acceptodds
Withdrawn from ICLR 2027

Uncovering Separable and Composable Skill Mechanisms in Vision-Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent work has sought to localize the internal mechanisms underlying visual reasoning skills in vision-language models. We posit two important desiderata, separability and composability, for skill localization. Separability requires identified mechanisms to contribute primarily to their target skill while having limited effects on other skills, enabling precise diagnosis and targeted intervention. Composability requires skill mechanisms to be combinable to support complex tasks, allowing a small set of localized skill mechanisms to account for broader capabilities. However, existing methods overlook separability and composability in both mechanism identification and evaluation. To address these limitations, we propose a multi-skill interpretability framework DISentangled and COmposable Skill Mechanisms (DISCO) for identifying and evaluating separable and composable attention heads. For mechanism identification, we propose a novel technique called Separable Head Selection, which selects skill-specific heads based on their contributions across skills. For evaluation, we are the first to design metrics for evaluating separability and composability of identified heads. Experiments on three VLMs and four fundamental visual reasoning skills show that the heads identified by DISCO achieve the highest overall separability on single-skill tasks, outperforming the strongest baseline by 22.75, 10.95, and 9.96 points across the three models, while also achieving the highest average composability on multi-skill tasks across all models.

Primary area:
interpretability and explainable AI
Keywords:
Vision-Language Models, Mechanistic Interpretability, Skill Mechanisms

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

This paper introduces DISCO, a framework for identifying skill-specific attention heads in vision-language models. The method estimates head contributions through counterfactual activation patching across different answer formats and image styles, then selects heads based on their relative contributions across four visual reasoning skills. The authors introduce separability and composability metrics and evaluate the selected heads on both single-skill and two-skill tasks across three VLMs.

Critical strengths

The paper addresses a worthwhile question: identifying components that matter for a task is insufficient to establish that they are specific to the underlying skill. Explicitly considering effects on other skills is a useful direction. I also appreciate the distinction between preceding and final token positions, the question-level identification/evaluation split, and the detailed intervention results in the appendix. The selection ablations provide encouraging evidence that cross-skill information helps identify more selective head sets.

Critical weaknesses
  1. My main concern is the composability metric does not establish the property that the paper claims to measure. Equations (9)–(11) compare the joint intervention effect against averages of individual or mixed-random intervention effects. A positive score does not require both skill-specific head sets to contribute. For example, suppose ablating (H_A) changes 50% of originally correct predictions, ablating (H_B) has no effect, and ablating their union still changes 50%. If the random controls have no effect, both the absolute and relative composability scores are 25 percentage points, even though (H_B) contributes nothing. Conversely, if two components are both necessary and removing either one already causes failure, joint ablation may add no further effect. I would therefore be cautious about interpreting these measurements as evidence of compositional computation. Controlled interventions that independently change the two task factors, or restoration experiments linking each head set to its expected subcomputation, would provide more convincing evidence.

  2. The proposed scoring procedure averages effects across several surface forms, but averaging does not ensure that an individual head contributes consistently across them. A head with a strong effect in one format can still receive a high pooled score. Filtering for originally correct predictions may also change the relative representation of each surface form.

  3. Subtracting a random-head baseline is a reasonable attempt to account for differences in task sensitivity. However, the resulting score measures selectivity relative to that baseline. In this case, it does not directly establish that the selected heads have small absolute effects on non-target tasks.

Questions for authors

See Weaknesses

3 comments

We thank the reviewer for recognizing that our paper addresses a worthwhile question and for appreciating our separable head selection method, the distinction between different token positions, and the selection ablations. Below, we address the reviewer’s concerns. All the experimental results reported below have been incorporated into the latest version of the paper.

Q1. The composability metric does not establish the property that the paper claims to measure...if two components are both necessary and removing either one already causes failure, joint ablation may add no further effect.

A1. We thank the reviewer for pointing out the limitations of our original composability metrics. We revise Equations (9)–(10) to define composability as the joint intervention effect minus the maximum of individual or mixed-random intervention effects. Under this definition, a positive score requires heads of both skills to make contributions. In the reviewer's example, under the new metric, the composability score becomes 0 instead of 25. As shown in the tables below, our method still achieves the highest average composability across all three models (7.82, 9.68, and 3.90), substantially outperforming the strongest baselines (1.81, 2.48, and 2.71). These results demonstrate that our method identifies more composable attention heads.

In practice, disrupting the heads for one skill alone usually has limited effects rather than causing the corresponding multi-skill task to fail completely. For example, heads that contribute to Skill A but primarily support other skills are not selected for Skill A. However, these heads may still serve as weak substitutes, partially supporting Skill A when heads that we identified are disrupted.

Moreover, even if disrupting the heads for one skill causes complete failure on the corresponding multi-skill task, there are two possible cases. First, these heads may support necessary basic functions shared across skills (e.g., output formatting), as existing methods only focus on heads with the largest effects on the target skill. Such heads do not reflect the skill composability we aim to identify. Therefore, assigning them a low score under our metric is reasonable. Second, these heads may genuinely support the distinctive function of the target skill, which is the case the reviewer is concerned about. As shown in Table 2, the baselines achieve very low separability, which suggests the first case is true because basic functional heads have substantial effects on multiple skills.

On Qwen2.5-VL:

MethodPosition+Size ComposabilityCounting+Position ComposabilityDistance+Size ComposabilityAverage Composability
VFL-0.43-2.35-0.93-1.24
V-SEAM0.000.000.000.00
NOTICE0.271.173.991.81
Ours13.589.080.797.82

On Qwen3-VL:

MethodPosition+Size ComposabilityCounting+Position ComposabilityDistance+Size ComposabilityAverage Composability
VFL0.00-2.81-0.80-1.20
V-SEAM-4.82-1.98-1.78-2.86
NOTICE0.00-0.057.502.48
Ours12.254.2812.509.68

On Gemma4:

MethodPosition+Size ComposabilityCounting+Position ComposabilityDistance+Size ComposabilityAverage Composability
VFL1.91-2.15-5.29-1.84
V-SEAM-6.07-0.20-6.59-4.29
NOTICE-2.581.399.342.71
Ours3.29-1.469.893.90

Q2. The proposed scoring procedure averages effects across several surface forms, but averaging does not ensure that an individual head contributes consistently across them.

A2. We agree with the reviewer that simply averaging may mask variance. That is, a head may be important for some surface forms but not others and still gets identified.

That is exactly why we introduced the coverage mechanism in head identification (Lines 228–238). Coverage is defined as the proportion of instances for which the head receives a positive score, i.e., the head contributes to that instance. The final score is the product of the average score and coverage. In this way, a head can only be identified if it has a high average score and contributes to many data points, which employ different surface forms.

Q3. Subtracting a random-head baseline when computing the separability score does not directly establish that the selected heads have small absolute effects on non-target tasks.

A3. Absolute effects alone may be poor metrics, since tasks may have different sensitivity to intervention on random heads. As an example, disrupting a set of heads may cause the performance to drop 50% on Task A and 70% on Task B. This may suggest the heads affect Task B more. However, due to different sensitivity, disrupting an equal number of random heads may degrade performance by 10% and 50% on the two tasks, respectively. Subtracting these random baselines yields adjusted effects of 40% and 20%, and reveals a stronger contribution to the Task A / Skill A. The larger absolute effects of these heads on Task B do not necessarily indicate greater contribution to Skill B, but rather higher sensitivity of Task B to interventions.

Thank you for your valuable and responsible review, which has been very inspiring for us.

We would appreciate it if you could let us know whether our response has addressed your concerns. We look forward to further discussion with you.

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

DISCO studies whether attention heads associated with visual skills are both selective for those skills and reusable in combined tasks. It scores heads through counterfactual activation patching across visual styles and answer formats, then selects them based on their effects across skills. Experiments on three VLMs evaluate counting, positioning, size comparison, distance comparison, and tasks combining pairs of these skills.

Critical strengths
  1. Cross-skill selection addresses specificity rather than target-task importance alone. The criterion in Section 3.3 evaluates a head’s target-skill effect relative to its effects on other studied skills. Table 4 provides a controlled comparison of selection strategies, supporting the value of cross-skill selection.
  2. The answer-format analysis reveals an important localization sensitivity. In Table 5, heads identified using Option Content obtain average separability of 40.06 under that format but 6.67 under Option Label. This is useful evidence that a localized intervention’s effect can depend strongly on the surface form, motivating cross-format evaluation.
Critical weaknesses
  1. The main separability comparison mixes evaluation conditions. Section 4.1 states that the baselines use natural photographs with Option Content for both identification and evaluation, whereas DISCO uses all surface forms. Table 2 therefore does not isolate the localization improvement under common evaluation inputs; the mismatch alone does not establish the direction of the bias. Question: Please evaluate all identified head sets on the same natural-image, Option Content subset and report the resulting comparison.
  2. Positive composability does not guarantee the stated property. Section 3.4 expects joint erasure to exceed either individual effect, but Equation (9) compares it with their average. In Table 13’s Gemma Counting + Position case, joint erasure is 58.33 versus 59.72 for Counting alone, while absolute composability remains positive at 5.56. Question: Please report joint erasure minus the stronger individual effect and align the composability interpretation with the condition actually tested.
  3. Surface-form robustness is not sufficiently established. Table 5 supports combining surface forms, but mixed-format identification still yields average separability of 32.53 on Option Content and 14.91 on Option Label. This difference does not itself prove format-specific heads, but the aggregate scores do not establish the stronger claim that the identified mechanisms reflect skill function rather than particular surface forms. Question: Please report per-form adjusted intervention effects and test the jointly identified heads on an answer format or visual style excluded from identification.
Questions for authors

see weaknesses

3 comments

We thank the reviewer for acknowledging the value of our separable head selection method and the importance of our answer-format analysis. Below, we address the reviewer’s concerns. All the experimental results reported below have been incorporated into the latest version of the paper.

Q1. The main separability comparison mixes evaluation conditions. Please evaluate all identified head sets on the same natural-image, Option Content subset.

A1. We thank the reviewer for suggesting comparison on the same surface form. The reason we chose to test different methods differently was to maximize the performance of the baselines and minimizing the performance gap, which may have led to underestimation of the strengths of our technique.

As requested, we report the separability of all methods evaluated under the same surface form (Natural images + Option Content answer format), with scores averaged over Patch and Erasure, in the tables below. Our method achieves the highest overall separability across all three models, scoring 34.74, 39.62, and 22.98 on Qwen2.5-VL, Qwen3-VL, and Gemma4, respectively, outperforming the strongest baseline by 29.09, 17.99, and 12.91 points. Furthermore, our method achieves the best results on 12 out of 15 metrics across the three models. These substantial improvements under the same single-surface-form setting further demonstrate the effectiveness of our method in identifying separable attention heads.

On Qwen2.5-VL:

MethodCounting SeparabilityPosition SeparabilitySize SeparabilityDistance SeparabilityOverall Separability
VFL-1.55-0.890.86-2.09-0.92
V-SEAM10.0016.73-16.61-14.97-1.21
NOTICE-1.225.014.1414.685.65
Ours50.9930.9723.2033.8034.74

On Qwen3-VL:

MethodCounting SeparabilityPosition SeparabilitySize SeparabilityDistance SeparabilityOverall Separability
VFL-2.25-0.39-0.71-0.20-0.89
V-SEAM19.4730.34-16.740.808.47
NOTICE22.86-0.103.8159.9421.63
Ours66.0815.3330.7546.3439.62

On Gemma4:

MethodCounting SeparabilityPosition SeparabilitySize SeparabilityDistance SeparabilityOverall Separability
VFL-0.47-6.79-0.47-2.01-2.44
V-SEAM-8.9612.88-16.33-21.50-8.48
NOTICE5.227.5710.1817.3010.07
Ours23.7512.3932.4423.3322.98

Q2. Positive composability does not guarantee the stated property. Please report joint erasure minus the stronger individual effect.

A2. We thank the reviewer for pointing out the issue with our original composability metric. Following the reviewer's suggestion, we redefine composability in Equations (9)-(10) as the joint Erasure effect minus the stronger individual Erasure effect. The results of final composability are reported in the tables below. Our method achieves the highest average composability across all three models, scoring 7.82, 9.68, and 3.90 on Qwen2.5-VL, Qwen3-VL, and Gemma4, respectively, outperforming the strongest baseline by 6.01, 7.20, and 1.19 points. Furthermore, our method achieves the best results on 10 out of 12 metrics across the three models. These results demonstrate that our method identifies more composable attention heads than baselines.

On Qwen2.5-VL:

MethodPosition+Size ComposabilityCounting+Position ComposabilityDistance+Size ComposabilityAverage Composability
VFL-0.43-2.35-0.93-1.24
V-SEAM0.000.000.000.00
NOTICE0.271.173.991.81
Ours13.589.080.797.82

On Qwen3-VL:

MethodPosition+Size ComposabilityCounting+Position ComposabilityDistance+Size ComposabilityAverage Composability
VFL0.00-2.81-0.80-1.20
V-SEAM-4.82-1.98-1.78-2.86
NOTICE0.00-0.057.502.48
Ours12.254.2812.509.68

On Gemma4:

MethodPosition+Size ComposabilityCounting+Position ComposabilityDistance+Size ComposabilityAverage Composability
VFL1.91-2.15-5.29-1.84
V-SEAM-6.07-0.20-6.59-4.29
NOTICE-2.581.399.342.71
Ours3.29-1.469.893.90

Q3. Surface-form robustness is not sufficiently established. Please report per-form adjusted intervention effects and test the jointly identified heads on an answer format or visual style excluded from identification.

A3. We thank the reviewer for pointing out the limitations of our surface-form robustness experiment. As requested, we introduce a new True or False answer format for evaluation, where the model is given a question and a candidate answer and must determine whether the candidate is correct. In the table below, we report both Adjusted Effect on target skills and Separability, both averaged across skills and over Patching and Erasure interventions.

First, single-format identification methods suffer substantial cross-format drops in separability. On Option Content, Content-only identification achieves 40.06, whereas Label-only identification achieves only 11.10 (27.7% of the best score). Similarly, on Option Label, Label-only identification achieves 18.49, whereas Content-only identification achieves only 6.67 (36.1% of the best score). In contrast, our method achieves 32.53 on Option Content and 14.91 on Option Label, reaching over 80% of the best separability under both formats. Second, on the unseen True or False format, our method also achieves the highest separability (10.78 vs. 8.10 and 8.55). These results demonstrate that our method is more robust to surface forms in identifying separable heads. Although our method does not always achieve the highest Adjusted Effect on target skills, this is consistent with our objective of identifying separable heads rather than those with the largest effects on target skills.

Effect denotes Adjusted Effect on target skills, and Sep. denotes Separability.

Format for IdentificationEffect on ContentSep. on ContentEffect on LabelSep. on LabelEffect on True/FalseSep. on True/False
Option Content47.7740.0624.246.6721.218.10
Option Label21.3111.1027.8818.4926.168.55
Content + Label (Ours)41.9732.5328.5314.9122.5910.78

Thank you for your valuable and responsible review, which has been very inspiring for us.

We would appreciate it if you could let us know whether our response has addressed your concerns. We look forward to further discussion with you.

Reviewer_zP2JOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes novel metrics for assessing the separability and composability of internal components in LLMs/VLMs (the work is applied to VLMs, but the methods apply equally well to LLMs). Separability is defined as the extent to which a particular component uniquely contributes to a particular 'skill' - a term used to refer to a particular computational ability. Composability refers to the extent to which two components combine together to perform a joint function. The separability metric is used to identify attention heads that are shown to be more separable and composable (according to these same metrics) than the heads identified by other automated methods for identifying attention heads.

Critical strengths
  • The paper is clearly written, the methods are clearly described, and the experiments are easy to understand.
  • The notions of separability and composability are well-defined, and these are interesting and well-motivated properties to look for in models.
  • An ablation study is performed to assess the importance of each component of the head selection procedure.
Critical weaknesses
  • Although the notions of separability and composability are well-motivated from a normative perspective, the paper seems to assume that attention heads in VLMs will actually be separable and composable, and that the identification of heads that aren't separable or composable reflects a weakness of interpretability method used to identify the heads. This ignores the very significant possibility that a substantial portion of the causally relevant heads for a given task may not be separable or composable. The proposed methods could be used to evaluate the question, by iteratively ablating heads according to their separability and composability scores, but this is not done in the present work.
  • It is not clear that we should really expect attention heads to be factorized at the level of the somewhat arbitrary set of skills that are explored in this work. Some of these skills, such as counting, are more complex and likely require sequential computation (chain of thought) implemented by multiple sets of heads performing distinct functions. The paper assumes that these skills should be implemented by a single group of attention heads, but the skills themselves are likely decomposable into more basic functions that would ideally be subserved by separable sets of attention heads. This dependence on pre-defined skills is a major weakness of the method.
  • The proposed methods are compared against baselines that have completely different goals and are not designed to identify separable or composable heads. It is therefore not surprising that the proposed method does a better job.
Questions for authors
  • Can the proposed method be used to determine the extent to which the causally relevant attention heads actually are separable or composable?
  • Why should the functionality of attention heads be factorized at the level of the specific proposed skills?
  • Why would we expect the baseline methods, which were designed with different goals in mind, to identify separable or composable heads?
4 comments

We thank the reviewer for recognizing the clear writing, well-defined notions, interesting motivation, and ablation study of our work. We address the concerns below.

Q1. (Part 1) Although the notions of separability and composability are well-motivated from a normative perspective, the paper seems to assume that attention heads in VLMs will actually be separable and composable, and that the identification of heads that aren't separable or composable reflects a weakness of interpretability method used to identify the heads. This ignores the very significant possibility that a substantial portion of the causally relevant heads for a given task may not be separable or composable.

Q3. The proposed methods are compared against baselines that have completely different goals and are not designed to identify separable or composable heads. Why would we expect the baseline methods, which were designed with different goals in mind, to identify separable or composable heads?

A1. We thank the reviewer for recognizing separability and composability are desirable. On the other hand, the reviewer seems to assume a dichotomy of head localization methods: (1) those identifying all heads causally relevant to a skill, and (2) those identifying separable and composable heads only.

However, existing methods do not identify all causally relevant heads; they employ a vague notion of separability though they do not explicitly articulate the idea. In the three baselines, V-SEAM (Wang et al., 2025) selects the Top-K heads that contributes the most to a skill, excluding lower-ranked heads that still contribute positive. This exclusion implicitly reflects a consideration of separability, as these lower-ranked heads may contribute more strongly to other skills. Similarly, VFL (Shi et al., 2025) selects only the layer with the strongest causal contribution to the target skill. NOTICE (Golovanevsky et al., 2025) first identifies candidate heads with high causal contributions to each skill and then removes heads that overlap across skills, implicitly favoring separability. Thus, it is reasonable to consider them as having separability elements in the design, but they did not explicitly explain the rationale behind such design and did not identify separability as the driving force.

Unlike these methods, the current paper explicitly aims for both separability and composability and proposes techniques and evaluation metrics for them. That is our novel contribution.

We are not aware of methods that explicitly attempt to identify all causally relevant heads. In practice, such attempts may be undermined by spurious correlation between head activities and the target skill.

Q1. (Part 2) Can the proposed method be used to determine the extent to which the causally relevant attention heads actually are separable or composable?

A. If Technique X could identify all heads causally relevant to a skill, we could use them as a reference. We could then measure the proportion of separable or composable heads by the ratio of (# heads identified by both our method and Technique X) / (# heads that Technique X identified).

Although existing methods do not identify all heads causally relevant to a skill, at the request of the reviewer, we conduct this analysis as is. The results on Qwen2.5-VL are shown in the tables below. As shown in Table 1, the heads identified by V-SEAM and NOTICE have substantial overlap with those identified by our method. This further supports our argument in A1 that these baselines actually select heads based on a vague notion of separability. However, as shown in Table 2, the overlapping heads account for only a small proportion of the heads identified by our method. This suggests that by explicitly considering separability and composability, our method can identify a more comprehensive set of heads exhibiting these properties.

Table 1. (# heads identified by both our method and Technique X) / (# heads that Technique X identified)

MethodCounting HeadsPosition HeadsSize HeadsDistance HeadsOverall
VFL1/28 (3.57%)5/28 (17.86%)0/28 (0%)1/28 (3.57%)7/112 (6.25%)
V-SEAM6/20 (30%)6/20 (30%)2/20 (10%)6/20 (30%)20/80 (25%)
NOTICE2/2 (100%)4/4 (100%)6/6 (100%)9/9 (100%)21/21 (100%)

Table 2. (# heads identified by both our method and Technique X) / (# heads that our method identified)

MethodCounting HeadsPosition HeadsSize HeadsDistance HeadsOverall
VFL1/33 (3.03%)5/22 (22.73%)0/21 (0%)1/28 (3.57%)7/104 (6.73%)
V-SEAM6/33 (18.18%)6/22 (27.27%)2/21 (9.52%)6/28 (21.43%)20/104 (19.23%)
NOTICE2/33 (6.06%)4/22 (18.18%)6/21 (28.57%)9/28 (32.14%)21/104 (20.19%)

Q2. The dependence on pre-defined skills is a major weakness of the method. Why should the functionality of attention heads be factorized at the level of the specific proposed skills?

A2. We do not claim that the four skills we used are the best decomposition / factorization. Our algorithm can be applied to skill composition at any level, but it requires a dataset with problems labeled with skills and skills compositions. We chose these four skills because an existing dataset, SPHERE (Zhang et al., 2025), provides problems that human annotators determine can be solved by a single skill and problems that require joint applications of two skills.

The reviewer suggests the four skills we used may be composed of more basic skills. That hypothesis is entirely possible. But we are not aware of any evidence that supports it. In contrast, the authors of SPHERE proposed these four skills are basic skills in spatial understanding and showed that their composition can solve a wide range of problems. That is evidence, in terms of human expert judgments and compositional power, that supports the hypothesis that the four skills may be fundamental.

Again, we recognize that research on skills and skill composition in AI is at an early stage, and our understanding of what is the best level of factorization is limited. However, when there are two competing hypotheses, the rational thing to do is to favor the hypothesis with slightly more evidence.

Thank you for your valuable and responsible review, which has been very inspiring for us.

We would appreciate it if you could let us know whether our response has addressed your concerns. We look forward to further discussion with you.

Thank you to the authors for these replies. I still have the following concerns:

  • Determining whether skills actually are separable and composable at the level of attention heads: To address this question, the authors have provided statistics on the number of overlapping identified heads for their method and other methods, but this is not what I was proposing, and I don't see how this addresses the question. It should be straightforward to ablate the heads identified by their method and to quantify how much this impairs task performance.
  • Whether this decomposition of skills is the right one: The authors argue that there is no evidence against this particular decomposition being the right one, but this reverses the burden of proof -- the burden is on the authors to establish that this decomposition is what the models actually use. They also point to evidence that 1) humans find this decomposition sensible and 2) in principle this decomposition can solve this task. Neither of these points are relevant to the question of whether this is the actual decomposition that models employ. There is also reason to think it is not very likely to be the right decomposition, specifically that counting is a difficult task that requires chain-of-thought to perform precisely, so it is not likely to be something an attention head implements in a single forward pass. In general, it probably doesn't make sense to identify high-level, human-interpretable 'skills' and assume that individual attention heads should carry these skills out. For these reasons, I will stick with my current score.

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

Related papers

Discussion (0)

Sign in to comment.