acceptodds
Under review as a conference paper at ICLR 2027

LoAD: Low-Rank Distributional Alignment and Test-Time Decision Adaptation for Zero-Shot Skeleton Action Recognition

Abstract

Generative methods for zero-shot skeleton action recognition mostly follow an align-then-classify paradigm, where a cross-modal autoencoder aligns skeleton sequences with class semantics in a shared latent space, upon which a classifier is built. This paper revisits both phases. On the alignment side, class-conditional latent distributions are typically modeled as diagonal Gaussians, leaving the correlation structure within each distribution unexploited. On the decision side, different decisions usually read the same generative score, whereas they in fact impose different requirements on the evidence. Accordingly, we propose LoAD, which improves this paradigm in both alignment depth and decision structure. First, Low-Rank Covariance Alignment (LRCA) extends the latent distribution to a diagonal-plus-low-rank structure, and jointly aligns the means, variances, and low-rank correlation factors of both modalities during training. Second, Test-time Decision Adaptation (TDA) refines how decisions are made at test time, with two components. Decision-Decoupled Inference (DDI) constructs two complementary per-class evidences and reads them in a decoupled manner, where class ranking in the zero-shot setting fuses a flow-matching velocity error with a deterministic reconstruction error, while seen/unseen discrimination in the generalized setting relies on the former only. Test-time Shrinkage Calibration (TSC) further adapts the class latents with statistics from unlabeled test batches, modulated by the number of seen classes and requiring no gradient updates. On standard splits of NTU-60, NTU-120, and PKU-MMD, LoAD achieves competitive performance under both zero-shot and generalized zero-shot settings, and systematic diagnostic studies verify the effectiveness of each design.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Zero-shot learning, Skeleton-based action recognition, Cross-modal latent alignment, Low-rank covariance modeling, Test-time adaptation.

Reviews

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

This paper proposes a new method for zero-shot skeleton action recognition. Specifically, the overall framework is based on Flora. First, it incorporates low-rank covariance estimation for cross-modal alignment. Furthermore, it designs a test-time adaptation technique to calibrate unseen action-class centers. Experimental results show that the proposed method achieves competitive performance.

Critical strengths
  1. The introduction of low-rank covariance estimation for cross-modal alignment is appreciated, as it models the intra-class variance of skeleton features at a fine-grained level, which is beneficial for more precise alignment.
  2. The paper is well organized. Meanwhile, the experiments are relatively thorough and comprehensive, and the proposed method achieves competitive performance.
Critical weaknesses
  1. The motivation for incorporating LRCA is not well explained, and there is insufficient evidence or justification to demonstrate its effectiveness beyond recognition accuracy. Meanwhile, the performance gains brought by this design are limited. A more thorough analysis at the feature level would help justify the necessity of this module.
  2. The test-time adaptation requires at least a batch of unseen samples to estimate class means and variances. However, during inference, each unseen action sample should ideally be recognized independently rather than relying on a batch of test data. Thus, I believe the assumption underlying this inference setting is too strong and may not be suitable for standard inductive zero-shot learning.
  3. The proposed framework, including both training and ZSL/GZSL inference, is largely based on Flora, while the newly introduced designs are relatively incremental, raising concerns about the technical contributions and novelty of this work.
Questions for authors
  1. Why are flow-matching errors computed over a grid? Flora has demonstrated that single-step inference is sufficient. However, the motivation and benefits of this grid-based design are not clearly explained.
  2. It is unclear why cross-reconstruction errors are incorporated into the flow-matching errors. Meanwhile, the two types of errors may have different scales, raising concerns about the appropriateness of directly adding them together.
  3. The GZSL inference framework is derived from Flora, so an appropriate citation should be provided for Eq (12).
1 comment
AuthorsOpenReview ↗

Response to Reviewer WQRm

We thank the reviewer for the constructive feedback. W and Q denote the numbered weaknesses and questions in the review.

1. Why is LRCA necessary, and what does it learn? (W1). Flora's diagonal Gaussian conditionals cannot represent dependencies between latent dimensions. LRCA introduces the low-rank covariance component LL⊤LL^\top to model these correlations and align them across skeleton and semantic representations.

The effectiveness is supported by both recognition and structural evidence.

  • Recognition: LRCA alone improves reproduced Flora on all four NTU splits (+1.0/+1.0/+1.9/+0.8 pp, Table 3). Under matched inductive inference, a separate three-seed experiment also improves ZSL from 59.42% to 61.11% (+1.69 pp) on 48/12.
  • Capacity control: With matched active projection-head dimensions, LRCA outperforms an ordinary projection baseline on all four splits, including 64.65% versus 59.14% on 48/12.
  • Structural alignment: With the low-rank architecture held fixed, factor alignment reduces the induced cross-modal covariance discrepancy from 0.1592 to 0.0638, approximately 60%, in a 128-sample diagnostic.

These results show that LRCA contributes both improved recognition and better alignment of the learned cross-dimensional covariance structure, beyond an additional projection head.

2. Does TSC impose an unrealistic batch requirement? (W2). No requirement is imposed on the core ZSL model. LRCA and DDI classify individual queries. TSC is an optional enhancement for deployments where an unlabeled query collection is available, not a prerequisite for zero-shot recognition. Its value is measurable when that information exists: holding the full pipeline fixed, Table 5 shows TSC raises 48/12 ZSL from 62.45% to 66.06% and GZSL-H from 58.36% to 59.68%. The TSC-off LRCA-only model independently outperforms reproduced Flora (64.7% versus 63.7%, Table 3). A separate matched-access three-seed control also favors low-rank over diagonal without TSC (61.11% versus 59.42%). This directly answers the deployment concern: batch access is a condition for using the optional TSC enhancement, not a requirement to run LoAD. The same trained model remains usable for independent, single-query ZSL inference.

3. What advances beyond Flora? (W3). LoAD uses the Flora flow predictor but changes what the model aligns (correlated class-conditionals) and how it makes decisions (class-conditioned reconstruction for ZSL; flow evidence for GZSL). This is supported by the LRCA-only gain and structural controls above, rather than only a final-model score. Reconstruction fusion also helps under one fixed λ=0.1\lambda=0.1, not just split-specific weights (see Q2). TSC contributes a separate optional inference improvement.

4. Is multi-time flow scoring necessary? (Q1). No. Our reported benchmark results use single-time scoring at t=0.5t=0.5, consistent with Flora's single-step inference. The time-grid formulation is a generalization included for analysis, not a requirement of LoAD.

We additionally evaluated multi-time scoring independently, using the same fixed checkpoint without TSC. It changes ZSL accuracy from 58.96% to 59.02% on 48/12. Thus, LoAD's reported improvements do not depend on multi-time scoring or additional ODE integration.

5. Why combine errors of different scales? (Q2). FM tests velocity consistency; deterministic reconstruction tests whether the class semantics explain the observed skeleton feature. They are combined as Efm+λEvaeE^{\rm fm}+\lambda E^{\rm vae}, where λ\lambda calibrates their influence. Rather than requiring split-wise tuning, one fixed λ=0.1\lambda=0.1 yields:

NTU ZSL gain over FM only55/548/12110/1096/24
FM + reconstruction (pp)+0.89+0.61+0.97+1.03

On 48/12, fusion corrects 33 FM-only errors versus 13 newly incorrect predictions. The consistent net benefit supports the scores' complementary evidence, despite their different raw scales.

6. GZSL attribution (Q3). The GZSL seen/unseen ratio gate in Eq. (12) is inherited from Flora; we will cite the source alongside the equation. Reconstruction is used for ZSL class ranking; the GZSL gate and class scores remain FM-only.

Reviewer_hCH6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes LoAD, an extension of Flora for zero-shot skeleton action recognition. Low-Rank Covariance Alignment (LRCA) adds and aligns low-rank covariance factors in the skeleton and text distributions. Test-time Decision Adaptation (TDA) calibrates class representations using unlabeled test data and uses different scoring rules for ZSL and GZSL, without updating network weights. The method is evaluated on NTU RGB+D 60, NTU RGB+D 120, and PKU-MMD.

Critical strengths
  1. The paper identifies clear limitations of diagonal distribution alignment and test-time scoring, and proposes reasonable solutions to these problems.

  2. The paper provides detailed experiments, including component ablations, parameter sensitivity, and per-class error analysis. These help explain when the proposed components are useful.

Critical weaknesses
  1. The novelty over Flora needs stronger support. LoAD keeps the main flow-matching framework of Flora (CVPR 2026 Findings) and adds low-rank alignment, test-time calibration, and score fusion. These changes are reasonable, but the technical contribution appears incremental. The paper needs to explain more clearly what these additions contribute beyond the existing Flora framework.

  2. The gains are modest in GZSL, and the test-time comparison needs clarification. LoAD consistently improves ZSL accuracy over the reproduced Flora baseline, but the GZSL gains are small. For example, on NTU-120 96/24, H only increases from 52.6% to 52.9%. Repeated runs would help establish whether these small differences are stable. TSC uses statistics from the whole unlabeled test set to adjust class representations. This may give LoAD an unfair advantage over baselines that do not use test-set statistics for adaptation. On NTU-60 48/12, LoAD without TSC reaches 62.45% ZSL accuracy, below the reproduced Flora result of 63.7%. The authors should clarify which baselines use test-time adaptation and provide comparisons under the same test-data access conditions.

Questions for authors
  1. Could the authors clarify whether the compared baselines also use the unlabeled test set for adaptation? If not, how do the authors ensure a fair comparison with TSC?

  2. Could the authors expand the related work and explain more clearly how LoAD differs from Flora and other closely related methods?

  3. Could the authors define the notation more clearly when it first appears and use consistent symbols throughout the paper?

1 comment
AuthorsOpenReview ↗

Response to Reviewer hCH6

We thank the reviewer for the careful assessment. W and Q denote the numbered weaknesses and questions in the review.

1. Contribution beyond Flora and related work (W1, Q2). LoAD retains Flora's flow backbone but changes the aligned distribution and the decision evidence. LRCA learns and aligns cross-dimensional covariance factors between skeleton and semantic class-conditionals; this is not equivalent to widening a layer. DDI adds deterministic semantic-reconstruction evidence for ZSL class ranking, while retaining FM evidence for GZSL gating. TSC provides an optional, gradient-free calibration of class latents. In contrast, T3A adjusts classifier templates and SkeletonCache performs retrieval, rather than aligning generative class-conditionals.

LRCA alone, without test-set calibration, improves reproduced Flora on all four NTU splits by +1.0/+1.0/+1.9/+0.8 pp (Table 3). A matched-active-size ordinary projection is inferior to covariance modeling on each split; one fixed fusion weight improves all four splits independently. Full-model ZSL gains over Flora are +1.8/+2.4/+1.1/+1.9 pp (Table 1) and +3.3 pp on PKU-MMD (Table 2).

2. Fairness: separating representation gains from calibration (W2, Q1). LRCA's improvement is present even when the baseline and LoAD have identical inference access. Table 3 shows 64.7% for LRCA-only versus 63.7% for reproduced Flora on NTU 48/12, with no TSC in either model. The reviewer's 62.45% refers instead to LRCA + DDI without TSC. A further three-seed, fixed-endpoint check compares diagonal and low-rank modeling under matched inference settings:

NTU 48/12 ZSLDiagonalLow-rankGain
TSC off59.42%61.11%+1.69 pp
TSC on (same batches of 64)59.43%61.66%+2.23 pp

The reproduced Flora baseline does not use TSC; we therefore substantiate the LRCA claim with TSC-off and matched-access experiments, not the full-model score alone. Separately, toggling TSC within the same full model raises ZSL from 62.45% to 66.06% (Table 5), measuring calibration's additional contribution. Thus, covariance modeling improves under standard independent inference, and TSC provides an independently ablated enhancement when unlabeled queries are available. The additional controls use different fixed-endpoint checkpoints from Table 3.

3. GZSL improvements and repeated-run evidence (W2). The GZSL benefit is particularly evident in unseen-class recognition. In Table 1, LoAD improves NTU-120 unseen accuracy from 68.9% to 71.8% (+2.9 pp) on 110/10 and from 51.6% to 54.6% (+3.0 pp) on 96/24. These improvements are important because GZSL must recognize unseen classes while competing with seen-class predictions; the harmonic mean additionally reflects the seen/unseen balance.

We further evaluate the contribution of low-rank modeling on 96/24 using three paired training seeds, fixed 1,200-iteration checkpoints, and a common γ=0.75\gamma=0.75:

Low-rank vs. diagonalMean gainPositive seeds
Unseen accuracy+1.99 pp3/3
GZSL harmonic mean+1.16 pp2/3

The consistent unseen-class improvements across all three runs support LRCA's contribution to unseen-action discrimination. The harmonic mean also improves on average, complementing the results reported in Table 1.

4. Notation (Q3). We will define on first use rr (rank), λshape\lambda_{\rm shape} (factor alignment), λ\lambda (ZSL fusion), γ\gamma (GZSL gate), KK (seen-class activation) and T\mathcal T (flow-time grid), and cite Flora beside the inherited GZSL gate.

Reviewer_7UpcOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LoAD extends a cross-modal flow-matching framework with low-rank covariance alignment and frozen-model decision adaptation. It combines flow-matching and reconstruction evidence for ZSL and conditionally calibrates class latents using unlabeled test data. Experiments cover NTU-60, NTU-120, and PKU-MMD under ZSL and GZSL protocols.

Critical strengths

1.The paper revisits both the alignment and the decision stages of ZSAR. Diagonal Gaussian distributions cannot explicitly capture latent correlations, and a single generative score may not be suitable for different inference decisions. Improving both distributional alignment and test-time decision making is a reasonable and well-motivated direction.

2.LRCA introduces a low-rank covariance structure to capture latent correlations without modeling full covariance matrices. TDA combines flow-matching and reconstruction evidence while using unlabeled test statistics for calibration. The test-time components require no gradient updates or additional trainable parameters, and the reported overhead is modest.

3.The authors evaluate LoAD on NTU-60, NTU-120, and PKU-MMD, including a reproduced Flora baseline under the same implementation pipeline. The ablations examine individual components, hyperparameter sensitivity, and the conditional effectiveness of TSC. Computational cost and negative experimental results are also reported, improving transparency.

Critical weaknesses
  1. Diagonal + low-rank Gaussian modeling, gradient-free test-time adaptation, prototype calibration using unlabeled data, and multi-score fusion have all been explored in prior work. For example, low-rank variational inference and T3A (NeurIPS 2021) provide relevant precedents. Within the field, SkeletonCache (NeurIPS 2025) is already a training-free test-time method. LoAD mainly combines and adapts these established techniques to a generative ZSAR framework. Stronger comparisons with simpler covariance alignment, prototype calibration, and score-fusion alternatives are needed to establish the novelty and necessity of its specific designs.

2.TSC uses statistics from the entire unlabeled test set, introducing a transductive setting that is not clearly distinguished from other evaluation protocols in the main comparisons. On NTU-60 48/12 split, the full model without TSC (LRCA + DDI) reaches only 62.45%, below the reproduced Flora baseline (63.7%), whereas enabling TSC raises it to 66.06%. Thus, the final improvement on this split depends heavily on test-time calibration rather than better distributional alignment alone. Comparisons under matched test-data access conditions are needed.

  1. The TSC activation rule (|S| ≤ 60) is justified using the negative results in Table 5, but it is unclear whether this threshold was selected using an independent validation set. Any threshold between 55 and 95 would produce identical decisions on the reported splits. Moreover, the explanation based on candidate-class count is insufficient for ZSL, where 110/10 has fewer unseen candidates than 48/12, yet TSC is disabled on 110/10. The fusion weight λ is also specified separately for each split without a clear validation protocol. Given that hyperparameter variations can exceed some component gains, stronger validation are needed.

  2. LRCA aligns per-sample low-rank factors rather than directly matching class-conditional covariance matrices, and its factor-level loss is not invariant to equivalent covariance decompositions. Figure 5a reports nearly uniform mean support values (5.1–5.4), but their relationship to the whole test-set statistics used by TSC is unclear, leaving the claimed adaptive shrinkage insufficiently explained. Furthermore, DDI reduces ZSL accuracy on 110/10 (81.2% to 80.4%), while its contribution to GZSL is not independently isolated, particularly since reconstruction evidence is excluded from GZSL inference.

Questions for authors
  1. Which held-out classes and selection criteria were used to choose the fusion weights, GZSL thresholds, and TSC activation rule?
  2. How do comparisons change with matched access to test data, and does TSC remain effective on smaller or class-imbalanced test batches?
  3. What happens when the low-rank heads are retained but the factor-alignment loss is disabled?
  4. What are the separate contributions of time averaging, independent class-latent sampling, and reconstruction fusion?
3 comments
Reviewer_7UpcOpenReview ↗

Additional Reviews about the Submission 37783

  1. The authors state that the improvement on the NTU-60 48/12 split is 1.4 (Lines 371–374). However, Table 3 seems to show an increase from 63.7 to 66.1, corresponding to 2.4. Could there may be a manual error in the Table 3?

  2. The authors directly minimize the Frobenius distance between the low-rank factors of the two modalities (Lines 202–207). However, different low-rank factors can represent the same covariance matrix due to rotational ambiguity. It is unclear why directly aligning these factors is preferable to aligning the corresponding covariance matrices. Furthermore, the paper should provide more direct evidence that the learned low-rank components capture meaningful inter-dimensional correlations.

  3. The improvements from LRCA are not sufficiently isolated from the increase in model capacity. The number of trainable parameters increases from 15.81M to 18.96M after introducing LRCA (Lines 693–696). Without a parameter-matched diagonal Gaussian baseline, it is difficult to determine whether the reported improvements result from better covariance modeling or simply additional model capacity.

  4. The proposed methods show the marginal gains, even lower than those of state-of-the-art methods from 2025 shown in Table 1.

  5. The contributions of the different test-time inference components are not adequately disentangled. The proposed inference procedure combines many strategies (Lines 239–298). However, the current ablation study does not fully isolate the individual effects of these modifications. It is therefore difficult to determine which components are primarily responsible for the reported improvements.

AuthorsOpenReview ↗

Response to Reviewer 7Upc

We thank the reviewer for the additional questions.

A1. Is the 1.4-point statement a Table 3 error? The table is correct: reproduced Flora is 63.7%, LRCA-only is 64.7%, and full LoAD is 66.1% on 48/12. Therefore 63.7 to 66.1 = +2.4 pp overall; 64.7 to 66.1 = +1.4 pp beyond LRCA alone. We will correct the accompanying sentence, not the table.

A2. Why align low-rank factors rather than covariance matrices? Our factor-alignment loss serves as a shared-coordinate regularizer for the jointly trained skeleton and semantic encoders. It does not assume that low-rank covariance factors have a unique decomposition. Instead, it encourages both modalities to represent correlated latent variation in a consistent basis.

Importantly, factor alignment also constrains the induced covariance difference. For two factor matrices AA and BB,

∥AA⊤−BB⊤∥F≤(∥A∥2+∥B∥2)∥A−B∥F.\|AA^\top-BB^\top\|_F \leq (\|A\|_2+\|B\|_2)\|A-B\|_F.

Thus, reducing factor mismatch directly controls the discrepancy between their low-rank covariance contributions. The loss operates on the compact D×rD\times r factors without constructing full D×DD\times D covariance matrices.

We further verify this relationship experimentally. With identical low-rank heads and parameter counts, enabling factor alignment reduces the induced LL⊤LL^\top discrepancy from 0.1592 to 0.0638 (approximately 60%) in a 128-sample diagnostic. Therefore, the alignment effect is observed in the resulting covariance structure, not merely in the factor coordinates.

A3. Is LRCA's improvement simply a parameter-count effect? We tested this explanation with an ordinary projection using the same active head dimensions/rank, but without correlated covariance modeling. Archived best-checkpoint comparisons show:

NTU split55/548/12110/1096/24
Projection ZSL85.1059.1479.5365.29
Low-rank ZSL86.3664.6581.2365.69
Projection GZSL-H76.4557.8366.2552.48
Low-rank GZSL-H78.2959.3366.8452.92

The low-rank formulation wins on all four splits under both metrics, including +5.51 pp ZSL on 48/12. It therefore offers more than an ordinary similarly sized projection. In addition, FQ2 holds the complete architecture and parameter count fixed and shows that the alignment loss alone substantially reduces induced covariance mismatch. The two ablations separately support the contributions of low-rank covariance modeling and factor alignment.

A4. Are the gains marginal compared with 2025 methods? Not across the reported benchmarks. Against the best 2025 result listed in Table 1 for each split, LoAD improves 3/4 ZSL splits and 3/4 GZSL-H splits. The gains are especially pronounced on NTU-120 110/10: 80.4% versus 74.9% ZSL (+5.5 pp) and 66.6% versus 63.9% GZSL-H (+2.7 pp); on NTU-60 48/12, ZSL improves 66.1% versus 62.7% (+3.4 pp).

LoAD minus best listed 2025 (pp)55/548/12110/1096/24
ZSL-2.2+3.4+5.5+1.7
GZSL-H+1.4+0.6+2.7-1.8

Crucially, the claim of methodological progress is backed by a more tightly controlled comparison: with the same backbone and released features, LoAD exceeds the reproduced Flora baseline on all four NTU ZSL splits (+1.8/+2.4/+1.1/+1.9 pp) and on PKU-MMD by +3.3 pp. The broader comparisons establish competitive recognition accuracy; the matched-backbone gains substantiate the contributions of LoAD's designs.

A5. Which inference components drive the improvements? We provide component-specific controls to distinguish the contributions of calibration, flow-time scoring, sampling, and reconstruction fusion.

TSC: With the other components fixed, enabling TSC improves NTU-60 48/12 from 62.45% to 66.06% ZSL (+3.61 pp) and from 58.36% to 59.68% GZSL-H (+1.32 pp) (Table 5).

DDI scoring: We evaluate each inference choice independently on a fixed checkpoint without TSC:

NTU 48/12 scoringZSL accuracyGain
Single-time FM (independent draws)58.96%--
Multi-time FM59.02%+0.06 pp
Shared-noise FM59.11%+0.15 pp
FM + reconstruction59.57%+0.61 pp

Reconstruction fusion provides the largest isolated scoring gain. This improvement is not limited to the 48/12 split: with a single fixed λ=0.1\lambda=0.1, reconstruction improves ZSL on all four NTU splits by +0.89/+0.61/+0.97/+1.03 pp. In particular, the +0.97 pp gain on 110/10 isolates reconstruction's contribution from the sequential, combined-component ablation in Table 3. These controlled results identify TSC and reconstruction fusion as the principal inference contributions, while time averaging and the sampling choice have smaller effects. In GZSL, reconstruction is not used; decisions retain Flora's FM-based gate.

AuthorsOpenReview ↗

Response to Reviewer 7Upc

We thank the reviewer for the detailed feedback. W and Q refer to the numbered weaknesses and questions.

1. Novelty and simpler alternatives (W1). LoAD advances Flora through cross-modal covariance modeling (LRCA) and complementary generative inference (DDI/TSC). LRCA models correlations absent from diagonal Gaussians. Against an ordinary projection with matched active-head dimensions, low-rank modeling performs better on all four NTU splits (e.g., 64.65% vs. 59.14% ZSL on 48/12). Holding the low-rank heads fixed, factor alignment further reduces the induced covariance discrepancy by ~60%. DDI's reconstruction fusion improves all four splits with one fixed λ=0.1\lambda=0.1; TSC separately calibrates class latents without gradient updates. Most importantly, LRCA alone, without DDI/TSC, exceeds reproduced Flora on all four splits (+1.0/+1.0/+1.9/+0.8 pp, Table 3). These controls establish gains beyond combining existing inference techniques.

2. Fairness and batch dependence (W2, Q2). The cited 62.45% is LRCA + DDI with TSC off, not LRCA-only. Table 3 shows 64.7% vs. 63.7% for LRCA-only versus reproduced Flora on 48/12, with TSC absent in both. An additional three-seed comparison uses fixed 1,200-iteration checkpoints and matched inference access:

NTU 48/12 ZSLDiagonalLow-rankGain
TSC off59.42%61.11%+1.69 pp
TSC on (batch 64)59.43%61.66%+2.23 pp

Thus LRCA's benefit persists without TSC. Table 5 independently isolates TSC within the same full model (62.45%→66.06% ZSL). The supplementary fixed-endpoint results are separate from Table 3. Batch statistics are optional: LRCA and DDI still classify individual queries when TSC is disabled, so small or imbalanced batches do not prevent LoAD inference.

3. Hyperparameters and TSC activation (W3, Q1). The ∣S∣≤K|\mathcal S|\leq K rule counts seen training classes, not unseen candidates. Table 5 supports stronger calibration for smaller seen-class sets and reduced strength for larger ones. K=60K=60 is not a sensitive optimum: any 55≤K<9655\leq K<96 makes the same activation decisions on the reported NTU splits. Table 4 evaluates five ranks and five alignment weights. To test fusion without split-specific weights, we use one λ=0.1\lambda=0.1 and obtain ZSL gains of +0.89/+0.61/+0.97/+1.03 pp across the four splits. NTU GZSL also uses a shared γ=0.75\gamma=0.75; a three-seed 96/24 control yields +1.99 pp mean unseen accuracy for low-rank over diagonal modeling. The original parameter values are explicitly listed in Appendix Table 7.

4. Factor alignment and adaptive shrinkage (W4, Q3). Low-rank factors are non-unique, but factor alignment acts as a shared-coordinate regularizer for jointly trained skeleton and semantic encoders. Its effect extends to the resulting covariance contributions: with identical heads and parameter counts, enabling only the alignment loss reduces the LL⊤LL^\top discrepancy from 0.1592 to 0.0638 (~60%) (128-sample, one-seed diagnostic). Three-seed ZSL also changes from 61.02% to 61.11%. The module-level LRCA gains are separately shown in Table 3.

In Fig. 5(a), the average support near 64/12=5.3364/12=5.33 follows from 64 normalized soft votes over 12 classes; adaptivity is reflected in the support–anchor-movement correlation (r=−0.83r=-0.83), consistent with larger correction for less-supported classes. The TSC ON/OFF gain in Table 5 confirms the recognition benefit of this calibration.

5. Independent DDI contributions (W4, Q4). A fixed-checkpoint, no-TSC 48/12 study isolates each scoring choice:

ScoringZSL
Single-time FM58.96%
Multi-time FM59.02%
Shared-noise FM59.11%
FM + reconstruction59.57%

Reconstruction is the main isolated scoring gain (+0.61 pp); with fixed λ=0.1\lambda=0.1 it improves all four splits, including +0.97 pp on 110/10. Thus the Table 3 cumulative change (81.2%→80.4%) should not be attributed to reconstruction alone. For GZSL, LoAD retains Flora's FM-only gate, so GZSL gains arise from LRCA/TSC, not reconstruction (TSC ON/OFF: 58.36%→59.68% GZSL-H, Table 5).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.