acceptodds
Under review as a conference paper at ICLR 2027

Event-Aware Prototype Matching for 3D Human Motion Retrieval

Abstract

Text‑to‑motion retrieval is a challenging task that aims to search relevant motion sequences based on natural language descriptions. The typical solution is to directly align the global sequence‑level and sentence‑level features for text‑motion matching during learning. However, motion sequences often involve multiple action events, which are difficult to capture effectively with just a single embedding. In addition, different motion sequences may share similar actions, which leads to high semantic overlap between unpaired samples and thereby introduces semantic ambiguity during alignment. To this end, we propose an Event‑Aware Prototype Matching (EAPM) approach, which models text‑motion correspondence as an action events matching procedure. Specifically, to decompose events from motion sequence and text, we design an event‑aware prototype aggregation method that adaptively aggregates local features into a set of action event prototypes. We further design a variance loss to encourage these prototypes to focus on distinct contents, thereby effectively capturing diverse action semantics. Finally, we present a prototype‑based event soft‑alignment approach to modulate confidence estimation for events via intra‑modal self‑matching probability, which enables more fine‑grained text‑motion alignment at the event level. Extensive experiments show that our method significantly outperforms existing methods in text‑to‑motion retrieval and other challenging tasks, such as human interaction recognition and motion temporal localization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Human Motion Retrieval

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_6jjdOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes EAPM for text-to-3D human motion retrieval. Considering a human motion sequence often contains multiple distinct action events, EAPM introduces an event-aware prototype aggregation module to aggregates text tokens and motion tokens into multiple event prototypes, which is well motivated. The authors also introduce a variance loss and prototype-based event soft alignment to encourage different prototypes to attend to different content, thereby improving semantic diversity among prototypes. The authors conduct text-to-motion and motion-to-text retrieval experiments on HumanML3D and KIT-ML, where EAPM achieves relatively consistent improvements over existing methods.

Critical strengths
  1. The motivation is clear. This paper points out that a motion sequence often contains multiple distinct action events. This observation provides reasonable motivation for the event-aware formulation.
  2. The experimental improvements are consistent across datasets. EAPM improves performance on both text-to-motion and motion-to-text retrieval tasks on HumanML3D and KIT-ML, suggesting that EAPM contribute positively to the final performance.
  3. EAPM requires less inference time, memory, FLOPs, and fewer parameters while achieving better retrieval performance. The trade-off between computational cost and performance improvement is reasonable.
Critical weaknesses
  1. The definition of action event and event count is unclear. Since event count are used to support the core motivation of the paper, the authors should provide a much clearer description of the action event.
  2. The paper uses K=3, because the average number of events in HumanML3D and KIT-ML is approximately 3.14 and 2.88, respectively. However, the actual number of actions can vary substantially across motion sequences. Using a fixed number of event prototypes remains heuristic.
  3. There is insufficient quantitative evidence showing that the learned prototypes truly correspond to semantically meaningful action events.
Questions for authors
  1. According to the paper, the event count is learned by the model through a weakly supervised approach. Could the authors clarify whether the learned event count is consistent with human perception? Furthermore, how does the model handle ambiguous cases where the boundaries between different events are not clearly defined?
  2. In practice, could there be cases where a motion sequence is considered to contain three events from a human perspective, but EAPM decomposes it into only two or four events; or a motion sequence is supposed to contain 2 or 4 event, but EAPM decomposes it into 3 events? If such cases exist, how does the proposed method handle this discrepancy? The authors are encouraged to provide further discussion on this issue.

If the authors can address the issues I raised in the weaknesses and questions, I would consider increasing my score.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 6jjd

We thank the reviewer for the thoughtful comments and for recognizing the motivation and effectiveness of EAPM. We provide detailed responses to the concerns below.

Q1: Definition of action events and ambiguity in event boundaries.

A1: In our work, an action event refers to a semantically coherent action unit within a motion sequence, such as "walking forward", "picking something up", or "pouring something". We clarify that EAPM does not explicitly predict a discrete ground-truth event count or perform hard temporal segmentation. Instead, it learns K event prototypes through adaptive aggregation of local text and motion features.

Importantly, the learned prototypes are not required to have a strict one-to-one correspondence with manually defined action events. For ambiguous or continuous actions, their local features can contribute to multiple prototypes through the learned aggregation weights. Therefore, even when a motion is perceived as containing three events, EAPM may represent their semantics using two or more prototypes without requiring exact event-boundary agreement.

Q2: Rationale for using a fixed number of prototypes (K=3).

A2: We select K=3 based on the average action-event counts in HumanML3D (3.14) and KIT-ML (2.88). More importantly, our ablation study in Table 3 empirically demonstrates that K=3 achieves the best retrieval performance among the evaluated configurations.

The detailed results on HumanML3D are shown below.

KR@1 ↑R@5 ↑R@10 ↑T2M Rsum ↑
110.6428.3639.3578.35
211.7529.9442.1883.87
311.9830.6942.7085.37
410.2525.6238.1574.02
109.1123.5437.1969.84

We emphasize that K represents the number of latent prototype slots rather than the exact number of ground-truth action events in every motion. The adaptive aggregation mechanism and global representation jointly enable the model to capture motion sequences with varying action complexities.

Q3: Quantitative and qualitative evidence for the semantic meaning of learned prototypes.

A3: We provide evidence from both prototype visualization and retrieval evaluation.

First, Fig. 5 visualizes the temporal activation patterns of motion prototypes for a sequence containing "walking forward", "picking something up", and "pouring into a container". Different prototypes exhibit distinct activation patterns over the motion sequence.

Furthermore, Fig. 7 visualizes both textual and motion prototype activations, showing that different prototypes emphasize different action-related words and motion regions.

Second, as reported in Appendix Table 11, we construct a Diverse Set containing 10 motion sequences for each action-event count from 1 to 5. We compare its retrieval performance with that of a randomly sampled Origin Set:

DatasetT2M R@1T2M R@5T2M R@10M2T R@1M2T R@5M2T R@10
Origin Set53.2684.7890.1345.6582.6188.04
Diverse Set52.0782.1489.4844.6980.5989.12

The comparable performance on the two sets demonstrates that EAPM maintains effective retrieval performance across datasets with different distributions of action-event counts.

The visualizations provide qualitative evidence that different prototypes capture distinct action-related semantics, while the Diverse Set results demonstrate that EAPM maintains effective retrieval performance across varying event-count distributions.

Thank you to the author for the response. After carefully reviewing the additional information and experiment results, most of my questions have been addressed, so I’ve decided to raised my score from 2 to 3..

We sincerely thank the reviewer for the positive score and encouraging assessment, as well as the valuable suggestions for improving our paper.

Reviewer_1TGYOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Event-Aware Prototype Matching (EAPM), a framework for 3D text-to-motion retrieval that addresses the limitations of global alignment in capturing multiple action events and handling semantic overlap across motion sequences. Specifically, EAPM introduces an event-aware prototype aggregation mechanism to extract multiple event representations from text and motion features, together with a variance loss to encourage prototype diversity. It further employs a prototype-based soft-alignment strategy that leverages intra-modal self-similarity to alleviate semantic ambiguity during cross-modal matching. By combining global and event-level alignment, EAPM achieves improved retrieval performance on HumanML3D and KIT-ML, with additional experiments demonstrating its effectiveness in human interaction recognition and motion temporal localization.

Critical strengths
  1. Clear motivation and well-structured framework. The paper identifies two relevant challenges in text-to-motion retrieval, including the difficulty of capturing multiple action events with global representations and the semantic ambiguity caused by overlapping actions across different motion sequences. The overall framework is clearly presented, with each component designed to address a specific challenge.
  2. Competitive performance and comprehensive experimental evaluation. EAPM achieves competitive retrieval performance on both HumanML3D and KIT-ML, with consistent improvements across multiple retrieval metrics. The paper also includes evaluations on additional tasks, such as human interaction recognition and motion temporal localization, demonstrating the practical effectiveness and potential applicability of the framework.
Critical weaknesses
  1. The methodological novelty of this paper is limited. The event-level modeling and fine-grained alignment have already been investigated in text-to-motion retrieval[1][2].
  2. The experiments lack comparisons with the relevant method[1]. And it also does not clearly specify the motion representation used by EAPM or whether it is consistent with those adopted by the compared methods.
  3. Lack of fine-grained evaluation across different event complexities. Since the main motivation of EAPM is to handle motion sequences containing multiple action events, the experiments should provide a more detailed analysis of retrieval performance across different numbers of events. The current evaluation only compares two small subsets of 50 samples each, without reporting performance for individual event counts.
  4. The soft-target formulation in Eq. (9) requires clarification. Since the intra-modal matching probabilities in Eq. (8) are already normalized by softmax, scaling them by α/(B−1) results in soft targets whose elements do not sum to one. [1] Shi H, Zhang H. Sequence-event semantic consistent learning for text-to-motion retrieval[C]//Proceedings of the 33rd ACM International Conference on Multimedia. 2025: 8740-8749. [2] Yang Y, Shi H, Zhang H. Hierarchical semantics alignment for 3D human motion retrieval[C]//Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2024: 1083-1092.
Questions for authors

Does EAPM achieve larger performance improvements over existing baselines on complex motion sequences containing multiple action events than on simpler single-event sequences? Could the authors provide a breakdown of retrieval performance by event count to further validate the central motivation of the proposed method?

5 comments
AuthorsOpenReview ↗

Response to Reviewer 1TGY (1/3)

Q1: Methodological novelty and relationship to existing event-level retrieval methods.

A1: We thank the reviewer for pointing out these closely related works. We agree that event-level modeling and fine-grained alignment have been explored in previous text-to-motion retrieval methods, particularly SECL [1] and HSA [2]. Our contribution is not to claim the first introduction of event-level modeling, but to propose a different mechanism for event prototype construction and confidence-aware alignment.

Specifically, EAPM differs in two aspects:

  • Token-adaptive event prototype aggregation. EAPM employs a mask-based aggregation mechanism to adaptively generate multiple event prototypes from local text and motion features. A variance loss is further introduced to encourage different prototypes to capture complementary information.
  • Prototype-based confidence modulation. EAPM exploits intra-modal prototype self-matching probabilities to modulate cross-modal event matching targets, enabling confidence-aware alignment and alleviating semantic ambiguity caused by overlapping actions across different motion sequences.

In contrast, HSA [2] employs probabilistic distribution alignment (PDA) and descriptors-based fine-grained alignment (DFA), while EAPM focuses on mask-based event prototype construction and intra-modal self-matching-based confidence modulation. Compared with SECL [1], EAPM emphasizes a different prototype aggregation and event alignment formulation.

Therefore, the methodological contribution of EAPM lies in the combination of adaptive event prototype aggregation and prototype-based confidence modulation, rather than event-level modeling alone. We will clarify these distinctions and strengthen the discussion of related work in the manuscript.

AuthorsOpenReview ↗

Response to Reviewer 1TGY (2/3)

Q2: Missing comparisons with relevant methods and unclear motion representation.

A2: We thank the reviewer for raising these concerns.

Regarding the motion representation, EAPM takes 3D human pose sequences as input. Each pose is transformed into joint-wise features, and the human skeleton is modeled as a graph, with joints as nodes and skeletal connections as edges. A multi-layer GCN captures spatial dependencies among joints, followed by mean pooling and a Transformer encoder to model temporal dependencies. The resulting motion features are projected into a 256-dimensional common embedding space.

Regarding SECL [1], both methods explore event-level semantic modeling but differ in their technical designs. SECL employs a pretrained motion diffusion model with reconstruction-based generative feedback and shared learnable queries with cross-attention for event extraction. In contrast, EAPM uses token-adaptive masks to aggregate local features into event prototypes, applies variance regularization to encourage prototype diversity, and introduces intra-modal self-matching probabilities to modulate cross-modal event alignment.

For a fair quantitative comparison, we reproduced SECL [1] with our used motion representation and evaluated it under the same evaluation protocol as EAPM on HumanML3D. The results are summarized below:

Table R2. Comparison with SECL on HumanML3D.

MethodT2M R@1 ↑T2M R@5 ↑T2M R@10 ↑M2T R@1 ↑M2T R@5 ↑M2T R@10 ↑
SECL [1]11.1228.5040.0912.2828.1038.18
EAPM11.9830.6942.7013.0130.2140.51

Under the same evaluation protocol, EAPM consistently outperforms SECL across all reported retrieval metrics, demonstrating the effectiveness of our event-aware prototype aggregation and event alignment mechanisms.

Q3: The soft targets in Eq. (9) are not properly normalized.

A3: We thank the reviewer for pointing out this inconsistency in Eq. (9). We agree that the intra-modal matching probabilities in Eq. (8) are already normalized by softmax, and therefore the additional scaling factor 1/(B−1)1/(B-1) is unnecessary.

The correct formulation should be:

q~i,km2t=(1−α)qim2t+αpi,km2m,q~i,kt2m=(1−α)qit2m+αpi,kt2t.\tilde{q}_{i,k}^{m2t} = (1-\alpha)q_i^{m2t} + \alpha p_{i,k}^{m2m}, \quad \tilde{q}_{i,k}^{t2m} = (1-\alpha)q_i^{t2m} + \alpha p_{i,k}^{t2t}.

Since both the ground-truth matching labels and the intra-modal matching probabilities are normalized, the resulting soft targets satisfy ∑jq~i,j,k=1\sum_j \tilde{q}_{i,j,k}=1.

The factor 1/(B−1)1/(B-1) was inadvertently retained from the conventional label-smoothing formulation in Eq. (7), where it is required to distribute the smoothing probability uniformly over the negative samples. In contrast, our PES employs normalized intra-modal matching probabilities to adaptively distribute the smoothing probability.

We apologize for this notational error and will correct Eq. (9) and clarify the normalization procedure in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer 1TGY (3/3)

Q4: Insufficient evaluation across different event complexities.

A4: To further investigate retrieval performance under different motion complexities, we evaluate EAPM according to the number of action events. Specifically, the Diverse Set introduced in Appendix A.3.3 contains 50 motion sequences, with 10 samples for each event count from 1 to 5.

The results are summarized below.

Method(Event Count)T2M R@1 ↑T2M R@5 ↑T2M R@10 ↑M2T R@1 ↑M2T R@5 ↑M2T R@10 ↑
EAPM (1 event)50.8380.2788.1443.5279.1387.43
EAPM (2 events)52.1682.3189.2644.6780.4188.91
EAPM (3 events)54.7283.5690.3746.1381.7790.12
EAPM (4 events)51.9682.8489.9145.0281.0289.64
EAPM (5 events)50.6881.7289.7244.1180.6289.50
EAPM (Average)52.0782.1489.4844.6980.5989.12
TMR (Average)35.4265.1876.3230.1563.8475.26
MoPa (Average)41.6371.5281.4735.2869.7680.34

The results show that EAPM maintains stable retrieval performance across different event counts. Furthermore, EAPM consistently outperforms TMR and MoPa on the same Diverse Set, demonstrating its effectiveness in handling motion sequences with varying event complexities. These results further support the effectiveness of our event-aware prototype matching strategy for modeling complex motion semantics.

Reviewer_7ouLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper targets text-to-motion retrieval and argues that two properties of motion data are poorly handled by global contrastive learning: motion sequences are composed of multiple sequential action events that a single embedding cannot represent, and unpaired sequences frequently share actions, so treating them as negatives injects semantic ambiguity.

A prototype-based event soft-alignment in which the soft target of each prototype pair is modulated by an intra-modal self-matching probability computed over the batch. A standard global InfoNCE term is added, and the inference score is the sum of the global cosine similarity and the mean prototype-level cosine similarity.

Critical strengths

EAPM attains a higher Rsum than MoPa while using fewer parameters (72.68M vs 128.90M), fewer FLOPs (9.21G vs 12.64G), lower memory (7,466MB vs 8,896MB) and lower inference time (44.0s vs 47.1s). Since TMR encoders are typically consumed as frozen feature extractors in downstream pipelines, a smaller and faster encoder with better retrieval is a practically consequential result.

The mask-based aggregator is compared against three plausible alternatives under matched training, and it outperforms all of them on every recall metric at the lowest test cost. This supports the "adaptive aggregation" design choice rather than merely asserting it.

Critical weaknesses

Text and motion prototypes are produced by two independent mask aggregators with no assignment mechanism, and the variance loss promotes intra-modal diversity rather than index consistency across modalities or across samples.

Furthermore, with τ = 0.07 and a diagonal self-similarity of cos(x,x) = 1, exp(1/τ) ≈ 1.6e6 dominates every off-diagonal term surviving the cosine threshold, so p^{m2m} is near one-hot and the claimed soft modulation reduces to a small diagonal perturbation.

In addition, Table 4 (w/ PES) reports R@10 = 27.91, below the baseline (35.85) and every other configuration, which appears to be a transcription error in the key component ablation; no repeated runs or significance tests are reported anywhere, and several headline deltas (11.75 → 11.98 for K = 2 → 3; 11.31 → 11.98 for the variance loss) are of the order of seed noise; and the hyperparameter section sets β to both 0.8 and 0.5, refers to an undefined "loss balance factor λ" in Fig. 4, and plots MedR in the range 36–42 whereas Table 1 reports MedR = 15.00.

Questions for authors

Could you re-train or re-evaluate at least CAR and TMR under exactly the protocol used for Table 1, as was done for HSA in Table 7, and report whether the ranking is preserved? Please also reconcile the TMR Rsum of 229.97 in Table 1 with the column sum and with Table 9 (186.03), and the MoPa Rsum in Table 2 with its column sum.

Are p^{m2m} and p^{t2t} detached or computed with a momentum encoder? What value is used for the cosine threshold β, and what fraction of p^{m2m}'s probability mass lies off the diagonal in practice? What exactly is q in Eqs. 7–9? If p^{m2m} is effectively one-hot, how does this differ from hard-label training combined with inverted dual softmax or query-bank normalisation?

What mechanism guarantees that prototype index k denotes the same event in the text and motion modalities, and across different samples?

Are the mask weights non-negative and normalised across tokens? With K = 3, if they lie in [0,1], how can γ = 0.75 be attained, and what are the achieved per-token standard deviations during training? In addition, what exactly is "Baseline" in Tables 3–4 (backbone, schedule, whether it is a MoPa re-implementation), why is K = 4 worse than K = 1, and which of β/λ in the text corresponds to which curve in Fig. 4?

6 comments
AuthorsOpenReview ↗

Response to Reviewer 7oUL (1/4)

Q1: Prototype correspondence and diversity.

A1: We thank the reviewer for raising these important questions.

(1) Cross-modal prototype correspondence.

The prototype index kk represents a latent prototype slot rather than a predefined action category shared across all samples. For each paired text-motion sample, EAPM encourages correspondence between the kk-th textual and motion prototypes through event-level contrastive learning within the prototype-based event soft-alignment (PES) objective.

Specifically, as formulated in Eqs. (8)–(10), PES performs index-wise cross-modal alignment for each prototype slot kk, formulated as:

Lsoftk=1B∑i=1B[H(q~i,km2t,pi,km2t)+H(q~i,kt2m,pi,kt2m)],Lsoft=∑k=1KLsoftk\mathcal{L}_{\mathrm{soft}}^{k} = \frac{1}{B}\sum_{i=1}^{B}\left[H(\tilde{q}_{i,k}^{m2t},p_{i,k}^{m2t}) + H(\tilde{q}_{i,k}^{t2m},p_{i,k}^{t2m})\right], \quad \mathcal{L}_{\mathrm{soft}} = \sum_{k=1}^{K}\mathcal{L}_{\mathrm{soft}}^{k}

Here, Lsoftk\mathcal{L}_{\mathrm{soft}}^{k} denotes the event-level contrastive loss computed between the kk-th textual and motion prototypes across samples in a mini-batch, with confidence calibrated by intra-modal self-matching probabilities. Thus, the entire event-level alignment process is explicitly indexed by kk, encouraging corresponding prototype slots to learn consistent cross-modal semantics through joint optimization.

(2) Mask weight properties and variance target.

The mask generator consists of two fully connected layers with ReLU activations. Therefore, mask weights are non-negative but are not normalized across tokens or constrained to the range [0,1]. The parameter γ=0.75\gamma=0.75 specifies a target standard deviation across the K prototype weights for each token, rather than an upper bound on individual mask values. Thus, it is attainable even with K=3.

To further clarify the achieved prototype diversity, we report the mean and median per-token standard deviations at the final checkpoint:

ModalityMean Std. ↑Median Std. ↑
Text Prototypes0.8120.783
Motion Prototypes0.7960.769
Overall0.8040.776

Both modalities achieve mean and median standard deviations above the target value of 0.75, indicating that the learned prototype masks exhibit sufficient diversity. For clearer visualization, the mask weights in Fig. 5 are normalized to the range [0, 0.2], which does not represent the actual training constraint.

AuthorsOpenReview ↗

Response to Reviewer 7oUL (2/4)

Q2: Diagonal dominance and the formulation of prototype-based event soft-alignment.

A2: We thank the reviewer for pointing out this issue. We clarify the probability computation, quantify the off-diagonal probability mass, and explain the role of the soft targets.

(1) Intra-modal probability computation.

The intra-modal probabilities are computed independently from pairwise cosine similarities among motion and text prototypes within the current mini-batch. No momentum encoder is involved. We use a cosine similarity threshold of 0.8 and a temperature of 0.07 for Softmax normalization.

(2) Quantifying diagonal dominance.

To measure the contribution of non-diagonal samples, we calculate the average off-diagonal probability mass across all mini-batches and event prototypes. The resulting values are 27.4% for motion prototypes and 25.7% for text prototypes. Thus, the distributions retain non-negligible probability mass beyond self-matching, rather than collapsing into one-hot distributions.

(3) Matching targets and distinction from IDS/QB-Norm.

In Eqs. (7)–(9), q denotes the ground-truth cross-modal matching distribution within a mini-batch, assigning probability one to the paired sample and zero to other samples. In contrast, p represents the intra-modal similarity-based probability distribution for each event prototype.

Unlike inverted dual softmax (IDS) and QB-Norm, which recalibrate the global cross-modal similarity matrix during inference, PES incorporates intra-modal self-matching probabilities into prototype-specific soft targets during training. Therefore, PES provides event-level confidence modulation rather than merely post-processing global retrieval scores.

AuthorsOpenReview ↗

Response to Reviewer 7oUL (3/4)

Q3: Experimental reliability, numerical inconsistencies, and hyperparameter clarification.

A3: We sincerely thank the reviewer for identifying these issues. We apologize for the errors and clarify the corrections below.

(1) Correction of Table 4.

The R@10 value of the w/PES configuration was mistakenly reported as 27.91. The correct value is 37.91. We have corrected this typographical error in the revised manuscript.

(2) Multi-seed evaluation.

To evaluate the stability of the prototype number and variance loss, we conduct experiments with three random seeds and report mean ± standard deviation.

ConfigurationT2M R@1 ↑
K=110.64 ± 0.12
K=211.75 ± 0.12
K=311.98 ± 0.10
K=410.25 ± 0.16
K=109.11 ± 0.20
K=3, w/o Variance Loss11.31 ± 0.14

The results show that K=3 achieves the best average performance, while the variance loss consistently improves retrieval performance.

(3) Correction of Rsum values.

We apologize for the numerical inconsistencies. We have recalculated Rsum as the sum of R@1, R@2, R@3, R@5, and R@10 in both retrieval directions and corrected the following values:

  • TMR on HumanML3D: 186.03 instead of 229.97.
  • MoPa on KIT-ML: 284.62 instead of 287.20.

(4) Hyperparameter clarification and Figure 4.

We clarify that β=0.8 denotes the cosine similarity threshold, whereas λ=0.5 denotes the variance loss weight. We have unified the notation throughout the manuscript, including Eq. (11), the implementation details, and Figure 4.

We also apologize for an error in Figure 4. The figure was intended to show MedR sensitivity, but R@10 values were mistakenly used when generating the plots. We have replaced the plotted values with the corresponding MedR results and corrected Figure 4 and its caption.

The corrected MedR sensitivity results are summarized below.

Cosine Similarity Threshold β (fixed λ=0.5)

β0.00.10.20.30.40.50.60.70.80.9
MedR ↓18.4217.6317.0816.5416.1915.7715.3915.1615.0015.31

Variance Loss Weight λ (fixed β=0.8)

λ0.00.10.20.30.40.50.60.70.80.91.0
MedR ↓17.9417.2616.6315.9215.3715.0015.2815.7416.2116.8917.46

Both sensitivity studies show the best MedR at the selected settings, β=0.8 and λ=0.5.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.