acceptodds
Under review as a conference paper at ICLR 2027

Semantic Fields Are Interfaces: Consumer-Typed Functional Migration for Representation Autoencoders

Abstract

Representation autoencoders (RAEs) use pretrained visual features as generative states. These dense fields also support retrieval and downstream heads, forming shared interfaces whose consumers may persist across producer releases. Producer upgrades may preserve tensor shape while changing coordinate semantics, rendering retained consumers incompatible. We study post-release field migration, in which an already-trained successor serves retained consumers through one release-wide map while the generator, representation models, and consumers remain frozen. Consumer-Typed Functional Migration (CTFM) learns corrections around a host-specific calibrated base and supervises repair with archived consumer responses rather than field matching alone. Its contract separates behavior to recover from coordinates to preserve: IQ-Safe repairs decoding while exactly preserving the base-calibrated pooled coordinate, whereas Quality permits full-field correction. RAEv2 optionally employs an ownership-preserving, affine-initialized low-rank refinement. Across DINO-based RAEv2 and Perception Encoder (PE) RAEs, CTFM improves recognition, fidelity, semantic consistency, and distribution quality over matched post-hoc alternatives. On 3,232 new RAEv2 fields, a 1.05M-parameter map outperforms identical-capacity feature matching on all five primary metrics. With one quarter of the refinement samples, CTFM gains 3.87 ResNet-50 points and 1.82 dB PSNR over full-data matching; with half the PE residual-fitting classes, it gains 10.88 ResNet-50 points over the feature-to-task procedure. On 7,200 real-image queries, IQ-Safe preserves every calibrated-base retrieval decision while improving decoded PSNR by 5.95 dB. These results support consumer-defined compatibility for practical reuse of dense RAE fields under fixed consumer contracts.

Primary area:
generative models
Keywords:
RAE, representation autoencoders, backward compatibility, model stitching, dense visual representations, image retrieval

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_D7vhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the “post-release migration” problem in Representation Autoencoders (RAE). RAE uses a pre-trained visual feature grid (semantic field) as a generative latent variable, which is simultaneously read by downstream components such as the decoder and the retrieval gallery. After a producer upgrade, the shape of the field remains unchanged but the meaning of its coordinates changes, rendering the old components obsolete. The authors propose CTFM: under the premise of freezing all models, it learns a globally shared mapping to transform the new field into the old field space. This method first performs a baseline calibration, then uses the output responses of the old components to supervise a small residual, and strictly preserves the pooling coordinates read by the retriever through zero-mean projection (IQ-SAFE). Optionally, a low-rank refinement stage can be added on top of RAEv2. Experiments were conducted on two model pairs, DINOv3→DINOv2 and PE, and the results show that CTFM outperforms baselines of the same capacity in terms of recognition, fidelity, and distribution quality. It requires less training data, leaves retrieval decisions completely unchanged, can be transferred to distribution-shifted data without retraining, and incurs a deployment overhead of approximately 1% of the decoder’s runtime.

Critical strengths

1. Transparent experimental documentation. The paper reports data splits, checkpoint and initialization selection rules in unusual detail, and evaluates with independent observers excluded from fitting. This makes the claims verifiable and the experiments reproducible.

2. Practical feasibility. On RAEv2, a lightweight map (~1M learned parameters, ~1% added latency) lifts decoding from near failure under direct reuse close to the archived field, without modifying any released model. The same map also transfers to real-image distribution shifts without refitting. This shows that post-hoc migration of dense RAE fields can be done cheaply in practice.

Critical weaknesses
  1. Weak motivation. Prior compatibility work retains consumers that are costly to replace, e.g., plugin ecosystems [1] or large indices [2]. Here, keeping the old decoder is justified by one sentence (“decoder replacement can require client updates”) without a concrete scenario. Moreover, the paper’s own results suggest that giving each consumer its own map would dominate: the unconstrained QUALITY map decodes better (e.g., PE ResNet-50 58.25 vs. 54.44, Tabs. 12 and 33), while retrieval only needs the base mean. The authors should clarify when a single shared field is actually required.

  2. Limited novelty. The base calibration follows the latent translation recipe of [3], which is not cited. On PE, the residual objective is essentially the FFM objective of the VFM baseline (matching the frozen target's final hidden features).

  3. Marginal gains over the strongest baseline. The pooled-coordinate guarantee is shared by the projected affine VFM baseline (Tab. 4), over which gains are marginal (53.28 vs. 52.70 Gallery Top-1, Tab. 2) or slightly negative (75.65 vs. 75.897 prospective ResNet-50, Tab. 10).

  4. Weak identical-capacity control. A central piece of evidence is the gain over identical-capacity feature matching, but this control degrades the shared initialization (ResNet-50 74.88 → 71.84, PSNR 16.967 → 15.939, Tab. 2). Relative to the initialization, CTFM improves only modestly (to 75.96 and 17.816), so this comparison overstates the benefit of response supervision. Moreover, the reported intervals and seed repeats (Tabs. 11 and 24) only cover CTFM vs. feature matching, so it is unclear whether the 1-point gain over the initialization exceeds variance. Since the reduced-data comparison in Tab. 3 shares the same initialization, the initialization should be reported there as well.

Minor:

I found the main method’s pipeline (figure 2) hard to follow.

[1] Ran et al. X-Adapter: Adding Universal Compatibility of Plugins for Upgraded Diffusion Model. CVPR 2024.

[2] Vejendla. Drift-Adapter: A Practical Approach to Near Zero-Downtime Embedding Model Upgrades in Vector Databases. EMNLP 2025.

[3] Maiorca et al. Latent Space Translation via Semantic Alignment. NeurIPS 2023.

Questions for authors
  1. In what concrete deployment scenario can the old decoder not be replaced, and when is a single shared field required instead of separate maps for the decoder and the gallery?

  2. How does the base calibration differ from latent space translation, and how does the PE residual objective differ from VFM's FFM?

  3. Given that projected affine VFM offers the same retrieval guarantee, what practical advantage justifies CTFM's small recognition gains over it?

  4. How much does CTFM improve over the shared initialization on the prospective sets (Tab. 3), and is this gain significant?

6 comments
AuthorsOpenReview ↗

Response to Reviewer D7vh (1/2)

Thank you for requesting a direct comparison with the shared initialization and a closer examination of projected affine VFM. We now report both, including validation-based retuning and uncertainty estimates.

W4 and Q4. How much does CTFM improve over the shared initialization, and is the gain significant?

On the same prospective states as Tab. 10, and on the second prospective set of Tab. 3, the submitted CTFM improves its compiled initialization on all seven endpoints, with statistically significant effects (recognition in percentage points):

EndpointProspective set 1 (Tab. 10)Prospective set 2, 25% data (Tab. 3)
Decoded Gallery Top-1 ↑+1.27+0.96
ResNet-50 Top-1 ↑+1.55+1.21
Independent Top-1 ↑+1.35+1.58
PSNR (dB) ↑+0.84+0.74
LPIPS reduction ↑0.01660.0085
Cycle reduction ↑0.07040.0508

Every entry is statistically significant under the class bootstrap of Tab. 11. The initialization is fixed before refinement and shared by CTFM and the feature-matching control, so these gains isolate the refinement. As a separate robustness check, Tab. 31 (Tab. 24 in the submission) refits the refinement with three data orders, and its paired effects against feature matching vary with standard deviation at most 0.143 percentage points and 0.006 dB. The revised Tabs. 3 and 10 include the initialization rows, and the new Tab. 12 marks the significance of every effect.

We agree that the margin over feature matching includes the control's own decline. We examined whether additional fitting or learning-rate selection explained this decline. In these runs, validation field error falls monotonically (1.000 to 0.856) and its selector picks the last step, while ResNet-50 and PSNR fall at every saved checkpoint (74.10 to 72.25, 16.90 to 15.89 dB).

Retuned over the same seven learning rates as our affine map, its best run by its own loss ends 7.05 ResNet-50 points below the initialization. Selected by the declared response risk instead, it returns the initialization at every tested rate. The table above therefore measures response refinement directly against the shared starting point.

W3 and Q3. The margin over projected affine VFM, and the practical advantage at the same guarantee.

The submitted prospective result slightly underperformed projected affine VFM on ResNet-50. We therefore repeated refinement with validation-based learning-rate selection. The submitted rate was 10−510^{-5}. Selection over 10−510^{-5} to 10−310^{-3} chooses 10−410^{-4}. The compiled CTFM then improves over projected affine VFM on all seven endpoints on both prospective sets (ResNet-50 +1.27 and +2.97, PSNR +1.41 dB on both sets), every gain significant. This retuning followed evaluation of the affine map to compare both at equal tuning. App. A.6 documents the sequence.

The full affine map trained with the same response loss yields larger gains over projected affine VFM (ResNet-50 +2.32 and +3.06, about +2.0 dB on both sets). It also outperforms projected VFM in all 24 endpoint and domain comparisons on four real shifts (revised Tab. 23). The revision adds per-population class-bootstrap comparisons alongside the submitted macro and observer analyses (Tabs. 22 and 37, numbered 19 and 30 in the submission).

Practical advantage under the same guarantee. Projected affine VFM and CTFM preserve the same pooled coordinate. The matched comparison shows that response supervision improves recognition, fidelity and cycle consistency while retaining this guarantee at the same weight count.

AuthorsOpenReview ↗

Response to Reviewer D7vh (2/2)

Additional control: continued fitting of projected affine VFM.

We continued VFM fitting on its own feature loss with the same update budget, selecting checkpoints by either feature loss or the declared response risk. This separates response information used for optimization from information used only for selection. App. D.2 documents the shared learning-rate grid and its further extension for this baseline.

Selected by that loss, it leaves ResNet-50 unchanged on prospective set 1 (75.90) and adds 0.62 dB. Selected by our response risk, it shows no significant difference from CTFM (affine) in 14 of 18 recognition comparisons across two generated sets and four real shifts, and is significantly higher in the other 4.

Starting CTFM's response refinement from this stronger feature fit improves PSNR by 0.15 to 1.07 dB and cycle error in all six populations. Recognition shows no significant change in 13 of 18 comparisons, is significantly better in 4, and significantly worse in 1. LPIPS, which neither response supervises, is 0.01 to 0.02 worse. Relative to continuing the feature fit to 2,400 updates, the two-stage map improves PSNR by 1.10 and 1.04 dB on the two generated sets. App. D.2 specifies its longer, 3,600-update budget.

These controls identify benefits from consumer-response information both during optimization and during checkpoint selection. Revised App. D.2 (Tabs. 13 and 14) reports both selection rules and the two-stage map, including the recognition and LPIPS trade-offs.

W1 and Q1. When is a single shared field required?

A shared field is relevant when one stored or broadcast tensor must remain executable by multiple retained readers. CTFM provides IQ-Safe for this requirement and Quality when a serving layer can deliver separate inputs (Sec. 2.2). Scale-RAE [1] illustrates readers beyond decoding: "the multimodal model can directly reason over generated latents".

On prospective set 1, the Quality affine map trained with the same response loss gains 1.83 ResNet-50 percentage points and 0.21 dB PSNR, and reduces LPIPS by 0.028 relative to the IQ-Safe map. The decoded-gallery difference (+0.09) is not significant, and cycle error slightly favors the shared field. These results quantify the decoder-quality trade-off when separate inputs can be routed to different consumers.

Decision stability under small margins. Unprojected affine VFM changes 99.78% and 99.95% of installed retrieval decisions on ImageNet-R and Sketch (App. H.10, Fig. 10). The smallest top-two score gaps are 1.36×10−51.36\times10^{-5} and 3.83×10−63.83\times10^{-6} (Tab. 41). Releasing 1% of the affine map's pooled displacement flips 0.9% and 1.0% of retrieval decisions while recovering 2.6% and 1.9% of the available PSNR gain. A soft penalty alone does not certify zero flips. The projected maps satisfy the retrieval-margin check on all 7,200 queries.

W2 and Q2. Base calibration and the PE objective.

Thank you for the pointer to Maiorca et al. [2]. Our base starts from an affine alignment between frozen latent spaces of the kind they study, then adds host-specific corrections fitted before repair (App. B.1). We have added this attribution in Sec. 3.1. Base calibration is shared by matched controls. The contribution evaluated here is subsequent response-guided repair under ownership constraints.

We have also made the PE objective explicit in Sec. 3.1: it is final-feature matching under the ownership constraint (App. B.2). PE tests the contract and actuator under this established functional (Tab. 38), gaining 9.06 ResNet-50 percentage points over the matched FFM-to-TLT procedure while preserving the pooled coordinate (Tab. 15). RAEv2 compares the response loss through decoder output and archived re-encoding with feature matching under matched conditions.

Minor. Fig. 2 is hard to follow.

We apologize for the lack of clarity in the previous figure. We have revised Fig. 2 to separate compact repair from affine refinement, highlight the protected pooled-coordinate path, and clarify each host's supervision.

References

[1] Tong et al. Scaling Text-to-Image Diffusion Transformers with Representation Autoencoders. arXiv:2601.16208.

[2] Maiorca et al. Latent Space Translation via Semantic Alignment. NeurIPS 2023.

Thanks for the response. The added experiment has resolved some of my concerns. I still have some followup questions.

  1. According to your own results, routing separate inputs to the decoder and the retrieval head is not worse than the shared IQ-Safe field. So when exactly do we need IQ-Safe? Could you give a concrete example in the real world that separate inputs cannot be routed and the archived decoder can't be replaced?

  2. IQ-Safe will keep the retrieval input unchanged, but in the submitted version, this only gives only about 3% Top-1 for the linear pooled head, so most of it is wrong. Retrieval only gets good after the Procrustes calibration you added in the rebuttal. Is that calibration now part of your method?

Reviewer_PLZDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies compatibility in representation autoencoders when a new representation producer must serve retained consumers. The proposed Consumer-Typed Functional Migration (CTFM) learns a field mapping using frozen consumer responses as supervision. Its IQ-Safe mode restricts corrections to preserve the calibrated pooled coordinate, allowing decoding to improve without changing the input to a pooled retrieval consumer. The RAEv2 implementation additionally introduces affine-initialized low-rank refinement.

The main paper evaluates the approach on DINOv3-to-DINOv2 and PE-Spatial-to-PE-Core pairings, reporting improvements in decoded-image recognition and fidelity, reduced-data fitting, and transfer without refitting.

Critical strengths
  1. A coherent formulation of a plausible compatibility problem. Producer and consumer upgrades need not occur simultaneously, and retaining existing consumers can be a reasonable deployment requirement. CTFM explicitly separates the responses to recover from the coordinates to preserve. Its projection-based construction directly enforces the latter requirement, providing a clear mechanism for improving decoding without further perturbing calibrated pooled retrieval.

  2. Meaningful empirical improvements within the stated setting. The main paper shows that CTFM improves upon capacity-matched feature matching. In Table 3, the matched-data RAEv2 comparison improves ResNet-50 accuracy from 72.246% to 75.650% and PSNR from 15.885 to 17.744 dB. Table 4(a) also directly demonstrates the intended trade-off: retrieval metrics remain unchanged while decoding improves. These results support the effectiveness of the approach for reducing compatibility loss.

Critical weaknesses
  1. The practical significance of the specific compatibility setting is insufficiently established. The scenario is plausible, but the main paper does not clearly demonstrate why the shared-field and frozen-consumer constraints are necessary, or what producer-upgrade benefits justify the remaining migration loss. Even the selected favorable examples in Figure 4 retain noticeable visual discrepancies from the references. The relative improvements are convincing, but the overall benefit of switching producers remains unclear.

  2. The methodological contribution is incremental. CTFM combines consumer-response distillation with linear coordinate preservation. Given the competitive projected affine baseline in Table 2, the main paper does not sufficiently establish the additional value of the multi-stage construction over a simpler affine map trained with the same response objective and preservation constraint.

Questions for authors
  1. What concrete deployment requirement or quantified maintenance cost makes retaining the old consumers and sharing a single migrated field important in the evaluated setting?

  2. What specific advantage does the new producer offer over the old system, and how much of that advantage remains after migration? This would help contextualize the residual quality gap visible in the quantitative results and Figure 4.

  3. How does CTFM compare with a full affine map trained using the same RGB-plus-semantic response objective and pooled-coordinate preservation constraint?

4 comments
AuthorsOpenReview ↗

Response to Reviewer PLZD (1/2)

Thank you for suggesting this direct comparison. We evaluated a full affine map with the same response objective and ownership constraint, and used the result to simplify the scalable parameterization.

Q3 and W2. How does CTFM compare with a full affine map trained with the same response objective and preservation constraint?

The map has 1,049,600 weights and starts from projected affine VFM. It is trained with the paper's RGB plus semantic cycle loss (Eq. 9) under the same projector P0P_0, on the same 640 fitting states, for the same 2,400 updates. Its learning rate was chosen from seven values jointly with the checkpoint, on validation states only, and our compiled map was retuned under the same rule (App. A.6 gives the full grid). Absolute results on prospective set 1 (3,232 fresh states, Tab. 10, Top-1 in %):

Map (1.05M weights each)Decoded Gallery ↑ResNet-50 ↑Indep. ↑PSNR (dB) ↑LPIPS ↓Cycle ↓
Projected affine VFM53.8175.9076.9817.200.2410.563
CTFM, compiled (paper)54.5575.6577.4917.740.2380.528
CTFM, compiled, rate chosen on validation55.4877.1779.3218.600.2280.471
CTFM, affine55.5178.2279.5619.170.2210.440

Among these four configurations, the full affine map achieves the highest recognition accuracies and PSNR, with the lowest LPIPS and cycle error. Across three seeds its standard deviations are 0.11, 0.28 and 0.12 points and 0.10 dB, and its worst seed still improves over projected VFM by 1.70, 1.76 and 2.55 points and 1.77 dB. Improvements over projected affine VFM also hold on the second prospective set, the system set of Tab. 2 and four real shifts (general response).

Against the compiled form at equal tuning, the affine map is better on PSNR (+0.57 and +0.55 dB), LPIPS and cycle error on both prospective sets. It is better on ResNet-50 on set 1 (+1.05), with no significant difference on set 2, and decoded-gallery and independent Top-1 show no significant difference on either set. Refined on a quarter of the refinement data (40 classes), it still improves over projected affine VFM on all seven endpoints of set 2 (ResNet-50 +1.55, PSNR +0.85 dB).

The requested experiment shows that response-guided repair remains effective with a simpler affine parameterization. The affine initialization is the full-rank case of the submitted construction (App. A.5). We therefore present this variant as the scalable implementation (Sec. 3.3, Tabs. 2, 3 and 10) and retain the compiled form as an ablation, with Proposition 2. The consumer contract and exact ownership guarantee remain unchanged, preserving the separation of behavior and coordinates highlighted in your review.

AuthorsOpenReview ↗

Response to Reviewer PLZD (2/2)

Q1 and W1. What deployment requirement makes retained consumers and one shared field important?

We study stored or broadcast fields that retained consumers must read as the same tensor. IQ-Safe preserves the pooled coordinate within that shared field while repairing decoding. When a serving layer can route separate inputs, CTFM's Quality contract permits full-field correction. Sec. 2.2 states these deployment conditions.

The costs can be stated per query. In FLOPs, the map adds 0.5% to 0.6% of one decoder pass. Rendering plus retrieval through the retained decoder and EoldE_{old} costs about 1.75 decoder passes in total. Field repair raises this route's ImageNet-R R@1 from 3.6% with the calibrated base to 25.0% with CTFM. The successor-decoder route is evaluated separately. Decoder replacement changes the installed renderer, while gallery recomputation requires the stored items' original pixels. App. E.2 reports the measured accuracy and compute for each route.

Q2. What does the new producer offer, and how much of that advantage remains after migration?

The RAEv2 transition is one the RAE line itself made. The original RAE used DINOv2-B, and RAEv2 moved its default encoder to DINOv3-L. In the RAEv2 encoder ablation of Singh et al. [1] (k=1k=1, 20 epochs, no guidance), these two encoders reach gFID 2.81 vs 2.61 and ImageNet linear-probe accuracy 83.9 vs 87.0. Our archived encoder is the larger DINOv2-L.

The PE pair is a specialization. In the same paper, PE-Spatial-L improves on PE-Core-L (listed there as PE-L) in the LDS spatial metric (0.34 vs 0.14) and in RAEv2 gFID (3.97 vs 4.09), at lower global linear-probe accuracy (78.4 vs 85.5). It models a team adopting a spatially stronger producer while consumers built on PE-Core stay installed. Sec. 4.1 and App. B now describe it that way.

The published encoder ablations motivate the release choices. Our measurements quantify compatibility recovery after migration. On the system set of Tab. 2, the following percentages give the share of the gap from direct reuse to the archived field that each map closes (higher is better for every column):

MapDecoded Gallery ↑ResNet-50 ↑PSNR ↑LPIPS ↑Cycle ↑
CTFM, compiled (paper, Sec. 4.2)95.6%93.9%83.5%86.3%85.0%
CTFM, affine98.3%96.2%97.6%89.0%90.7%

Our comparisons measure how well retained consumers execute fixed successor outputs. Neither evaluated graph contains a generator for the archived representation, so we do not measure the old end-to-end pipeline for these pairs. Sec. 4.1 now states this evaluation scope.

Fig. 4. The favorable examples still show visible differences from the references.

We agree that residual distortions matter for deployment. The references in Fig. 4 come from DnewD_{new}, while every compared method uses the same frozen DoldD_{old}. Tab. 2 therefore includes an archived encode-decode control, which reconstructs the native images imperfectly even from exact archived fields, providing a reference for the retained renderer's own reconstruction error. CTFM enables successor fields to serve the installed decoder and pooled consumers while preserving the calibrated pooled input, without consumer retraining or gallery backfilling. App. E.2 compares this shared-field option with native-decoder re-encoding and routing, so deployments that can replace consumers or route separate inputs can assess the quality and compute trade-off.

References

[1] Singh et al. Improved Baselines with Representation Autoencoders. arXiv:2605.18324.

Thank you for the detailed response and the additional affine comparison. The response clarifies two important points:

  1. The deployment scenario is plausible. Producer transitions, retained consumers, shared fields, and migration costs are now described more concretely.
  2. Response-guided repair is effective with a simpler full-affine parameterization. Under the same ownership constraint and response objective, the full-affine CTFM reaches 78.22% ResNet-50 accuracy and 19.17 dB PSNR, compared with 75.90% and 17.20 dB for projected affine VFM. This substantially strengthens the empirical evidence for compatibility recovery.

My remaining concerns are:

  1. The practical importance of the problem is still not fully established. The response shows that the scenario is feasible and that migration is efficient, but it does not clearly demonstrate how prevalent or consequential this deployment setting is, or whether the benefits of switching producers justify the residual quality loss.
  2. The affine comparison clarifies that the main empirical gain comes from response-guided repair under coordinate preservation. This makes the method simpler and more convincing, but also suggests that the methodological contribution is an incremental combination of functional distillation and a linear ownership constraint.

Base on above, I therefore maintain my original rating.

Thank you for recognizing the stronger affine result. LINE Part Time Jobs documents producer retraining disrupting downstream reranking and uses post-processing alignment to allow independent updates [1]. LINE adopted the embedding-stabilization approach developed by Netflix researchers and published at RecSys 2025 [2]. Published RAE encoder comparisons motivate the DINOv3 transition and PE's spatial specialization [3]. Our experiments measure the compatibility loss and serving cost of migrating fixed successor outputs for retained consumers. The archived encode-decode control in Tab. 2 provides a reconstruction reference for the retained decoder, and App. E.2 compares routing and decode-and-re-encode alternatives. With the same affine parameterization and pooled-input preservation constraint, response-guided CTFM improves over projected affine VFM by 2.32 ResNet-50 points and 1.97 dB PSNR on 3232 fresh states. The simpler result strengthens the evidence for response-guided repair under the consumer contract without relying on the original multi-stage architecture. Could you clarify the closest prior work underlying the incremental-novelty concern and how you weigh the new affine evidence alongside the contract formulation?

References

[1] Kihara and Yuan. Embedding stabilization for search re-ranking: Overcoming cold start at LINE Part Time Jobs. LY Corporation Tech Blog, 2026.

[2] Zielnicki and Hsiao. Orthogonal Low Rank Embedding Stabilization. RecSys 2025.

[3] Singh et al. Improved Baselines with Representation Autoencoders. arXiv:2605.18324.

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

This paper proposes Consumer-Typed Functional Migration (CTFM) to address representation incompatibility when an RAE producer is upgraded while downstream consumers remain frozen. The method combines base calibration, consumer-response supervision, and ownership-constrained residual correction to restore downstream functionality while preserving selected representation coordinates. Experiments on RAEv2 and Perception Encoder demonstrate improvements in recognition, reconstruction fidelity, and semantic consistency over existing alignment baselines.

Critical strengths
  1. The paper addresses an important problem of maintaining compatibility between upgraded representation models and existing downstream consumers.

  2. The consumer-typed interface contracts provide a clear framework for separating functional recovery from coordinate preservation.

  3. The proposed projection mechanism ensures exact preservation of protected pooled representations.

  4. Experiments demonstrate consistent improvements over several baselines, with relatively low additional parameter and inference costs.

Critical weaknesses
  1. The main components, including functional distillation, residual adaptation, null-space projection, and low-rank refinement, are closely related to existing techniques. The contribution appears to be primarily their combination.

  2. Experiments cover only two RAE systems and specific release pairs, leaving robustness to larger distribution shifts, different architectures, and more diverse consumers unclear.

  3. The multi-stage pipeline makes it difficult to determine how much improvement comes from response supervision versus initialization, model capacity, or refinement. More targeted ablations would strengthen the claims.

Questions for authors

none

2 comments
AuthorsOpenReview ↗

Response to Reviewer qZvk (1/2)

Thank you for recognizing the contract formulation and for highlighting novelty, experimental breadth, and attribution. We address these concerns with the additional affine comparison and pooled-head evaluation, alongside the existing evidence.

W1. The main components are closely related to existing techniques, and the contribution appears to be their combination.

CTFM studies compatibility at a dense-field interface where decoder-response recovery and preservation of a declared pooled statistic must be satisfied together. The contract τ=(F,O)\tau = (F, O) states both requirements, and the construction enforces the second exactly (Proposition 1). The matched comparisons below establish the additional recovery provided by consumer-response guidance under the same capacity and ownership constraints.

CTFM builds on established alignment, distillation, and projection techniques. Secs. 2.3 and 3.1 now make these connections explicit. The table summarizes their roles in the formulation and comparisons.

Prior lineWhat it providesConnection to CTFM
Latent-space translation [1]Linear or orthogonal maps between frozen latent spaces, estimated from paired samplesOur base A0A_0 starts from such an alignment. These maps are fitted on latent correspondences, without supervision from what a frozen decoder renders
Model stitching, VFM [2]A stitch trained to match the target model's final featuresA strong alignment baseline evaluated under the same ownership constraint. Selected by its own loss, its continuation adds at most 0.6 ResNet-50 points on the two prospective sets (W3)
Knowledge distillation [3]Supervision through a teacher's outputsWe distill archived consumer responses into a release-wide field map, with producer and consumers frozen
Null-space-constrained updates [4, 5]Updates restricted to subspaces that limit interference with preserved knowledgeWe keep a declared linear consumer statistic exactly invariant. The projector is analytic, its guarantee covers every ψ∘Q\psi\circ Q reader, and the construction composes with regional means (Tab. 16)
Backward-compatible representations [6, 7]Embeddings that stay comparable across model versionsThese are global embedding vectors. We need one dense field that an immutable decoder can render and pooled readers can read at the same time

W3. How much of the improvement comes from response supervision, versus initialization, capacity or refinement?

Results on prospective set 1 (3,232 fresh states) separate compact repair from the compiled and affine branches. ResNet-50 Top-1 is in %, and weight counts exclude the shared frozen base.

StageWeightsResNet-50 ↑PSNR (dB) ↑
Direct reuse00.069.36
Calibrated base08.1411.55
Compact response-trained residual163.8K36.3616.47
Compiled branch: initialization1.05M74.1016.90
+ feature-matching refinement1.05M72.2515.89
+ response refinement (submitted)1.05M75.6517.74
Affine branch: projected VFM1.05M75.9017.20
+ its feature loss, continued1.05M75.9017.82
+ response refinement, CTFM (affine)1.05M78.2219.17

Within each branch, the refinement comparison holds capacity, initialization and update budget fixed. Response refinement gains 1.55 ResNet-50 points and 0.84 dB from the compiled initialization, and 2.32 points and 1.97 dB from projected affine VFM (Tabs. 10 and 12). On this set, feature-loss refinement selected by its own loss leaves ResNet-50 unchanged or lower. Response refinement improves all seven endpoints from both initializations, with statistically significant gains under the class bootstrap. Across three seeds of CTFM (affine), standard deviations are at most 0.28 ResNet-50 points and 0.10 dB.

AuthorsOpenReview ↗

Response to Reviewer qZvk (2/2)

W2. Only two RAE systems and specific release pairs. Robustness to larger shifts, different architectures and more diverse consumers is unclear.

The revision adds affine-map evaluations across four shifts and a frozen pooled-head test, complementing the existing evidence from the DINO and PE hosts.

Larger distribution shifts. The maps are applied without refitting to four real-image shifts, 25,383 images in total. CTFM (affine) minus projected affine VFM (recognition gains in percentage points, PSNR in dB):

PopulationImagesDecoded Gallery ↑ResNet-50 ↑Indep. ↑PSNR (dB) ↑
ImageNet-R3,200+1.47+2.25+2.47+2.15
ImageNet-Sketch4,000+0.97+2.15+1.44+4.46
ImageNet-A7,500+1.32+1.30+0.90+1.64
ImageNetV2-D10,683+1.60+3.22+3.23+1.61

LPIPS and cycle error improve on every shift as well, and all 24 effects across these six endpoint families favor CTFM (affine) and are statistically significant (revised Tab. 23). The shifts range from renditions and sketches to adversarially filtered photographs.

Cross-host evidence. Alongside the DINO host, PE provides a vision-language encoder family with its own base estimator, functional and actuator. With half of the residual-fitting data, its compact CTFM improves over the full-data feature-to-task procedure by 10.88 ResNet-50 points (59.19 vs 48.31, Tab. 3).

Existing regional and industrial tests. The frozen compiled map supports four- and sixteen-region ownership, improving all seven endpoints over identical-capacity feature matching without refitting (Tab. 16). On all 1,568 MVTec LOCO test images, the frozen PE map improves PSNR, LPIPS and cycle error over FFM-to-TLT without domain-specific fitting (Tab. 27).

Additional consumer evaluation. We added the frozen pooled-head test identified as an extension in App. H.12 of the submission. A linear 1,000-class head is trained only on archived pooled features of 50,000 ImageNet training images, then applied unchanged to every migrated field (Top-1 in %):

Field given to the headSet 1 ↑ImageNet-R ↑Sketch ↑Same decision as base ↑
Any IQ-Safe map with the calibrated gg3.342.283.35100%
IQ-Safe CTFM with real-pair g′g'70.2033.2242.83(new protected coordinate)
Ownership-free affine VFM0.060.060.050.1 to 0.2%
Archived field (reference)76.2438.7548.35

The original-gg rows test invariance: every IQ-Safe map keeps the base head's decision on all 10,432 inputs, as Proposition 1 requires, while the ownership-free map reduces accuracy to near chance. Replacing the pooled map with g′g', fitted on real training pairs, changes calibration while leaving the residuals frozen. The head then reaches 92%, 86% and 89% of its accuracy on archived fields (revised App. E.2).

Exact pooled-coordinate preservation covers every deterministic pooled head, including nonlinear MLPs. Protecting attention- or max-pooled statistics requires a separate ownership construction (App. H.12). The added shift and pooled-head tests provide further evidence within the two RAE release pairs studied.

References

[1] Maiorca et al. Latent Space Translation via Semantic Alignment. NeurIPS 2023.

[2] Mai et al. Revisiting Model Stitching In the Foundation Model Era. CVPR 2026.

[3] Hinton et al. Distilling the Knowledge in a Neural Network. arXiv:1503.02531.

[4] Saha et al. Gradient Projection Memory for Continual Learning. ICLR 2021.

[5] Fang et al. AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models. ICLR 2025.

[6] Shen et al. Towards Backward-Compatible Representation Learning. CVPR 2020.

[7] Vejendla. Drift-Adapter: A Practical Approach to Near Zero-Downtime Embedding Model Upgrades in Vector Databases. EMNLP 2025.

Reviewer_2fU2OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper tackles a problem of post-release feature migration in Representation Autoencoders (RAEs), where updated visual representations become incompatible with previously trained downstream consumers. It proposes Consumer-Typed Functional Migration (CTFM), which learns a lightweight release-wide mapping using archived consumer responses, explicitly separating behavioral recovery from coordinate preservation. Across DINO-based RAEv2 and Perception Encoder RAEs, CTFM consistently outperforms feature-matching baselines in recognition, reconstruction fidelity, semantic consistency, and distribution quality.

Critical strengths
  1. Methodological and Experimental Rigor
  • The ownership-preserving affine compilation is an interesting and well-motivated component. Rather than discarding a stronger affine alignment that violates the ownership constraint, the method reformulates it into a compact refinement that preserves the required coordinate structure while retaining much of its benefit.
  • The experimental setup is thoughtfully designed around two distinct RAE release pairs with different producer and consumer configurations. In particular, the use of heterogeneous migration settings and the separate evaluation of the compact-map regime provide a reasonable testbed for examining the core migration problem without conflating it with a single architecture or scaling choice.
  • The evaluation is extensive and generally well controlled. In particular, the principal RAEv2 comparison isolates response-based supervision using matched capacity, initialization, and optimization conditions.
Critical weaknesses
  1. Practicality of the Problem Setting
  • The paper defines a clear post-release compatibility problem and conducts generally sound, well-controlled experiments within this setting. However, the claimed motivation of enabling “practical reuse of dense RAE fields” appears substantially stronger than what is supported by the proposed deployment scenario. The setting becomes relevant only under the intersection of several restrictive assumptions: the producer has been upgraded, existing consumers must remain frozen, old producer components cannot be retained or routed to, archived model calls are disallowed. While each constraint may have precedent individually, the paper provides limited evidence that they naturally co-occur in realistic deployments. More importantly, it does not compare against several straightforward alternatives that could substantially weaken the need for migration.

  • First, the comparison against per-consumer routing is incomplete. Although the paper acknowledges routing as an alternative deployment regime, it does not establish when routing is infeasible or compare its cost-quality trade-off against CTFM. If legacy consumers cannot be updated, continuing to serve (GoldG_{\text{old}}) to them is also plausible. Moreover, because the evaluation is largely defined relative to (Dnew(Z)D_{\text{new}}(Z)), it remains unclear whether migration is preferable to retaining the original (Gold→DoldG_{\text{old}} \rightarrow D_{\text{old}}) pipeline.

  • Second, the decode-and-reencode alternative seems weakens the practical motivation for IQ-SAFE. The protected retrieval path seems substantially weaker than alternatives that decode the field and re-encode it with (EoldE_{\text{old}}), yet the paper does not quantify whether avoiding one additional archived-model call justifies this loss. While IQ-SAFE correctly preserves the calibrated-base retrieval behavior, exact invariance is only compelling when the protected coordinate remains sufficiently useful. If a simple re-encoding path yields much stronger retrieval, the motivation for sacrificing decoding quality relative to QUALITY becomes unclear. More broadly, the paper also leaves under-justified why generation and retrieval must directly consume the same dense field rather than decoded images or separately maintained representations.

Questions for authors

I encourage the authors to carefully address the critical concerns raised above during the rebuttal phase.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 2fU2 (1/2)

Thank you for emphasizing the need to compare shared-field migration with routing and decode-and-re-encode alternatives. We now quantify these alternatives and clarify the deployment conditions addressed by each.

Practicality. When do the shared-field and frozen-consumer constraints arise together, and how do alternatives compare?

Deployment conditions addressed by CTFM. Deployment applies the migration map followed by the retained consumers, including DoldD_{old}. The no-additional-call requirement applies to computing A(Z)A(Z). App. E.2 separately evaluates decode-and-re-encode routes.

An illustrative deployment stores or broadcasts one successor field shared by multiple retained readers. Srivastava et al. [1] show that a model update which raises accuracy can still introduce downstream errors. Scale-RAE [2] illustrates readers beyond decoding: "the multimodal model can directly reason over generated latents". Sec. 2.2 distinguishes this shared-field requirement from serving layers that can deliver different tensors to different readers.

Measured routing comparison. CTFM provides IQ-Safe for one shared field and Quality for separate consumer inputs. By Proposition 1, routing can send the same g(Q(Z))g(Q(Z)) to pooled readers while releasing the decoder's ownership constraint. We evaluate this option by training the same affine map with the same response loss without the projector. On prospective set 1 (3,232 states), using a linear 1,000-class head trained only on archived pooled features (Top-1 in %):

Serving optionResNet-50 ↑PSNR (dB) ↑LPIPS ↓Pooled head Top-1 ↑Same decision as base ↑
One IQ-Safe field, CTFM (affine)78.2219.170.2213.34100%
One Quality field, same loss80.0419.380.1930.060.2%
One field, ownership-free affine VFM79.6118.630.1980.060.2%
Routing: Quality field to decoder, g(Q(Z))g(Q(Z)) to pooled readers80.0419.380.1933.34100%

On this population, routing with CTFM's Quality map improves decoder-side ResNet-50 by about 1.8 points and PSNR by 0.2 dB over one IQ-Safe field. The decoded-gallery difference (+0.09) is not significant. A single ownership-free field, with either objective, reduces the pooled reader's accuracy to near chance. Routing preserves the pooled input separately and requires support for distinct consumer inputs. The pooled accuracies reflect the original calibration gg. Part 2 evaluates a replacement pooled map.

Retaining Gold→DoldG_{old} \rightarrow D_{old}. Retaining the old pipeline continues serving the old release. Our comparison holds successor outputs fixed and measures their execution by retained consumers. Neither evaluated graph contains a generator for the archived representation, so old-pipeline end-to-end performance is not measured here. Sec. 4.1 states this evaluation scope.

Evaluation relative to Dnew(Z)D_{new}(Z). A generated field has no paired real image, so paired metrics use Dnew(Z)D_{new}(Z). For real images the reference is the image itself (App. E), and the revised map improves over projected affine VFM in all 24 endpoint and domain comparisons on 25,383 images.

AuthorsOpenReview ↗

Response to Reviewer 2fU2 (2/2)

Decode and re-encode seems stronger, and exact invariance only matters if the protected coordinate stays useful.

With real-pair calibration, protected-coordinate retrieval has higher R@1 on both evaluated query sets and uses about 42.5% fewer rendering-plus-retrieval FLOPs than decoding and re-encoding the CTFM field (Tab. 26). We compare the following routes on the 7,200 real queries of Tab. 20 (Tab. 17 in the submission, same 1,000-centroid gallery). Compute is the total per query, rendering included, in FLOPs relative to one pass of the installed decoder:

Route to the installed galleryCompute ↓R@1 ImageNet-R ↑R@1 Sketch ↑
Protected coordinate, paper's pooled map gg1.0064.535.90
Protected coordinate, real-pair pooled map g′g'1.00626.2235.27
Decode and re-encode, calibrated base field1.743.563.40
Decode and re-encode, CTFM field1.7525.0029.62
Successor decoder, then EoldE_{old}1.7425.3730.65
Archived field Eold(x)E_{old}(x)needs original pixels25.8434.45

For the retained-decoder route, field repair substantially improves subsequent re-encoded retrieval: calibrated-base fields reach 3.6% and 3.4% on the two query sets. The table reports the successor-decoder route separately. The original pooled map gg was fitted on 768 generated calibration states.

Proposition 1 holds for any fixed pooled map. We therefore fit g′g' by orthogonal Procrustes on 50,000 ImageNet training pairs, selected on a training holdout, and replace the output mean without refitting the field maps. This belongs to the family of pooled translators that Drift-Adapter [3] studies for vector database upgrades. The ownership identity remains valid, allowing the new calibration to be combined with the frozen centered component. On generated states (prospective set 1):

FieldDirect pooled R@1 ↑ResNet-50 ↑PSNR (dB) ↑LPIPS ↓
CTFM (affine)7.6478.2219.170.221
CTFM (affine) with g′g'56.0080.1420.210.182
Archived field (reference)57.3382.0919.410.144

The protected coordinate now nearly matches archived-field retrieval on generated states and exceeds it on the real queries above, with no extra model call. Ownership stays exact (error below 10−710^{-7}), and every decoded endpoint on generated states improves as well. On the real queries, composing g′g' with the compiled map improves PSNR by 1.46 dB on ImageNet-R and 2.50 dB on Sketch (revised App. E.2).

Why must generation and retrieval consume the same dense field?

CTFM's Quality contract supports routing when distinct inputs can be delivered to different consumers, as evaluated in part 1/2. IQ-Safe serves retained readers of one shared stored or broadcast field. On the evaluated real-query sets, g′g' reaches archived-field retrieval performance while retaining exact preservation of the newly calibrated pooled coordinate.

The ownership-preserving compilation you highlighted remains in the paper with Proposition 2. At full rank it reduces exactly to projected affine VFM (App. A.5), so the new affine refinement retains the same consumer contract and ownership guarantee.

References

[1] Srivastava et al. An Empirical Analysis of Backward Compatibility in Machine Learning Systems. KDD 2020.

[2] Tong et al. Scaling Text-to-Image Diffusion Transformers with Representation Autoencoders. arXiv:2601.16208.

[3] Vejendla. Drift-Adapter: A Practical Approach to Near Zero-Downtime Embedding Model Upgrades in Vector Databases. EMNLP 2025.

I have carefully read the other reviewers' comments and the authors' rebuttal. I appreciate the authors' efforts.

However, my main concern about the practicality of the problem setup is still not fully resolved. I do not disagree with the idea of using semantic fields as shared interfaces. However, in the case of Scale-RAE mentioned by the authors, my understanding is that the sharable representation is possible because the same SigLIP-2 model representations are used. In this regard, I am still not fully convinced whether the main scenario considered in this paper, where different model releases need to share an interface, is a realistic and practically important setting. I think providing concrete examples of real-world systems or services operating under similar conditions would be the most direct way to address this concern.

The other two reviewers (PLZD and D7vh) also seem to have similar concerns about the practicality of the problem setup. Since this concern remains unresolved for me, I am now slightly more inclined toward rejection than before. Nevertheless, I decided to keep my original score of Weak Accept.

Thank you for continuing the discussion. LINE Part Time Jobs provides an operational precedent for cross-release representation compatibility [1]. Periodic retraining changes the two-tower producer's embedding coordinates, degrading downstream reranking when training and inference use different releases. LY Corporation applied post-processing alignment so producer and downstream models could update independently. The report includes offline comparisons and an online A/B test incorporating stabilized embeddings. LINE adopted the embedding-stabilization approach of Netflix researchers, whose RecSys study evaluates alignment across retraining cycles on proprietary recommendation data [2].

As you note, Scale-RAE’s consumers share the same SigLIP-2 representation within a release [3]. Published RAE development also includes changes in representation choice, as illustrated by RAEv2’s adoption of DINOv3 [4]. When existing consumers are retained across a representation change, they still expect the earlier coordinates. CTFM learns a field map to recover the retained decoder’s responses while preserving the calibrated pooled input under IQ-Safe.

The producer and retained consumers stay fixed while we fit the migration map. Archived models are available during offline fitting, and the retained decoder still executes at deployment. The restriction on additional model calls applies to computing the map. Maintaining the old producer remains an option for continuing the old release. IQ-Safe serves consumers of one shared field. Separate routing uses a Quality field for decoding and the calibrated input for pooled readers, with routing and decode-and-re-encode alternatives evaluated in Sec. 2.2 and App. E.2. The practical benefit is allowing a producer to advance without coordinating an upgrade of every dependent consumer.

References

[1] Kihara and Yuan. Embedding stabilization for search re-ranking: Overcoming cold start at LINE Part Time Jobs. LY Corporation Tech Blog, 2026.

[2] Zielnicki and Hsiao. Orthogonal Low Rank Embedding Stabilization. RecSys 2025.

[3] Tong et al. Scaling Text-to-Image Diffusion Transformers with Representation Autoencoders. arXiv:2601.16208.

[4] Singh et al. Improved Baselines with Representation Autoencoders. arXiv:2605.18324.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers

We thank the reviewers for highlighting two central questions: the value of response supervision and the deployment conditions for a shared field. We address both with additional experiments and a revised PDF (changes in blue). Table and section numbers refer to the revised PDF.

1. What does response supervision add beyond the simplest alternative under the same constraint? (PLZD Q3, qZvk W3, D7vh W3, W4, Q3, Q4)

Following PLZD's Q3, we trained a full affine map from projected affine VFM with the paper's response loss (Eq. 9), projector P0P_0, 160 fitting classes and 2,400 updates. CTFM (affine) improves over projected affine VFM at equal weight count on each evaluated population (recognition gains in pp, Indep. = mean of two independent classifiers).

Decoded Gallery denotes decoded-image Top-1. Direct pooled retrieval is reported separately.

PopulationStatesDecoded Gallery ↑ResNet-50 ↑Indep. ↑PSNR (dB) ↑
System set (Tab. 2)2,592+2.08+2.43+2.03+1.99
Prospective set 1 (Tab. 10)3,232+1.70+2.32+2.58+1.97
Prospective set 2 (Tab. 12)3,232+1.27+3.06+2.75+1.96
ImageNet-R (Tab. 23)3,200+1.47+2.25+2.47+2.15
ImageNet-Sketch4,000+0.97+2.15+1.44+4.46
ImageNet-A7,500+1.32+1.30+0.90+1.64
ImageNetV2-D10,683+1.60+3.22+3.23+1.61

Every gain in this table is statistically significant under the class bootstrap, and LPIPS and cycle error improve in every row.

The matched controls identify the additional contribution of consumer-response information. The compiled feature-matching control, retuned over the same seven learning rates, does not improve on its initialization under either tested selection rule. Continuing projected VFM on its own feature loss, selected by that loss, leaves ResNet-50 at 75.90 on prospective set 1. Response guidance improves recovery when used for optimization or checkpoint selection.

2. When must one migrated field serve every consumer? (2fU2's practicality concerns, D7vh W1 and Q1, PLZD W1 and Q1)

CTFM provides IQ-Safe for one stored or broadcast field shared by retained readers, and Quality (O=0O = 0) when a serving layer can route different inputs. The shared-field setting includes pooled readers unknown at migration time.

By Proposition 1, every IQ-Safe field carries g(Q(Z))g(Q(Z)), the vector a router can supply separately to pooled consumers. Quality then releases the decoder's ownership constraint.

A linear 1,000-class classifier trained only on archived pooled features makes the same decision under every IQ-Safe map built on the calibrated gg as under the base, on all 10,432 test inputs. The ownership-free affine VFM reduces its accuracy to 0.06%, near chance.

Pooled-coordinate usefulness also depends on calibration. Replacing gg with an orthogonal Procrustes map fit on 50,000 real training pairs, with the residual unchanged, lifts direct retrieval to the archived field's level (R@1 26.2 vs 25.8 on ImageNet-R, 35.3 vs 34.5 on Sketch) and improves every decoded endpoint on generated states.

3. How the revision relates to the submission

The revised manuscript retains the submitted results and reports the additional comparisons alongside them. On both prospective sets, the submitted CTFM improves its own initialization on all seven endpoints. It also outperforms feature matching at identical capacity, and on the system set it improves all five endpoints over projected affine VFM (Tab. 2).

CTFM (affine) refines the full-rank case of the submitted construction. Eq. (7) projects V(Z)−Ac(Z)V(Z)-A_c(Z) onto rr channel directions. r=432r=432 is the largest rank within the weight budget. At r=dr=d and α=1\alpha=1, ΠTd=P0\Pi_{\mathcal T_d}=P_0 and Q(Ac(Z))=g(Q(Z))Q(A_c(Z))=g(Q(Z)), so the initialization equals projected affine VFM (App. A.5). Refinement uses the same response loss and projector, preserving the contract and Proposition 1.

Comparison protocol and reproducibility. Matched controls share fitting states, update budgets and checkpoint grids. Rebuttal refinements, ours and the baselines', select learning rates and checkpoints on validation states only. Evaluations first reproduced the submitted per-state results exactly. Apps. D.2 and E.2 document protocols, tuning and seeds. We will release complete code, configurations and per-state results as open source to support reproduction.

Revised-paper locationMain revision
Secs. 2.2, 3.1, 3.3, 4.1, Fig. 2 and Alg. 1Deployment conditions, base/functional attribution and affine refinement
Tabs. 2, 3, 10 and App. D.2Initialization rows, matched affine comparisons, tuning controls and seed robustness
Tabs. 4, 23–26 and App. E.2Shift transfer, pooled-head preservation, real-pair calibration and serving-route costs

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.