-shot difficult/easy accuracy improvements from 47.2/72.1 to 56.6/82.1 over fixed-weight fusion. Second, we compare our support-estimated α\alpha with the best-performing coefficient from dataset-specific grid search under 16-shot CLIP. The former uses only support samples, while the latter is selected post hoc using test accuracy as a diagnostic reference. Results on five difficult datasets are shown below, with further analysis in Appendices A3.3–A3.4.

CoefficientAircraftSATDTDSUNCars
Grid-selected0.60.50.50.40.4
Estimated0.46300.41040.46840.39050.3754

Across all 11 datasets, the coefficients differ by only 0.055 on average, with differences below 0.1 on 9 of 11 datasets, supporting the effectiveness of our margin-based weighting without coefficient search.

One-shot augmentation. We clarify that one-shot augmentation is not intended to fully recover the variation across different images of the same class, which is inherently difficult with only a single support example. Instead, augmentations provide supplementary information about local visual variability, serving as a practical approximation for estimating visual-anchor reliability under scarce supervision.

To assess its robustness, we conduct a sensitivity analysis using 2, 4, 8, 16, and 32 augmentation views (See responses to Q5 and Appendix A3.3). The results show that performance remains largely stable across different view counts. This suggests that, although augmentation-based uncertainty cannot capture the full intra-class variation, it provides a reasonably stable and effective reliability signal in the one-shot setting.

AuthorsOpenReview ↗

Response to Reviewer FA1s (2/3)

3. Stability and preservation of class correspondence

We clarify that endpoint estimation, stability of learned dynamics, and class correspondence are distinct properties. Sec. 4.2 addresses endpoint estimation, while the stability of composed dynamics follows under standard regularity conditions.

For fixed α∈[0,1]\alpha\in[0,1], if vTv_T and vVv_V are continuous in time and state-Lipschitz with constants LTL_T and LVL_V on the relevant trajectory domain, then

L_\alpha\le(1-\alpha)L_T+\alpha L_V.$$ Provided trajectories remain in this domain, the composed field inherits **existence, uniqueness, and continuous dependence on perturbations**. Class correspondence additionally requires a **positive classification margin.** Let $q^*(\tau)$ denote a correctly classified reference transport output and $a_k$ the unit-normalized classifier anchors. For class $c$, define $$m=\min_{k\ne c}\langle q^*(\tau),a_c-a_k\rangle>0.$$ Since $\lVert a_c-a_k\rVert_2\leq 2$, the predicted class remains unchanged whenever $\lVert\widehat{q}(\tau)-q^*(\tau)\rVert_2<m/2$. If the velocity approximation error along the reference trajectory is bounded by $\varepsilon$, and $v_\alpha$ is state-Lipschitz with constant $L$, a sufficient condition is $$\varepsilon\frac{e^{L\tau}-1}{L} +\varepsilon_{\mathrm{solver}}<\frac{m}{2}.$$ where $\varepsilon_{\mathrm{solver}}$ denotes numerical integration error, and the first term becomes $\varepsilon\tau$ when $L=0$. Normalizing a nonzero final query does not change its class ranking against the same anchors. Thus, class preservation requires a **correctly classified reference, bounded transport and numerical errors, and sufficient classification margin**. Correct sample–class pairing provides supervision, but finite-sample flow-matching loss alone does not guarantee these conditions. We have clarified the theoretical discussion in **Sec. 4.2** and added the conditional stability analysis in **Appendix A1.5**. >**4. Additional parameters and computational cost.** We report the measured computational costs. All methods use **fixed one-step inference (NFE = 1)**. |Method|Params (M)|Train/step (ms)|PeakTrainMem (MiB)|Infer (ms)|PeakInferMem (MiB)| |---|---:|---:|---:|---:|---:| |FMA|17.217|22.976|344.881|4.566|108.125| |HFM|17.283|23.989|370.702|5.480|111.694| |DualFlow|18.004|27.058|361.036|4.669|112.135| Compared with FMA, DualFlow introduces 4.57% more parameters, while inference latency increases by only **2.3%** and peak inference memory by **3.7%**. Detailed analysis has been added in **Appendix A3.7**.
AuthorsOpenReview ↗

Response to Reviewer FA1s (3/3)

5. Implementation Details and Sensitivity Analysis

We now provide more detailed implementation specifications in the revised paper. Following FMA, we adopt the explicit Euler solver and we take a single integration step (NFE = 1). We have added an ODE-solver sensitivity analysis in Appendix A3.5, comparing Euler, Heun, and RK. Average test accuracies over difficult and easy datasets are summarized below where we obtain stable performance across different ODE solver.

SolverDifficultEasy
Euler72.788.4
Heun72.788.4
RK72.788.4

Furthermore, we provide an additional sensitivity analysis on the number of augmentation views, evaluated across three random seeds. The results report mean accuracy ± standard deviation, averaged over difficult and easy datasets.

ViewsDifficultEasy
256.63±0.7081.97±0.24
457.31±0.2881.97±0.22
857.39±0.3281.89±0.16
1657.34±0.3181.86±0.11
3257.29±0.2581.81±0.13

The results demonstrate that DualFlow is largely insensitive to the number of augmentation views. Performance remains stable across different augmentation budgets. This supports the robustness of our augmentation-based reliability estimation under scarce supervision. Detailed analysis are provided in Appendix A3.3.

6. Justification for Separate Velocity Fields

We clarify that, at the conditional-target level, combining textual and visual transport directions is mathematically equivalent to transporting toward the fused anchor (Sec. 4.2, Eq. A3). However, equivalent conditional targets do not necessarily imply equivalent learned dynamics, as the two formulations differ in their training objectives and interpolation paths.

Specifically, Single Flow combines textual and visual anchors before training and learns one velocity field along the fused interpolation path. In contrast, DualFlow separately supervises text- and visual-guided velocities along their respective interpolation paths before adaptive composition. This retains modality-specific transport supervision that early fusion may lose, potentially benefiting learning under limited data and model capacity. While our theoretical analysis motivates complementary anchor fusion rather than guaranteeing the superiority of separate fields, we empirically validate the benefit of this design.

We directly ablate the suggested single-flow alternative, denoted as Single, using the same adaptive coefficient and fused classifier anchors (Row 6 in Tab. 3). We further repeat this comparison across five random seeds under the 1-, 4-, and 16-shot settings (Appendix A3.2). The difficult-dataset averages are reported below as mean accuracy ± std, with gains measured in percentage points:

Method1-shot4-shot16-shot
Single55.60±0.2063.24±0.2972.52±0.34
DualFlow56.33±0.1663.98±0.1473.03±0.27
Gain+0.73+0.74+0.51

DualFlow achieves higher average accuracy and lower standard deviation across all three shot settings, providing empirical evidence that preserving separate modality-specific transport supervision offers benefits beyond early fusion, despite their conditional-target equivalence.

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

Related papers

Discussion (0)

Sign in to comment.