acceptodds
Under review as a conference paper at ICLR 2027

AnchorMAE: Bridging Temporal Correspondence and Task Semantics through Frame-Level Semantic Transfer

Abstract

Ego–exo video learning must recognize activities across changing viewpoints without losing the temporal detail needed to distinguish action phases. Masked cross-view reconstruction provides dense temporal supervision, while clip-level language alignment alone does not directly specify task supervision for individual deployable visual frames. Prior methods also combine language and temporal prediction; the question examined here is how to couple dense visual reconstruction with same-video frame-level task supervision without requiring language at inference in the unpaired ego–exo setting. We introduce AnchorMAE, a shared-encoder dual-path framework with an unconditioned visual pathway and a training-only task-conditioned pathway. Dual-stage Semantic Anchor Adapters condition the latter; task-aware clip alignment and detached, same-video frame-level transfer supervise the former while masked self-/cross-view prediction retains pre-semantic visual targets. Only the visual pathway is used at inference. On AE2, AnchorMAE reports 0.8687 Phase F1, 0.8264 frame mAP@10, and 0.9966 Kendall's . Evaluations on AE2 and Ego-Exo4D, reported ablations and semantic/temporal diagnostics examine whether the proposed coupling supports both phase-sensitive correspondence and activity discrimination, while revealing remaining rank-dependent retrieval trade-offs.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Egocentric–Exocentric Video Representation Learning, Cross-View Alignment, Masked Video Modeling, Vision–Language Learning, Semantic–Temporal Correspondence, Self-Supervised Representation 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_gj1SOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper proposes AnchorMAE for ego–exo video representation learning, aiming to capture both task semantics and fine-grained temporal correspondence. The framework uses a visual pathway and a training-only task-conditioned pathway that share a temporal encoder. It combines masked visual reconstruction, clip-level alignment, and same-video frame-level semantic transfer. Only the visual pathway is used at inference. The method is evaluated on AE2 and Ego-Exo4D, with component ablations and representation diagnostics.

Critical strengths
  1. The paper raises a relevant question within ego–exo representation learning: how to incorporate task semantics while preserving fine-grained temporal information. It proposes a concrete framework for investigating this question and reports useful empirical improvements, supported by evaluations and component ablations.
Critical weaknesses
  1. The specific limitation of prior work is not clearly established. It is difficult to determine whether existing methods fail to capture both task identity and temporal phase, or whether they capture both but combine their supervision inadequately. Since the paper acknowledges prior approaches incorporating semantic and temporal learning, this distinction is essential to understanding what problem remains unresolved and what constitutes the proposed advance.

  2. The connection between this problem and the architectural design is insufficiently explained. The two pathways differ in access to task text; they do not explicitly separate task identity from temporal phase. It remains unclear why transferring task-conditioned frame representations through this structure should achieve the desired balance. In particular, the paper could better explain how semantic transfer preserves phase distinctions rather than encouraging similarity across frames of the same activity. The ablations report benefits from the proposed components, but do not sufficiently clarify the reasoning that connects the design to the motivating limitation.

Questions for authors
  1. Would it be possible to include qualitative examples like Figure 3 for a wider range of labels in the main paper? Such examples, with more visible temporal progression cues, would help readers see the motivating problem and how the framework distinguishes task separation, temporal structure, and cross-view alignment.

  2. If visual-only inference is one of the contributions, could you indicate in the results tables which compared methods support visual-only inference and which require additional inputs at inference time? This would help clarify how much of the reported advantage comes from the framework itself rather than from differences in the inference setting.

3 comments
AuthorsOpenReview ↗

Response to Reviewer gj1S

Thank you for identifying that the original paper did not make the problem-to-design connection sufficiently explicit. We revised the framing, mechanism explanation, diagnostics, and qualitative evidence around task identity versus temporal phase.

W1 — What remains unresolved relative to prior work? The revision no longer suggests that prior methods lack temporal learning or semantics. Instead, the Introduction and Related Work explicitly acknowledge BYOV, MILES/TGM, VRS, PRISM, and feature distillation. Figure 1 now summarizes the narrower remaining interface question: how can task-level semantics supervise individual deployable visual frames while dense reconstruction remains pre-semantic, synchronized ego–exo frame matching is not assumed, and text is confined to training?

W2 — Why should the architecture address this limitation? We now state explicitly that AnchorMAE does not mathematically disentangle task identity and temporal phase. The visual path retains masked reconstruction against detached pre-semantic targets, while the training-only semantic path supplies task semantics; detached same-video frame transfer connects the two at matching frame indices within each video. Figure 2 shows this separation directly. Table 6 tests transfer removal, reconstruction-target type, and negative handling; Table 9 separately measures task/semantic and temporal/viewpoint diagnostics. We retain the modest temporal-distractor result (75.04% versus 74.97% Recon-only) rather than overclaiming it.

Q1 — Wider labels and visible temporal progression. We added a dedicated main-paper qualitative analysis in Fig. 3. It uses three Tennis Forehand ego queries from Early (Backswing), Mid (Hit), and Late (Front Swing) phases, each with its actual Top-3 ego->exo retrievals. The activity and retrieval direction remain fixed while the query phase changes, making the progression directly visible. Table 8 reports all four AE2 activities separately for Recon+Align and AnchorMAE. Supplement Fig. 7 further broadens the qualitative analysis to Break Eggs, Pour Liquid, Tennis Forehand, and Pour Milk in both ego->exo and exo->ego directions. Supplement Figs. 4–6 provide the model-comparison, embedding, and failure analyses.

Q2 — Visual-only inference. We revised this claim to avoid implying that competing representation-learning methods necessarily require text at inference. For AnchorMAE, visual-only inference is stated only as a deployment property: the text encoder, SAA modules, semantic forward pass, masked decoders, and training losses are removed, leaving the unconditioned visual pathway and downstream heads. This is stated in Sec. 3, illustrated in Fig. 2, and reiterated in the Ego-Exo4D Table 3 caption.

Together, Figs. 1–3, Tables 5–9, and Supplement Figs. 4–7 make the intended gap -> mechanism -> evidence chain explicit while keeping the qualitative claims appropriately scoped.

Reviewer_gj1SOpenReview ↗

Follow-up assessment

Thank you for the revisions. The explanation of how frame-level semantic transfer works alongside visual reconstruction clarifies the design. The revised Figure 3 addresses my question about temporal progression, and I appreciate the clarification that visual-only inference is a deployment property of AnchorMAE.

My remaining concern is that the paper does not yet clearly show where prior methods’ clip-level supervision falls short and how frame-level transfer addresses that limitation. Most of my initial concerns have been addressed, so I am raising my rating from 1 to 2 (Weak reject).

AuthorsOpenReview ↗

Follow-up on Clip-Level vs Frame-Level Supervision

Thank you for the follow-up and for raising the rating. We strengthened the remaining clip-level-vs-frame-level supervision point both conceptually and experimentally.

The revision now states explicitly that clip-level alignment constrains only the pooled video representation, whereas same-video transfer supervises each visual token from its task-conditioned counterpart at the same frame index.

Main-paper Table 6 tests this directly: clip-level-only/no-transfer gives 0.8096/0.7826/0.9072/0.9211 (F1/mAP@10/R²/τ), pooled broadcast gives 0.8145/0.7894/0.8993/0.9747, shuffled-index transfer drops to 0.7984/0.7554/0.8881/0.6947, and same-index transfer is strongest on all four metrics at 0.8687/0.8264/0.9247/0.9966.

This directly separates exposure to task semantics from preservation of frame-specific correspondence.

Reviewer_ojh5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3(+1)
Soundness
2
Read review
Paper summary

AnchorMAE combines dense reconstruction with task semantics for unpaired ego-exo video representation learning. A shared temporal encoder processes visual and training-only task-conditioned pathways, with Semantic Anchor Adapters before and after encoding. Detached semantic states supervise corresponding frames within each video, while reconstruction retains pre-semantic visual targets. Task-aware contrastive alignment excludes non-paired same-task negatives. Inference uses only visual features. Evaluation covers AE2 temporal correspondence and Ego-Exo4D retrieval/recognition, with mechanism ablations and semantic/temporal diagnostics.

Critical strengths
  1. This paper enables practical training and deployment setup. Same-video frame transfer removes the need for synchronized cross-view recordings, and deployment requires no text prompt.
  2. Targeted ablations support the method design. Removing frame transfer lowers F1 from 0.8687 to 0.8096. Supplementary comparisons of target type, placement, anchor content, gating, and conditioning provide further support.
Critical weaknesses
  1. Missing experimental details limit reproducibility. The paper does not specify the encoder checkpoints, temporal architecture, training schedule, loss weights, or exact evaluation splits. MSM token restoration and causal-attention masks are also unclear. A configuration table covering the main model and shared ablation settings would help resolve these gaps.
  2. Limited uncertainty and per-activity reporting. The paper reports aggregate scores but omits training-run counts and variability. Temporal-distractor accuracy improves by only 0.07 percentage points over reconstruction alone, from 74.97% to 75.04%, leaving the reliability of this gain unclear. More complete per-activity results would help establish how broadly the improvements hold.
Questions for authors
  1. Please see the weakness part.
  2. (minor question) It seems that the authors have reuploaded the manuscript (with appendix) as supplementary materials. Could the authors confirm that all materials related to this paper have been properly uploaded.
1 comment
AuthorsOpenReview ↗

Response to Reviewer ojh5

Thank you for highlighting the reproducibility, uncertainty, and supplementary-upload issues. We substantially expanded the implementation/evaluation description and added repeated-run and per-activity evidence.

Implementation / reproducibility. Table 2 now summarizes the core configuration: frozen CLIP ViT-B/16 and frozen Sentence-BERT encoders, patch selection → pooling → 256-D projection, a 12-block/4-head temporal encoder, T=32 frames, MSM/MCM mask ratios 0.40/0.80, optimizer/training settings, loss weights, temperature, and seeds. Appendix Table 11 gives the detailed configuration, including exact checkpoint names, projection dimensions, decoder depth, optimizer parameters, final-epoch checkpoint selection, and seeds 42/123/2024.

We also clarified the masking behavior requested in the review: masked tokens are removed before the temporal encoder and a learned mask token is restored before decoding; the encoder over visible tokens is bidirectional, the MSM decoder is causal, and the MCM decoder is bidirectional. We therefore no longer make an ambiguous end-to-end “past-only” claim. The frame-transfer coefficient is explicitly λ_ST = 1.0.

Exact evaluation protocols. For AE2, AnchorMAE is trained jointly across the four activities while respecting the released train/validation/test partitions, and the headline results are final held-out test-set macro-averages. For Ego-Exo4D, we follow the ViewpointRosetta cross-view benchmark/evaluation protocol: 188-class bidirectional retrieval and exo-train→ego-test recognition. The revision explicitly separates these benchmark choices from AnchorMAE-specific 32-frame sampling/training hyperparameters (Sec. 4 and Appendix A.2).

Run-to-run variability and per-activity results. Table 8 reports mean ± std over seeds 42/123/2024. Recon+Align obtains 0.8123±0.0020 F1, 0.7989±0.0030 mAP@10, 0.9010±0.0017 R², and 0.8365±0.0217 τ; AnchorMAE obtains 0.8707±0.0038, 0.8185±0.0097, 0.9111±0.0125, and 0.9954±0.0025. Table 9 now reports all four AE2 activities separately for the submitted reference run (seed 42); its macro averages reproduce the submitted Recon+Align and AnchorMAE values exactly. We also retain the limitation that the temporal-distractor gain over Recon-only is only 0.07 percentage points rather than presenting it as a large temporal improvement.

For transparency, Appendix A.1 and the Reproducibility Statement also state that per-run effective-negative counts and detailed batch manifests are not available; we do not infer those values.

Supplementary upload. Thank you for flagging the duplicate upload. We corrected the organization: the revised main PDF contains the paper and references, while the separate supplementary file contains Appendix A.1–A.4 only. No experimental material is omitted or duplicated across the two uploads.

We appreciate this feedback; it led to a substantially more reproducible and auditable revision.

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

The paper presents AnchorMAE, a dual-path framework for learning fine-grained ego-exo video representations that combine temporal correspondence with activity-level semantic supervision. Building on masked self-view and cross-view reconstruction, the method introduces a training-only task-conditioned pathway using pre- and post-encoder Semantic Anchor Adapters. This path shares temporal encoder weights with an unconditioned visual pathway, whose frame representations are supervised through detached same video semantic transfer. The method additionally combines task-aware clip alignment with visual reconstruction using pre-semantic targets, allowing the visual path to work without task descriptions at inference. Experiments on AE2 and EgoExo4D report improvements in phase classification, frame retrieval, temporal ordering, and cross-view recognition.

Critical strengths

[S1] A meaningful approach to combining temporal and semantic supervision. AnchorMAE addresses an important limitation of reconstruction-oriented ego-exo representation learning. While visual correspondence objectives do not necessarily provide strong discrimination between semantically different activities, the proposed training-only semantic pathway and detached frame-level transfer offer a concrete way to inject task supervision into deployable visual representations.

[S2] Ablations and complementary evaluation tasks. The paper evaluates both phase-sensitive correspondence on AE2 and activity-level transfer on EgoExo4D. Beyond aggregate benchmark comparisons, it studies reconstruction-only and alignment-only objectives, frame-level semantic transfer, reconstruction target choices, and contrastive negative handling. The task/view probes and semantic/temporal distractor evaluations are useful attempts to distinguish activity discrimination from temporal correspondence. The reported improvements in task discrimination are particularly encouraging.

Critical weaknesses

[W1] Novelty. The proposed framework combines several established techniques, including masked self-/cross-view reconstruction, language-conditioned feature learning, contrastive alignment, and teacher-student feature transfer. Although their particular integration is distinguishable from prior methods, the manuscript may not sufficiently establish what new scientific findings or intuitions emerge from this combination. In particular, the shared-encoder dual-path architecture and detached frame-level alignment appear to be relatively straightforward extensions of existing representation learning strategies. The distinction from closely related works, such as BYOV, MILES, and PRISM, is primarily explained through architectural differences rather than a deeper characterization of the underlying learning mechanism. Consequently, while the method is technically reasonable, its methodological contribution appears somewhat incremental, and the empirical evidence does not yet demonstrate why this particular integration is preferable to simpler alternatives.

[W2] Limited comparability. The improvements over previous approaches are encouraging, but their interpretation is complicated by differences in training supervision, model initialization, and potentially evaluation configurations. In particular, AnchorMAE uses explicit task-level textual supervision, whereas reconstruction-oriented approaches such as AE2 and BYOV do not necessarily have access to equivalent semantic information. The reported improvements therefore cannot be attributed solely to the proposed architecture or frame-level transfer mechanism. The manuscript acknowledges these differences, but the comparisons remain an important part of its empirical motivation. A more convincing evaluation would compare the proposed approach with existing methods under comparable settings, including backbones, training data, supervision, and evaluation settings.

Questions for authors

[Q1] The proposed framework combines masked reconstruction, task-aware alignment, semantic adapters, and frame-level feature transfer. Could the authors clarify which aspect of this combination constitutes the principal methodological contribution beyond existing language-guided video representation learning? In particular, how does the proposed frame-level transfer differ in its learning effect from conventional stop-gradient feature distillation between task-conditioned and unconditioned representations? A comparison with a simpler semantic-distillation baseline would be helpful in establishing the necessity of the proposed design.

[Q2] Could the authors provide additional comparisons with BYOV or other closely related methods using the same visual backbone, training data, and task-level semantic supervision? Since AnchorMAE introduces explicit task descriptions that may not be available to the compared reconstruction-based baselines, it is difficult to determine whether the observed improvements originate from the proposed architecture or from additional supervision. A controlled comparison would help clarify the significance of the proposed methodological contribution.

[Q3] Could the authors provide separate ablations for SAA-Pre, SAA-Post, and their combination while keeping the frame-level transfer objective unchanged? Additionally, how does performance change when replacing the dual-stage adapters with a simpler task-conditioned projection or direct feature-distillation objective? These experiments would clarify whether the proposed adapter configuration provides complementary benefits and whether both stages are necessary.

3 comments
AuthorsOpenReview ↗

Response to Reviewer fkKr

Thank you for the detailed review. We agree that the original submission did not isolate the contribution sharply enough from existing masked prediction, language supervision, and feature distillation. The revision therefore narrows the claim and adds stricter controls.

W1 / Q1 — Principal contribution and relation to conventional distillation. We no longer claim novelty from simply combining temporal learning and language, nor do we claim a new distillation operator. The revised Introduction/Related Work explicitly acknowledge that masked visual prediction, language-guided learning, semantic cross-view alignment, and feature distillation are established. Figure 1 now gives a compact supervision-interface comparison: the remaining question is how task semantics can reach individual deployable visual frames while reconstruction remains pre-semantic, ego–exo frame synchronization is not assumed, and text is confined to training. Section 3 then implements this through a shared visual/semantic encoder, training-only semantic anchoring, and detached same-video frame transfer.

To test whether ordinary task-conditioned distillation is sufficient, Table 5 adds a direct semantic distillation (DSD) control under the same backbone, semantic supervision, optimizer, and evaluator. DSD retains task text and the same detached frame-transfer objective but removes SAA-Pre/Post. It obtains F1/mAP@10/R2/tau = 0.8138/0.7945/0.9045/0.8957, versus 0.8687/0.8264/0.9247/0.9966 for AnchorMAE. We therefore position the contribution as the specific supervision interface/coupling, not as a new primitive.

W2 / Q2 — Backbone/supervision comparability and BYOV. We added a faithful rerun of the official BYOV implementation under our AE2 evaluator (0.7666 F1, 0.7948 mAP@10, 0.9015 R2, 0.7523 tau). We explicitly treat this as contextual rather than a one-factor architectural control because it retains BYOV-specific architecture/training choices. The stricter comparison is the common-pipeline chain Recon-only -> Recon+Align -> DSD -> AnchorMAE in Tables 4–5, which holds the AnchorMAE feature pipeline, optimizer, schedule, evaluator, and AE2 split fixed while varying access to task text, frame transfer, and SAA.

Q3 — Is dual-stage SAA necessary, and are simpler conditioning alternatives sufficient? Table 5 directly compares Recon+Align, DSD, SAA-Pre only, SAA-Post only, and full Pre+Post AnchorMAE with frame transfer held identical wherever enabled. The single-stage variants are metric-dependent; the full Pre+Post setting is the only row strongest across all four AE2 metrics. Supplement Table 13 additionally compares Addition, Concatenation, FiLM, and SAA, while Supplement Table 12 tests fixed versus learnable residual gating.

We also added the exact implementation configuration (Table 1; Supplement Table 10), three-seed variability (Table 7), and per-activity results (Table 8). The revised paper therefore makes a deliberately narrower claim: AnchorMAE is a specific supervision interface with controlled empirical support, not a claim that its individual ingredients are themselves new.

I would like to thank the authors for the detailed responses, manuscript revisions, and additional control experiments provided during the rebuttal period.

Several of my initial questions regarding internal ablations and implementation details have been clarified. In particular, adding the direct semantic distillation (DSD) control, separating SAA-Pre and SAA-Post alongside simpler conditioning alternatives, and reporting three-seed variability, per-activity AE2 scores, and exact encoder and masking configurations improve the internal empirical support and reproducibility of the paper. Narrowing the framing to focus on the supervision interface rather than claiming individual components as novel also makes the positioning more accurate.

At the same time, my primary concerns regarding methodological novelty and controlled comparability remain unresolved. Positioning AnchorMAE as a specific supervision interface confirms that the framework is largely an incremental integration of established masked cross-view prediction, adapter-based text conditioning, contrastive alignment, and stop-gradient feature transfer, without establishing deeper new insights into why this coupling preserves phase granularity better than alternatives, especially given that temporal-distractor accuracy improves by only 0.07 percentage points over reconstruction alone (75.04% vs. 74.97%). Furthermore, regarding [Q2], while rerunning the official BYOV implementation under the AE2 evaluator provides helpful context, it retains BYOV's original architecture and visual-only supervision rather than comparing against BYOV or another closely related baseline under matched visual backbones (CLIP ViT-B/16) and equivalent task-text supervision. The comparisons on AE2 and Ego-Exo4D (where HR@1 also trails VRS at 0.1504 vs. 0.2231) therefore remain confounded by supervision and backbone differences.

Taking everything into account, while the revisions strengthen the ablations and clarity of the manuscript, I still view the overall conceptual advance and controlled validation as insufficient for acceptance, and I will maintain my initial score.

AuthorsOpenReview ↗

Follow-up on Mechanistic Evidence and Matched Supervision

Thank you for the follow-up and for clarifying the two remaining concerns: mechanistic evidence for frame-sensitive supervision and controlled comparability to BYOV.

We added a new semantic-supervision granularity control in main-paper Table 6. Clip-level-only/no-transfer gives 0.8096/0.7826/0.9072/0.9211 (F1/mAP@10/R²/τ); pooled semantic broadcast gives 0.8145/0.7894/0.8993/0.9747; shuffled-index same-video transfer falls to 0.7984/0.7554/0.8881/0.6947; and correct same-index transfer is strongest on all four metrics at 0.8687/0.8264/0.9247/0.9966. This separates simply exposing frames to task semantics from preserving frame-index correspondence.

We also added Supplement Table 14 with a matched BYOV-style supervision comparison under the common frozen CLIP ViT-B/16 frame pipeline and evaluation setup. Matched BYOV, BYOV + clip-level text, DSD, and AnchorMAE progressively separate task-text access, frame-level transfer, and the full anchoring interface; AnchorMAE is strongest across all four metrics.

We hope these two new controls more directly address the remaining mechanism and comparability concerns.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Revision Summary

Thank you to all reviewers for the constructive feedback. We substantially revised the manuscript and supplementary material during the discussion period.

The final revision makes six main changes:

  • Sharper contribution framing. Fig. 1 and the Introduction now frame AnchorMAE as a supervision-interface question rather than claiming masked prediction, language supervision, attention, or feature distillation are individually new.

  • Direct semantic-interface controls. Table 5 compares Recon+Align, direct semantic distillation, SAA-Pre only, SAA-Post only, and the full Pre+Post configuration under the common pipeline.

  • New semantic-supervision granularity evidence. Table 6 now compares clip-level-only supervision, pooled semantic broadcast, shuffled-index same-video transfer, and correct same-index transfer. Shuffling the semantic frame targets degrades all four AE2 metrics, while correct same-index transfer is strongest on all four, directly testing whether frame-specific correspondence matters beyond simply exposing frames to task semantics.

  • New matched BYOV-style supervision control. Supplement Table 14 compares Matched BYOV, Matched BYOV + clip-level text, DSD, and AnchorMAE under the common frozen CLIP ViT-B/16 frame pipeline, AE2 split, sampling, optimization schedule, and evaluator. This separates task-text access, frame transfer, and the full anchoring interface while keeping the faithful official-BYOV rerun as a separate contextual reproduction.

  • Expanded reproducibility and uncertainty reporting. Table 1 and Supplement Table 10 provide the implementation configuration; Table 7 reports three-seed variability; Table 8 reports per-activity AE2 results; and the supplement specifies masking/restoration, causality, loss weights, seeds, and evaluation protocols.

  • More careful quantitative and qualitative analysis. Table 9 separates task/semantic and temporal/viewpoint diagnostics and retains the modest temporal-distractor result without overclaiming it. Fig. 3 shows early→mid→late temporal progression within a fixed Tennis Forehand activity, while Supplement Fig. 7 broadens retrieval examples across all four AE2 activities and both directions; Supplement Figs. 4–6 retain the model-comparison, embedding, and failure analyses.

The revised paper therefore makes a deliberately narrow claim: AnchorMAE studies a specific interface for transferring task semantics to deployable visual frames while retaining pre-semantic visual reconstruction targets, avoiding synchronized ego–exo frame assumptions, and requiring no task text at inference.

Detailed point-by-point responses and follow-ups are provided under the individual reviewer threads.

AuthorsOpenReview ↗

Follow-up on Remaining Contribution Concern

Thank you for the follow-up and for confirming that the added implementation details, repeated-run analysis, and design controls address the reproducibility and empirical-support concerns.

We understand that the remaining reservation concerns the significance of the methodological contribution. The final revision further clarifies this point in Fig. 1 and the Introduction: we explicitly acknowledge that masked reconstruction, semantic conditioning, feature distillation, and attention mechanisms are established. The claimed contribution is narrower—the supervision interface that transfers task semantics to deployable frame representations while retaining pre-semantic reconstruction targets and removing the semantic pathway at inference.

Table 5 then tests whether this reduces to a straightforward combination: direct semantic distillation and single-stage SAA controls do not reproduce the full model across the reported AE2 metrics.

If there is a specific remaining control or analysis that you believe would materially distinguish this interface from a straightforward combination of established components, we would appreciate knowing it. This would help us understand whether the remaining concern is empirically addressable or instead reflects a difference in the assessment of conceptual significance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.