acceptodds
Under review as a conference paper at ICLR 2027

HI-CLIP: Characterizing Functional Interactions among Attention Heads in CLIP

Abstract

As CLIP increasingly serves as the visual backbone of multimodal models, understanding its internal visual representations is important for building reliable and controllable models. Existing interpretability studies primarily focus on what concepts individual attention heads encode or what functional roles they play, while the functional relationships among heads remain underexplored, limiting the reliability of head-level interventions such as pruning and editing. Therefore, we propose HI-CLIP, a framework for interpreting inter-head functional relationships in CLIP-ViT. HI-CLIP identifies criterion-specific expert and collaborative heads, while characterizing inter-head relationships as supportive, redundant, or interfering. Experiments on ViT-B/16 and ViT-L/14 show that HI-CLIP reveals how criterion-specific visual representations are organized and formed through head interactions. Moreover, the discovered head roles and interactions improve customized retrieval without additional training and guide parameter-efficient adaptation. Overall, HI-CLIP provides a relational perspective for understanding how attention heads interact in CLIP, while demonstrating practical utility for customized retrieval and targeted fine-tuning.

Primary area:
interpretability and explainable AI
Keywords:
CLIP Interpretability, Attention Head Interactions, Mechanism Interpretability

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

This paper proposes HI-CLIP, a training-free framework for analyzing functional interactions among attention heads in CLIP vision transformers. The method first identifies criterion-specific expert and collaborative heads based on alignment between head representations and CLIP text embeddings. It then analyzes relationships between preceding heads and selected target heads by combining SVD-based representational subspace similarity with mean-ablation interventions. The resulting relationships are categorized as supportive, redundant, interfering, or non-interacting.

Critical strengths

Understanding relationships among attention heads is an important direction in mechanistic interpretability. The distinction between representational overlap and functional dependence is conceptually useful. Combining representation–text alignment, principal-angle subspace similarity, and mean ablation provides an intuitive analysis pipeline without training an additional interpretation model.

Critical weaknesses
  1. The central contribution is the characterization of supportive, redundant, and interfering relationships among attention heads. However, the definitions and experiments do not convincingly establish that the proposed classifications correspond to the claimed functional mechanisms. In particular, the relationship categories are largely determined by the effect of ablating a candidate head on the representation quality of a target head:
  • Supportive: removing the candidate decreases target representation quality.
  • Interfering: removing the candidate improves target representation quality.
  • Redundant: removing the candidate has little effect, but the two heads exhibit high representational overlap. While these definitions provide an operational categorization of intervention outcomes, they do not necessarily establish the underlying mechanisms claimed by the authors.
  1. The main experimental evidence for inter-head functional relationships is Table 2, which presents only three representative candidate–target head pairs corresponding to supportive, interfering, and redundant relationships. However, the reported intervention results are closely aligned with how the relationships are defined in Equation 11. Since the classifications themselves depend on the effect of ablation on target representation quality, observing corresponding performance changes after ablation does not constitute independent validation of the identified functional roles. It is therefore unclear whether the experiment verifies the proposed characterization or simply illustrates the criteria used to assign relationship labels.
Questions for authors

See weakness, and what additional benefit does inter-head interaction characterization provide beyond identifying criterion-specific heads?

1 comment
AuthorsOpenReview ↗

HI-CLIP combines representational analysis and causal intervention to characterize transferable inter-head functional relationships, supported by independent cross-dataset validation.

We thank the reviewer for the detailed and thoughtful comments. We clarify our methodology, present additional cross-dataset experiments, and explain the benefits of relationship characterization.

  1. HI-CLIP does not identify all relationships through causal ablation. Redundancy is determined by SVD-based subspace similarity, while ablation distinguishes supportive and interfering relationships.

Our framework follows three stages:

  • Expert/Collaborative heads: Representation–text alignment identifies heads covering most categories or complementary category subsets.
  • Redundant relationships: SVD-based subspace similarity identifies representational overlap, independent of ablation.
  • Supportive/Interfering relationships: Causal ablation measures positive or negative influences on target representation quality after excluding redundant heads.

Small ablation effects provide supporting evidence for redundancy rather than defining it. We also explored information communication analysis inspired by Talking Heads (Merullo et al., NeurIPS 2024). However, both supportive and interfering heads can exhibit strong communication, making communication strength insufficient to distinguish their effects. Likewise, subspace similarity measures overlap but not the direction of influence.

Thus, HI-CLIP combines representational overlap and causal intervention to characterize inter-head relationships. While this does not fully explain information-transfer mechanisms, it provides quantitative evidence from representations and interventions.

  1. We conducted additional LAS&T experiments to independently validate cross-dataset transferability.

We agree that Table 2 illustrates our characterization criteria rather than providing independent validation. We therefore evaluated the identified relationships on LAS&T (Eppel et al., CVPR 2026), directly transferring them from the original calibration data without head re-identification, repeated characterization, or training.

Across nine evaluation conditions, ablating interfering, supportive, and redundant heads produces relative mean performance changes of +0.756%, −1.330%, and +0.047% on ViT-B/16, and +0.773%, −4.051%, and +0.038% on ViT-L/14, respectively. In addition, compared with HI-CLIP, DAHIVT yields smaller interfering-head ablation gains (+0.16%/+0.26%) and less consistent redundant-head ablation effects (−1.00%/+0.31%).

These results provide independent evidence of transferable functional relationships, beyond the original characterization interventions. Although the underlying mechanisms remain unverified, the relationships retain their expected effects on unseen data.

  1. Relationship characterization provides intervention guidance beyond identifying criterion-specific heads.

Expert and collaborative head identification reveals where criterion-specific information is encoded, while relationship characterization reveals how other heads influence these representations.

For example, identifying a color expert does not reveal which preceding heads support, interfere with, or overlap with its representation.

HI-CLIP provides this information:

  • Supportive: Preserve heads to avoid performance degradation.
  • Interfering: Ablate or prune heads to improve performance.
  • Redundant: Identify representational overlap and potential parameter redundancy.

Our LAS&T experiments demonstrate this utility: Removing previously identified interfering heads improves mean performance by 0.756%/0.773% without training, while supportive-head removal degrades performance and redundant-head removal has negligible effects.

Moreover, relationships can be characterized once per model and criterion, then reused across datasets, avoiding repeated characterization.

Thus, HI-CLIP extends head identification to transferable relationship characterization, providing guidance for targeted interventions. The additional LAS&T experiments support both its generalizability and practical utility.

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

CLIP is the visual backbone of many multimodal systems, but its internal visual representations are still not well understood. Prior interpretability work asks what individual attention heads encode or what role they play. This paper asks how heads interact to form criterion-specific representations. HI-CLIP is a training-free framework with three stages. First, it identifies expert and collaborative heads for a visual criterion by aligning each head's representation with CLIP text embeddings of the criterion's categories, using coverage thresholds τe=0.7 and τc=0.8. Second, it characterizes inter-head relations by combining SVD subspace similarity through principal angles with a causal mean ablation, and labels relations as supportive, redundant or interfering. Third, it uses the selected heads for customized retrieval and for placing LoRA adapters. Experiments use Clevr-4-10k and CLIP ViT-B/16 and ViT-L/14. The paper reports clustering ACC, NMI and ARI, t-SNE and saliency visualizations, a three-pair ablation, retrieval Hit@1 and Hit@5, and Targeted LoRA versus Random LoRA over five epochs.

Critical strengths
  1. The relational framing is a genuine step beyond single-head interpretation. The expert and collaborative distinction, together with the supportive, redundant and interfering vocabulary, gives the paper a structural account of how criterion-specific information is organized. This matters because it turns head analysis into a guide for interventions rather than a catalogue of roles.

  2. The empirical gains are large and the applications are concrete. On ViT-B/16, texture ACC rises from 17.71 with the DAHIVT expert baseline to 68.63. On ViT-L/14, color ACC rises from 43.73 to 97.58. The selected heads reach Hit@1 of 0.9972 for color, 0.8870 for texture and 0.9981 for shape, and Targeted LoRA converges faster than Random LoRA. These results make the practical utility credible.

Critical weaknesses
  1. The evidence is narrower than the framing suggests. The experiments use one synthetic dataset, Clevr-4-10k, two CLIP backbones and three low-level criteria, namely color, texture and shape. The near-perfect retrieval on this setup does not establish real-image or real-retrieval utility. The paper frames the method as a general account of CLIP-ViT head interactions, but it does not test real images, other CLIP variants or semantic, spatial and relational criteria. The main text also reports single numbers, with no seeds, error bars or significance tests, so the large gaps are persuasive but not statistically anchored. This is why I land on a weak rejection rather than a weak acceptance.

  2. The baselines and controls leave some questions open. CHG is discussed as a close role-based framework but is not compared empirically. The thresholds τe=0.7 and τc=0.8, the top-K choice, the category vocabulary and the text prompt template are not varied or ablated. The paper also does not state how many heads the Random and DAHIVT baselines use, so it is unclear whether the clustering comparison is matched. Table 2 ablates only three representative head pairs, which supports the vocabulary but does not establish that the three relation types are stable across heads, criteria or backbones. The submitted PDF does not mention a code release or a reproducibility statement. It would help to add a CHG comparison, a matched-head-count baseline, a threshold and prompt sweep, a systematic pair analysis and a code release.

  3. The applications do not isolate the contribution of the inter-head relation analysis. The heads used for customized retrieval and Targeted LoRA come from the expert and collaborative identification stage. The supportive, redundant and interfering labels are quantitatively validated only by the three illustrative pairs in Table 2, and there is no human or quantitative faithfulness check of the labels. The quantitative gains therefore show that criterion-specific head selection works, but they do not show that the relational analysis improves any downstream metric beyond that selection. Because the relational perspective is the paper's central claim, this gap matters. It would help to compare heads selected with and without the relation analysis, or to use the relation labels to guide a pruning or LoRA intervention and report the downstream effect.

  4. The relation taxonomy is tested only with the mean-ablation choice. The paper replaces a candidate head's contribution with its calibration-set mean representation and reads the sign of the intervention effect as supportive or interfering. Other ablation choices, such as zero ablation or resampling, can produce different effects, and the paper does not compare them. The paper also does not state how large an effect counts as "little change" for redundancy, does not report the size of the calibration set, and does not say how the label is assigned when the subspace and ablation signals disagree. It would help to report the ablation variants, the size of the calibration set and the decision rule for the three labels.

Questions for authors
  1. How sensitive are the expert and collaborative decisions, and the downstream clustering and retrieval results, to τe, τc and K? Please report a sweep and the per-criterion head counts.

  2. Why is CHG not compared, given that it is a close role-based framework? A comparison on the same clustering and retrieval setup would clarify the delta.

  3. Do the clustering and retrieval gains hold across seeds and on a real-image dataset or another CLIP variant? Please report variance and at least one out-of-domain evaluation.

  4. How were the three representative pairs in Table 2 selected? How many candidate-target pairs were tested, and what is the distribution of the three relation types?

  5. [Future-work suggestion] A systematic study of when the three relation types are stable, and a release of the head-selection code, would make the framework easier to reuse.

1 comment

We thank the reviewer for the comments. We acknowledge the evaluation limitations and address the concerns below.

  1. Our evaluation follows CRL (Liu et al., NeurIPS 2025), using repeated random initializations to reduce experimental variability.

For customized clustering, we follow CRL's protocol, averaging 20 K-means runs to reduce initialization effects on ACC, NMI, and ARI. We will report standard deviations and clarify the retrieval evaluation protocol.

For hyperparameters, we use predefined thresholds τe=0.7\tau_e=0.7 and τc=0.8\tau_c=0.8, with criterion-dependent head selection. Although a sensitivity sweep is unavailable, we will provide selected-head counts and selection rules for each criterion.

  1. CHG is related to HI-CLIP but differs in training requirements and interpretation objectives. CHG requires additional training to identify head roles, whereas HI-CLIP is training-free, analyzing representations and interventions within frozen CLIP.

Moreover, CHG focuses on head–task relationships, while HI-CLIP characterizes inter-head influences on criterion-specific representations.

A direct comparison therefore involves different training requirements and objectives. We will clarify these distinctions and computational costs without claiming quantitative superiority over CHG. We acknowledge that controlled comparisons would provide further evidence.

  1. We conducted additional LAS&T experiments to validate cross-dataset transferability.

To address the limited evaluation on CLEVR-4-10k, we evaluated HI-CLIP on LAS&T (Eppel et al., CVPR 2026), transferring the identified relationships without re-identification or training.

Across nine conditions, ablating interfering, supportive, and redundant heads changes mean performance by +0.756%, −1.330%, and +0.047% on ViT-B/16, and +0.773%, −4.051%, and +0.038% on ViT-L/14, respectively.

These results show consistent trends:

  • Interfering: Removal improves performance.
  • Supportive: Removal degrades performance.
  • Redundant: Removal has negligible effects.

In comparison, DAHIVT yields smaller interfering-head ablation gains (+0.16%/+0.26%) and less consistent redundant-head ablation effects (−1.00%/+0.31%).

These results support the transferability and intervention utility of the identified relationships, although broader natural-image and semantic generalization remains unverified.

We will also report standard deviations across the 20 clustering runs.

  1. Table 2 presents three representative pairs from a broader set of candidate–target relationships.

Our characterization considers candidate heads beyond the identified expert and collaborative heads:

  • Redundancy: SVD-based subspace similarity identifies representational overlap.
  • Support/Interference: Mean ablation distinguishes positive and negative influences on target representations.

Redundancy reflects information overlap rather than functional compensation. The three Table 2 pairs illustrate these relationships, not the full set of evaluated pairs. Our analysis suggests a relatively balanced distribution of the three relationship types. We will report pair counts, distribution and protocol.

We acknowledge that assignment stability under zero ablation or resampling requires further investigation.

The additional LAS&T experiments provide cross-dataset evidence beyond these representative pairs, supporting the functional effects of the identified relationships.

In summary, HI-CLIP provides a quantitative characterization of transferable inter-head relationships without claiming a complete mechanistic explanation. We will strengthen the manuscript with LAS&T results, variability statistics, and implementation details.

Reviewer_6NZLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes HI-CLIP to characterize attention-head roles and interactions in CLIP. It identifies criterion-specific expert and collaborative heads using representation-text alignment and category coverage, then combines dominant-subspace overlap with mean-ablation effects to label preceding heads as supportive, interfering, or redundant. Experiments cover customized clustering on Clevr-4-10k with ViT-B/16 and ViT-L/14, selected head-pair interventions, criterion-based retrieval, and targeted LoRA adaptation.

Critical strengths

The paper addresses a useful interpretability question: how preceding heads affect a criterion-specific target representation. Separating representational overlap from functional dependence is a sensible design choice. Table 1 reports improvements across all three criteria and both backbones; for example, ViT-B/16 texture ACC increases from 17.71% for the stronger DAHIVT variant to 68.63%. Retrieval demonstrates a practical use of selected heads without additional training. The targeted-versus-random LoRA comparison explicitly matches the number of adapted heads, parameter budget, data split, optimization setting, and training duration, providing useful initial evidence for head selection.

Critical weaknesses
  1. Essential method definitions and evaluation protocols are missing. Equation (2) does not specify how an individual head representation is extracted and mapped into CLIP's text-embedding space. The quality measure Q_t, retained SVD rank k, collaboration size K, and relationship thresholds tau and tau_s are unspecified. Equation (3) uses category-labelled images, but the separation between head selection/calibration and final evaluation is not described. Disjoint query/gallery splits in Section 5.1 do not establish that head selection excludes evaluation labels. This also qualifies the claimed distinction between intrinsic and task-conditioned analysis. These omissions prevent reproduction and make comparison fairness difficult to assess.
  2. The central relationship analysis lacks independent, broad validation. Equation (11) assigns roles using ablation effects, while Section 4.3 illustrates those effects for only three pairs sharing one target, without explaining a separate validation protocol. High overlap and a small individual ablation effect alone do not establish compensation by alternative pathways. In Table 2, the redundant and interfering examples improve ACC by 0.64 and 0.91 percentage points, respectively; their separation needs the missing threshold and uncertainty estimates. Table 2's initial ACC/NMI/ARI of 93.32/91.41/86.99 also exactly match Table 1's DAHIVT Expert color results, whereas HI-CLIP reports 95.68/93.38/90.95. Since Table 3 identifies the same target, L11H10, as HI-CLIP's color head, the configurations require reconciliation.
  3. The applications do not isolate the value of inter-head relationship characterization. Sections 5.1 and 5.2 use expert/collaborative heads but do not explain how the supportive, redundant, and interfering relationships change retrieval or LoRA placement. Table 3 reports only HI-CLIP, so it does not establish retrieval improvements over CLIP or DAHIVT. Figure 6 compares only random LoRA placement, without an alignment-only or DAHIVT-guided control or variability across random selections. Quantitative validation is concentrated on three Clevr-4 criteria; the natural-image illustrations do not establish broader generalization.
Questions for authors
  1. Please define the head extraction/projection, multi-head aggregation, QtQ_t, KK, kk, tautau, and taustau_s, including how they are selected. What are the head-selection, calibration, clustering, query/gallery, and LoRA train/validation splits and sample counts? Are evaluation images or labels used for selection? Please specify the clustering procedure and whether DAHIVT receives comparable supervision and head budgets.
  2. Can you report relationship validation on held-out samples across more target heads and both backbones, including overlap scores and uncertainty across clustering seeds? What evidence distinguishes compensable redundancy from a head that is simply unnecessary for the target, for example through joint ablation or restoration? Please reconcile Tables 1-3 and explain the threshold separating the two positive ACC changes in Table 2.
  3. How are the inferred relationships actually used in the applications? Please compare selection from Section 3.1 alone against a variant that uses Section 3.2, alongside CLIP/DAHIVT retrieval and matched-budget LoRA controls. Please provide the LoRA objective, adapted projections, rank, learning rate, and variability across runs. Clear implementation details, independent relationship validation, and matched controls could materially change my assessment.
1 comment

We thank the reviewer for the comments. We have conducted additional experiments and will clarify the methodology and limitations.

  1. We conducted additional LAS&T experiments to validate the transferability of the identified relationships.

We agree that Table 2 illustrates our characterization criteria rather than providing independent validation. We therefore transferred the identified relationships to LAS&T (Eppel et al., CVPR 2026) without head re-identification or training.

Across nine conditions, ablating interfering, supportive, and redundant heads changes mean performance by +0.756%, −1.330%, and +0.047% on ViT-B/16, and +0.773%, −4.051%, and +0.038% on ViT-L/14, respectively. In comparison, DAHIVT yields smaller improvements from interfering-head ablation (+0.16%/+0.26%) and less consistent effects from redundant-head ablation (−1.00%/+0.31%). These results support the cross-dataset transferability of our relationships, although they do not establish the underlying mechanisms.

  1. We clarify that redundancy refers to representational overlap, not functional compensation.

HI-CLIP uses two criteria:

  • Redundancy: SVD-based subspace similarity measures representational overlap, independent of ablation.
  • Support/Interference: Causal ablation measures positive or negative effects on target representation quality.

Thus, small ablation effects support, but do not define, redundancy. Functional compensation would require joint ablation or restoration.

This also explains Table 2: positive ACC changes may occur for both redundant and interfering heads, but their classification depends on subspace similarity and target representation–text alignment, respectively, not ACC changes.

  1. We will clarify methodological details and experimental protocols.

The key specifications are:

  • Extraction: Each head's representation is extracted from its CLS-token contribution after attention output projection.
  • Mapping: We use full-pass LayerNorm statistics and CLIP's pretrained visual projection.
  • Aggregation: Selected-head representations are L2-normalized, concatenated, and normalized again.
  • Thresholds: SVD rank k=10k=10, subspace threshold τs=0.1\tau_s=0.1, and intervention threshold τ=0.005\tau=0.005.
  • Definitions: KK denotes category coverage size; QtQ_t measures mean target representation–text alignment.
  • Head selection: We use all 10,531 CLEVR-4-10k images and category labels without training.
  • Clustering: K-means uses one labeled initialization sample per class and 20 runs, reporting ACC, NMI, and ARI.
  • Data usage: Head selection and clustering share the same dataset; intrinsic refers to pretrained representations, not label-free selection.
  • DAHIVT: We follow its caption-based protocol and match head budgets where applicable.
  • LoRA: We train Q/V adapters on 8,424 images and validate on 2,107, freezing CLIP.
  • Optimization: Cross-entropy, AdamW, learning rate (10−310^{-3}), weight decay (10−410^{-4}), and five epochs.

We also acknowledge the Table 1–2 inconsistency caused by different reporting protocols (20-run average versus single run). We will recompute the results using a consistent protocol.

  1. Relationship characterization provides intervention guidance beyond criterion-specific head identification.

We agree that the original retrieval and LoRA experiments primarily validate head identification rather than relationship characterization.

However, head identification locates criterion-specific information, while relationship characterization reveals functional influences between heads.

Our LAS&T experiments validate this utility: interfering-head ablation improves performance by 0.756%/0.773% without training, while supportive-head ablation degrades performance and redundant-head ablation has negligible effects.

Moreover, identified relationships can be reused across datasets without repeated characterization. We acknowledge that alignment-only and DAHIVT-guided LoRA comparisons, along with random-selection variance, would provide stronger application-specific evidence. We will clarify this limitation and distinguish the contributions of head identification and relationship characterization.

In summary, HI-CLIP identifies transferable inter-head relationships supported by our additional LAS&T experiments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.