acceptodds
Under review as a conference paper at ICLR 2027

Simple Yet Effective Margin Semantic Extraction for Large Language Models enhanced Recommender Systems

Abstract

Recent advances in large language models (LLMs) have introduced rich semantic information into recommender systems. However, LLMs-derived semantics are not inherently compatible with collaborative signals, making it difficult to fully exploit their benefits and potentially causing information distortion. To address this issue, we shift from fusion or alignment to extraction and propose a model-agnostic framework, MarRec, which imposes a lightweight margin constraint as an implicit information bottleneck to effectively extract useful semantics. Additionally, we design two instance margin variants and verify that fine-grained constraints can further improve useful semantic extraction. Extensive experiments demonstrate that MarRec outperforms baselines while maintaining high efficiency and showing strong robustness under data sparsity, long-tail, and cold-start scenarios.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Recommender Systems, Large Language Models, Margin Constraint, Information Bottleneck

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 2 reviews, above 27% of the others.
Reviewer_ZnDEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MarRec, a two-stage recommendation framework by semantic extraction. In the first stage, semantic embeddings are projected through a MoE module and optimized by the margin loss. In the second stage, the learned semantic representations are used to initialize trainable ID embeddings and trained by standard recommendation loss. The paper also introduces adaptive and soft margin variants and theoretically demonstrates that MarRec embodies the information bottleneck. Experiments on four datasets and five recommendation backbones demonstrate improvements over existing semantic fusion and alignment methods.

Critical strengths
  1. MarRec introduces a lightweight margin loss without requiring complex additional modules. The framework can be applied to different recommendation backbones. This simplicity makes MarRec useful for practical recommendation systems.
  2. The paper evaluates MarRec on four datasets and five recommendation backbones. The reported results support the practical effectiveness of the proposed framework.
Critical weaknesses
  1. The motivation and semantic extraction mechanism are insufficiently justified. First, the paper claims that Stage I compresses irrelevant semantic information. However, the margin loss operates on the output of a collaborative recommendation model, rather than directly constraining semantic information. Although gradients propagate to the MoE, there is no guarantee that the margin constraint selectively removes irrelevant semantics instead of merely limiting collaborative optimization. Second, the claimed limitations of fusion and alignment methods are oversimplified. For example, AlphaFuse (SIGIR 2025) preserves semantic information while incorporating collaborative signals, and alignment methods can also employ fine-grained constraints. Since MarRec similarly optimizes semantic representations using collaborative signals, its fundamental advantage over existing fusion and alignment approaches remains unclear.
  2. The information bottleneck analysis lacks sufficient theoretical justification. First, theorem 4.2 assumes that the margin constraint implies (I(H1;Y)≤m)(I(H^1;Y)\leq m). However, the margin loss constrains ranking score differences, and no theoretical justification is provided for this mutual information bound. In fact, satisfying the margin does not generally imply such an upper bound. Second, even accepting this assumption, the derived bound also depends on (H(H1∣Y))(H(H^1\mid Y)) and (H(H1∣Es))(H(H^1\mid E^s)). Since the margin loss jointly optimizes the MoE and the collaborative model, both conditional entropy terms may change during training. The authors do not establish how these changes are controlled by margin constraint. Therefore, the derived bound does not guarantee semantic information compression.
Questions for authors
  1. How does the margin constraint guarantee the proposed semantic extraction rather than merely limiting collaborative optimization?
  2. What are the fundamental advantages of MarRec over existing controllable fusion and fine-grained alignment methods, such as AlphaFuse?
  3. How do the authors justify the bound (I(H1;Y)≤m)(I(H^1;Y)\leq m) and guarantee semantic compression when the remaining conditional entropy terms may change during joint optimization?
4 comments

We are grateful to the reviewer for the detailed feedback and insightful questions.

Weakness 1 (semantic extraction) and Question 1

MarRec uses collaborative supervision to learn a preference-useful projection of frozen semantic inputs, then transfers that projection into standard recommendation training. The margin controls how this supervision updates the semantic mapping.

Let Z=gθ(Es)Z=g_\theta(E^s) denote the MoE-projected semantic embeddings, and let Δt=s(u,i+)−s(u,i−)\Delta_t=s(u,i^+)-s(u,i^-) for triplet tt. For the fixed-margin loss in Equation 5, ℓt=[m−Δt]+\ell_t=[m-\Delta_t]_+, the gradient away from the hinge boundary is

=-\mathbf1\{\Delta_t<m\} \left(\frac{\partial Z}{\partial\theta}\right)^\top\nabla_Z\Delta_t.$$ Thus, although the loss is evaluated at the recommender output, it directly regulates the preference supervision received by the semantic projector. Once $\Delta_t>m$, that triplet contributes no direct main-loss gradient, whereas BPR continues to reward a larger gap at every finite $\Delta_t$. This removes the incentive from already-satisfied triplets to keep adapting the shared semantic mapping merely to increase their score gaps. Other triplets and auxiliary losses can still update shared parameters. The margin does not impose an upper bound on the score gaps themselves. This mechanism is particularly clear with LightGCN: its graph propagation has no trainable transformation weights, and the original semantic embeddings are frozen in Stage I. In the fixed-margin variant, the MoE projector is therefore the only trainable representation mapping. The margin directly controls the learning of this semantic mapping, rather than acting only on a separate set of collaborative-model weights. Stage II then transfers the projected vectors into independently trainable ID embeddings and resumes standard recommendation training (Section 3.4). The existing controls examine both the training objective and the semantic inputs. w/o Margin (Section 5.2 and Appendix D.2) keeps the semantic inputs, projection architecture, and two-stage transfer procedure, while replacing the Stage-I main loss with BPR. It is among the variants with the largest performance degradation. Appendix C.2 further compares MarRec with generic ID warm-start and frozen random-input controls. The reported LightGCN Recall@20 results are: | Method | Beauty | Sports | Steam | Yelp | |---|---:|---:|---:|---:| | Base | 0.1551 | 0.0845 | 0.1357 | 0.1061 | | ID-Standard | 0.1544 | 0.0846 | 0.1359 | 0.1065 | | ID-Margin | 0.1594 | 0.0863 | 0.1367 | 0.1037 | | Rand-Margin-MoE-Frozen | 0.1506 | 0.0809 | 0.1342 | 0.1088 | | **MarRec** | **0.1829** | **0.1142** | **0.1463** | **0.1273** | Together, these controls support the contribution of margin-based adaptation under the same semantic-transfer pipeline and the value of semantic content beyond generic warm-start. The scope of the information-theoretic interpretation is addressed in our response to Weakness 2 and Question 3 below. # Weakness 1 (fusion and alignment) and Question 2 We agree that controllable fusion and fine-grained alignment can explicitly regulate semantic adaptation. Our discussion of potential information distortion concerns specific designs and settings. Both AlphaFuse and MarRec use collaborative supervision, but they differ in how adaptation is controlled and which embedding coordinates remain trainable. AlphaFuse uses the language-embedding spectrum to partition the input item embedding space, keeping the standardized semantic-rich coordinates fixed while learning collaborative ID updates in approximate null-space coordinates. MarRec instead learns a semantic projection from preference supervision, with the margin controlling each triplet's direct score-gap gradient reaching the projector during Stage I. The learned projections then initialize ordinary ID embeddings in Stage II, where all transferred coordinates remain trainable under the backbone's standard objective. This provides a task-adapted initialization followed by optimization of the entire transferred representation, without a prescribed frozen semantic subspace. Stage II and inference use the standard backbone without the semantic projector or an additional semantic-alignment objective. Fine-grained alignment can also regulate individual examples or relationships. MarRec's contribution lies in combining margin-controlled semantic adaptation with subsequent transfer into fully trainable ID embeddings. The comparisons with KAR, RLMRec, and PAD across five backbones, together with the MF results in Appendix C.4, support the effectiveness and architectural flexibility of this design. The comparison with AlphaFuse above concerns the mechanism. The reported experiments do not establish empirical superiority over AlphaFuse. We will add a citation to AlphaFuse and discuss these methodological differences in the Related Work of the revised manuscript.

Weakness 2 and Question 3

Building on Appendix A.1, we model the representations using Gaussian distributions, a common distributional assumption in representation modeling, and establish the mutual-information bound under explicit shared-covariance and moment conditions. To address the concern about joint optimization, we further state uniform covariance bounds that allow both conditional entropies to change while preserving the same information-capacity bound.

Let Hθ=Hθ1H_\theta=H_\theta^1 and fix m>0m>0. Consider a fixed population distribution of semantic inputs and targets, with finite relevant entropies. Suppose that, at each parameter state considered,

\qquad\Sigma_\theta\succ0, \qquad\operatorname{tr}(\Sigma_\theta^{-1}B_\theta)\leq2m,$$ where the covariance is shared across labels, the conditional means have finite second moments, and $B_\theta=\operatorname{Cov}_Y(\mu_{\theta,Y})$. The moment condition bounds between-label mean variation relative to within-label variation. Taking $Q_\theta=\mathcal N(\mathbb E_Y\mu_{\theta,Y},\Sigma_\theta)$, the KL decomposition gives $$\begin{aligned} I(H_\theta;Y) &=\mathbb E_Y D_{\mathrm{KL}}(P_{H_\theta\mid Y}\|Q_\theta) -D_{\mathrm{KL}}(P_{H_\theta}\|Q_\theta)\\ &\leq\tfrac12\operatorname{tr}(\Sigma_\theta^{-1}B_\theta)\leq m. \end{aligned}$$ Equation 5 gives $[m-\Delta]_+=m-\min(\Delta,m)$, so a triplet's direct main-loss gradient vanishes for $\Delta>m$. The moment restriction above is an explicit additional modeling condition. Hinge saturation alone does not enforce it. The mutual-information bound therefore follows under the stated distributional conditions, without assuming that a small hinge loss verifies those conditions. To control the conditional-entropy difference during joint optimization, additionally assume that $H_\theta\mid E^s=e$ is Gaussian with covariance $\Sigma_{\theta,e}$, and require $$\Sigma_\theta\preceq\sigma_y^2 I_d, \qquad\Sigma_{\theta,e}\succeq\sigma_e^2 I_d, \qquad 0<\sigma_e^2,\sigma_y^2<\infty,$$ for every parameter state considered and almost every $e$, with the same constants across parameter states. The Gaussian entropy bounds then yield $$\begin{aligned} I(E^s;H_\theta) &=I(H_\theta;Y)+h(H_\theta\mid Y)-h(H_\theta\mid E^s)\\ &\leq m+\frac d2\log\frac{\sigma_y^2}{\sigma_e^2}, \end{aligned}$$ where $h$ denotes differential entropy and logarithms are natural. Under the stated Gaussian assumptions, the covariance bounds place a fixed upper limit on $h(H_\theta\mid Y)$ and a fixed lower limit on $h(H_\theta\mid E^s)$. Both conditional entropies may change during optimization, but their difference remains bounded. This is why the same information bound holds across the parameter states satisfying these conditions.
AuthorsOpenReview ↗

Request for Further Feedback

Dear Reviewer,

Thank you again for your time and feedback. We would be grateful if you could let us know whether our response has addressed your concerns and whether there are any remaining questions we could help clarify.

Thank you for the detailed response.

W1: The fundamental advantage of MarRec over AlphaFuse remains unclear, especially since the authors acknowledge that "the reported experiments do not establish empirical superiority over AlphaFuse."

W2: The additional theoretical analysis still does not demonstrate that margin-based training reduces semantic mutual information more effectively than BPR, nor does it establish that the retained semantic information is preference-relevant.

Therefore, I will maintain my original rating.

Reviewer_oTU2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposed a scalable margin semantic extraction framework, MarRec, to effectively extract preference-relevant semantics by introducing a controllable light-weight margin constraint. Three margin constrains are designed, namely fixed, adaptive and soft schema, to tested on MarRec. The authors provide a theoretical analysis connecting the margin constraint to the information bottleneck principle. Extensive experimental results verify the effectiveness of the proposed method in data sparsity, long-tail effect, and cold-start settings across different benchmarks with different backbones.

Critical strengths
  1. The proposed method trys to address a relevant problem in LLM-enhanced recommendation by minimising the semantic gap with collaborative signals in recsys by introducing a margin-based mechanism.
  2. The proposed method is reletively simple and model-agnostic, which can be easily applied to most existing recsys
  3. The empirical analysis shows promising results on effectiveness and efficiency of different backbones.
Critical weaknesses
  1. As the main idea of the proposed method is to learning a more informative user/item embedding with margin constrain as the initialisation of the stage 2, and the use of margins, to some extent, limits the learning of outliers. Then other initialisation strategies should also be experimented. For instance, a semantic initialization baseline using the same projected embeddings and initialization procedure should also be tested.
  2. Experiments covers cold-start item scenario but lack of cold-start users.
  3. Since margin settings play a crucial role in the proposed method, and according to Table 2, margins vary across different domains, will this method have an impact on cross-domain recommendations? What about new domain recommendation?
Questions for authors

My questions has been expressed in weaknesses.

2 comments

We sincerely thank the reviewer for the thoughtful comments and helpful suggestions.

Weaknesses 1

The existing w/o Margin ablation in Section 5.2 provides a matched semantic-initialization control. This variant uses the same frozen semantic inputs, MoE projection architecture, recommender backbone, and Stage-II initialization and training procedure as MarRec. The controlled change is that Stage I uses the standard BPR objective instead of the margin objective. Consequently, both methods transfer learned semantic projections into ID embeddings.

Figure 2 and the supplementary NDCG@20 ablation results report w/o Margin among the variants with the largest performance degradation. Since this variant retains semantic pretraining and the Stage-II transfer procedure, the comparison supports the contribution of margin-based training beyond semantic initialization alone.

As complementary evidence, the appendix's Warm-Start and Semantic-Input Analysis examines semantic content and generic initialization effects. The existing LightGCN Recall@20 results are:

Initialization / input controlBeautySportsSteamYelp
Base0.15510.08450.13570.1061
ID-Standard0.15440.08460.13590.1065
ID-Margin0.15940.08630.13670.1037
Rand-Margin-MoE-Frozen0.15060.08090.13420.1088
MarRec0.18290.11420.14630.1273

Taken together, these experiments examine complementary factors: w/o Margin evaluates the Stage-I objective under the same semantic-initialization pipeline, ID-Standard and ID-Margin examine generic ID warm-start, and Rand-Margin-MoE-Frozen evaluates the contribution of semantic content under the same projection and transfer procedure.

Weaknesses 2

Our focus on item cold-start follows the way semantic inputs are constructed in these datasets. Following prior work, we derive each user's semantic representation by aggregating the semantic embeddings of items they interacted with in the training set. Item semantic representations are obtained from item-side text and remain available even when the items have no training interactions.

For a strictly zero-history user, there are no training interactions from which to construct this semantic representation. The public datasets used here also do not provide independent personal information possibly due to privacy concerns. Given the information available in these datasets, an LLM-generated personalized user profile would likewise have to be inferred from interacted items. Using validation or test interactions for this purpose would expose the held-out preferences and introduce data leakage. This motivates our focus on item cold-start, where item-side text provides semantic inputs without using held-out interactions.

We additionally examine users with limited interaction histories in the Data Sparsity and Long-Tail analyses (Sections 5.3 and 5.4). These experiments assess the value of semantic information when user-side collaborative evidence is scarce. Together, the evaluations cover unseen items and sparse user histories under a consistent protocol that constructs user semantics exclusively from training interactions.

Weaknesses 3

Different optimal margins across datasets do not necessarily imply strong sensitivity to the margin. Table 2 reports the best-performing values, while the appendix's Hyperparameters Analysis directly examines robustness to this choice. The fixed-margin MarRec achieves similar performance across the tested margin values in most dataset-backbone settings. These results show that its effectiveness is not generally restricted to one precisely tuned margin, supporting its practical applicability across datasets with different characteristics.

Our two variants, Adaptive and Soft, automatically generate instance-specific margins. Section 5.6 further shows that their optimal α\alpha and β\beta are relatively similar across datasets under the same backbone, reducing the tuning effort.

For a new domain with available training and validation data, the same procedure can be applied using fixed-margin selection or either automatic margin variant. Cross-domain transfer is a promising extension, and we will investigate it in future work, including how the margin strategies behave under source-target preference shifts and how effectively they support adaptation to new domains. The present hyperparameter analysis and automatic margin mechanisms provide a practical basis for this investigation.

AuthorsOpenReview ↗

Request for Further Feedback

Dear Reviewer,

Thank you again for your time and feedback. We would be grateful if you could let us know whether our response has addressed your concerns and whether there are any remaining questions we could help clarify.

Comments on the paper

3 comments

We appreciate the reviewer's careful assessment and constructive feedback on our work.

Weakness 1 and Question 1

We provide a proof under Gaussian representation modeling, a common choice in deep learning, together with the explicit conditions below. Building on Appendix A.1, we state the shared-covariance and moment conditions needed for the mutual-information bound. All logarithms are natural.

Fix a parameter state and a margin m>0m>0, and consider a joint population distribution of (Es,H1,Y)(E^s,H^1,Y). Suppose

H1∣Y=y∼N(μy,Σ),Σ≻0,H^1\mid Y=y\sim\mathcal N(\mu_y,\Sigma),\qquad\Sigma\succ0,

where the within-label covariance is shared across labels. Assume that μY\mu_Y has finite second moments, and define μˉ=EYμY\bar\mu=\mathbb E_Y\mu_Y and B=Cov⁡Y(μY)B=\operatorname{Cov}_Y(\mu_Y). Impose the following additional restriction:

=\mathbb E_Y\!\left[(\mu_Y-\bar\mu)^\top\Sigma^{-1}(\mu_Y-\bar\mu)\right]\leq2m.$$ This condition bounds the between-label mean variation relative to the within-label covariance. In the isotropic case $\Sigma=\sigma^2 I_d$, it reduces to $\mathbb E_Y\|\mu_Y-\bar\mu\|^2\leq2m\sigma^2$. Take the label-independent reference distribution $Q=\mathcal N(\bar\mu,\Sigma)$. The KL decomposition and the KL formula for Gaussians with equal covariance yield $$\begin{aligned} I(H^1;Y) &=\mathbb E_Y D_{\mathrm{KL}}(P_{H^1\mid Y}\|Q)-D_{\mathrm{KL}}(P_{H^1}\|Q)\\ &\leq\mathbb E_Y D_{\mathrm{KL}}(P_{H^1\mid Y}\|Q)\\ &=\tfrac12\operatorname{tr}(\Sigma^{-1}B)\leq m. \end{aligned}$$ Thus, when the relevant entropies are finite, the entropy identity in Appendix A.1 yields the following conditional upper bound: $$\begin{aligned} I(E^s;H^1) &=I(H^1;Y)+H(H^1\mid Y)-H(H^1\mid E^s)\\ &\leq m+H(H^1\mid Y)-H(H^1\mid E^s)=m'. \end{aligned}$$ For the appendix's further Gaussian specialization, assume $\Sigma\preceq\sigma_y^2 I_d$ and $H^1\mid E^s=e\sim\mathcal N(\mu_e,\Sigma_e)$ with $\Sigma_e\succeq\sigma_e^2 I_d$, where $0<\sigma_e^2$ and $\sigma_y^2<\infty$. These nondegenerate conditional-distribution assumptions give $$I(E^s;H^1)\leq m+\frac d2\log\frac{\sigma_y^2}{\sigma_e^2}.$$ The bound is uniform over parameter states if the moment and covariance conditions hold with the same constants throughout the region considered. Both conditional entropies may change within these bounds. This is an information-capacity bound on the specified representation family. This Gaussian specialization characterizes a probabilistic representation model with nondegenerate conditional variability. Equation 5 gives $[m-\Delta]_+=m-\min(\Delta,m)$, so a triplet's direct main-loss gradient vanishes for $\Delta>m$. The moment restriction above is an explicit additional modeling condition. Hinge saturation alone does not enforce it. Consequently, the calculation establishes the mutual-information upper bound under the stated distributional conditions, without assuming that a small hinge loss verifies those conditions. Copying the selected projected embeddings and backbone state preserves $H^2_{\mathrm{init}}=H^1_{\mathrm{selected}}$ when the graph and deterministic forward settings match. For the subsequent information-increase statement, consider a fixed population distribution with $H(Y)<\infty$, and a normalized conditional predictor $q_\theta(Y\mid H^2_\theta)$ for the same target $Y$ as in the mutual-information expression. If this predictor equals the full true posterior at each parameter state being compared, then $$\mathcal L_{\mathrm{pred}}(\theta) :=\mathbb E[-\log q_\theta(Y\mid H^2_\theta)] =H(Y\mid H^2_\theta),\qquad I(H^2_\theta;Y)=H(Y)-\mathcal L_{\mathrm{pred}}(\theta).$$ A decrease in this population prediction loss therefore increases task mutual information. For BPR, this conclusion applies when its preference likelihood is the conditional likelihood of the same target $Y$, under the posterior-matching assumption above. The required decrease is in the expected BPR main loss. Under the stated assumptions, $I(E^s;H^1)\leq m'$, and hence $$\min\{I(E^s;H^1),m'\}-I(H^2;Y) =I(E^s;H^1)-I(H^2;Y).$$ The first term measures the semantic-input information retained in the Stage-I representation, while the second measures the task information contained in the Stage-II representation. Under the stated assumptions, these two terms define a two-stage information criterion motivated by the information bottleneck principle.

Weakness 2 and Question 2

The cold-start evaluation follows the item-disjoint positive-interaction split in Appendix B.4: held-out items have no observed positive training interactions, while their item IDs and text remain available.

Negatives are sampled uniformly from the full item set, excluding only each user's observed training positives. Held-out items can therefore be sampled as negatives in both stages. The sampler does not consult validation or test interaction labels. For graph-based backbones, the interaction graph is constructed from training positives. Held-out items remain graph nodes but have no training user-item edges, and sampled negatives do not add graph edges.

Following Section 3.4 and Appendix B.2, the selected projector initializes the Stage-II ID embeddings as eiII,0=gθ⋆(eis)e_i^{\mathrm{II},0}=g_{\theta^\star}(e_i^s) for every item, including held-out items. Any additional trainable backbone parameters are transferred from the same Stage-I checkpoint. In Stage I, the original semantic embeddings are frozen, but sampled cold negatives can contribute gradients to the shared projector and backbone. In Stage II, their trainable ID embeddings can receive ranking-loss and applicable regularization updates when sampled. Auxiliary losses follow the stage-specific settings described in the paper.

All candidate items, including held-out items, are scored using the same backbone and inner-product scoring function defined in Section 3.4. Evaluation ranks the full item set after masking each user's observed training positives.

Base and RLMRec-Con follow the same split, negative sampling, and full-item ranking protocol as MarRec, using their standard initialization and training procedures.

Base relies on ID embeddings learned from collaborative supervision. RLMRec-Con additionally uses semantic embeddings through contrastive alignment with collaborative outputs (Appendix B.3). MarRec's semantic initialization provides a task-adapted starting representation even for items without positive training interactions. The cold-start results support the effectiveness of this semantic transfer and subsequent recommendation training in ranking held-out items.

Question 3

The adaptive-margin network is jointly optimized with the semantic projector and trainable backbone through Equation 8, without a separate outer or validation-based optimization objective. Our use of meta-learning is motivated by prior work on learning components of the training criterion, such as Meta-Weight-Net (Meta-Weight-Net: Learning an Explicit Mapping For Sample Weighting, NeurIPS 2019), which learns a mapping from sample losses to training weights. In our formulation, the learned component is an instance-specific margin: a network maps the current user, positive-item, and negative-item representations to this threshold. The margin determines hinge activation and thereby controls each triplet's direct score-gap gradient during semantic adaptation.

The gradient paths follow directly from Equation 8. For a triplet, write ℓ=[m−Δ]++αe−m\ell=[m-\Delta]_++\alpha e^{-m}, where Δ=s(u,i+)−s(u,i−)\Delta=s(u,i^+)-s(u,i^-), and let a=1{Δ<m}a=\mathbf1\{\Delta<m\}. Away from the hinge boundary,

\qquad \frac{\partial\ell}{\partial m}=a-\alpha e^{-m}.$$ The margin-network parameters receive gradients through $m$. The representations supplied to the margin network are not detached. Therefore, gradients propagate to the projector and backbone through both the score difference and the predicted margin. For the projector and trainable backbone parameters $\theta$, the total gradient of the triplet loss is $$\nabla_\theta\ell =-a\nabla_\theta\Delta +(a-\alpha e^{-m})\nabla_\theta m.$$ These parameters additionally receive the auxiliary-loss gradients in Equation 16. The original semantic embeddings are frozen, as specified in Section 3.2, while the projected and backbone representations remain differentiable. For $\alpha>0$, the exponential penalty discourages margin collapse. When the hinge term is inactive, this penalty can still update the margin network and, through its input representations, the projector and backbone.
AuthorsOpenReview ↗

Request for Further Feedback

Dear Reviewer,

Thank you again for your time and feedback. We would be grateful if you could let us know whether our response has addressed your concerns and whether there are any remaining questions we could help clarify.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.