acceptodds
Under review as a conference paper at ICLR 2027

VFL: Visual Function Localization for Autoregressive Image Generation

Abstract

In autoregressive image generation, visual functions refer to hypothesized internal processes that support behaviors such as introducing attributes and refining local details. These processes remain poorly understood. Circuit analysis explains how internal computations contribute to behavior in language models, but extending it to image generation presents several challenges. Visual concepts typically span multiple image tokens, attention heads may interact, and intervention effects must be evaluated on the final image. We introduce VFL (Visual Function Localization), a framework for studying these functions through activation interventions, i.e., controlled changes to internal activations. VFL first identifies image-token positions where interventions most clearly produce the behavior associated with each visual function. It then learns how strongly to intervene at each attention head at those positions. We evaluate VFL on four visual functions using single-object images with simple backgrounds. Across all four functions, the interventions produce visual changes consistent with the corresponding functional hypotheses. The learned masks assign weight to multiple attention heads rather than concentrating on a single head.

Primary area:
interpretability and explainable AI
Keywords:
Intrepretability

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_h1KvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors use mechanistic interpretability techniques to study autoregressive image generation models. Their goal is to localize visual functions, internal processes like transmitting information from the prompt to the image, or binding color to within the boundaries of an object. They do this in a 2-step process for each target visual function. (1) they localize to a set of generated tokens, by manually coming up with a shortlist of candidate token sets (like the boundary of the object), and patching over the attention head outputs for every head of every layer in that set of tokens and seeing if the generated output image’s color changes. (2) they then localize to a sparse-ish set of attention heads by learning a soft mask over the attention heads when they do patching, supervised by a next token loss rather than by the final entire output.  They use only single-object, plain background images, and use two autoregressive image generation models, Janus Pro and Emu3. The localized interventions seem reasonably successful at transferring color and shadow properties of the objects, though it’s not clear if they exactly represent the stated visual functions.

Critical strengths
  1. The paper is, to the best of my knowledge, the first paper to apply mechanistic interpretability for localizing internal computations to autoregressive image generation models. It is debatable whether or not what they found is a circuit, but it is certainly a step in the right direction of interpreting AR image generation models. Work on doing interpretability for models other than language models is generally healthy for the interpretability ecosystem.
  2. The paper provides many qualitative examples of the intervention outputs, making it easy to judge whether or not their interventions worked.
Critical weaknesses
  1. The intervention experiments do not offer much evidence that the exact visual function has been captured. Fun. 1 to Fun. 3 all have the same final causal effect on the image of changing the object’s color, making it unclear that the localized heads are responsible for the claimed internal computation.

  2. The authors do not discuss the prompts/data that they use anywhere in the main paper. We don’t know what the prompts looked like or how many of them there were. This is pretty critical for interpreting the quantitative results, like in Figure 5, especially given the lack of error bars, to know whether or not their methods worked. The main qualitative results seem to only cover color and shadow of single objects; preferably the results include a more diverse set of prompts.

More minor concerns:

  • nearly all the figures are way too small, not readable without zooming in
  • nit: lines 222-224 seem duplicated
  • nit: line 266 incomplete sentence
Questions for authors

[Clarification] What components exactly of the method did the authors try to ablate in Figure 9b? The ablations seem to show that not every part of the method is necessary?

[Clarification] What are the differences between Fun.1, Fun.2, and Fun.3, and is there evidence showing that the localized heads capture these differences?

[Suggestion] The attention head masks still seem fairly dense. Do the interventions still work with a stronger sparsity regularization penalty on the masks?

9 comments

Weakness 1 + Question 2: Fun. 1–3 lead to similar color changes. What distinguishes their internal computations, and how do the localized heads capture those differences?

Response:

Thank you for asking us to distinguish functional localization from the appearance of an edited image. Color is a common experimental probe, while the three hypotheses concern different computational roles.

  • Different functions and observations (Sec. 3; Fig. 1). Fun. 1, Attribute transmission, concerns prompt-specified information entering the image-token sequence; Fig. 1(a) observes stronger prompt-dependent next-token differences early in generation. Fun. 2, Object-attribute ownership, concerns keeping an attribute associated with its object; Fig. 1(b) measures different hidden-state responses in object regions and far background. Fun. 3, Object-internal propagation, concerns the transfer of established information to later tokens of the same object; Figs. 1(c,d) examine within-object historical attention and local enrichment. These are three distinct internal observables rather than three names for color editing.

  • Different causal tests (Sec. 4.3; Fig. 3). FCP employs distinct intervention strategies: early-prefix positions for Fun. 1, object-neighborhood positions with equal-size spatial controls for Fun. 2, and future-object positions compared against onset/reference positions for Fun. 3. These tests distinguish early attribute transmission, spatial attribute ownership, and subsequent object-internal propagation, respectively. Their function-specific evaluation criteria (Sec. 4.3; Fig. 5(a)) further connect each intervention pattern to its hypothesized computational role.

  • From positions to heads (Secs. 4.4–4.5). Building on function-specific token positions identified by FCP, DML learns layer–head intervention configurations for each visual function. These configurations are evaluated through full autoregressive generation using function-specific measurements and causal controls (Figs. 5, 9). Together, the positional and head-level interventions provide complementary evidence that VFL localizes internal configurations associated with distinct computational roles.

In summary, the functions are distinguished by their generation-time signatures, intervention positions, and target/reference measurements. Their common color outcome is the controlled readout of different causal dependencies.

Weakness 2: Please clarify the prompts/data and sample sizes underlying Fig. 5, quantify uncertainty, and show greater prompt diversity.

Response:

Thank you for identifying where the experimental protocol needs to be made explicit. The evaluation uses manually constructed, controlled donor–recipient prompt pairs so that each intervention targets a specified visual change while shared conditions are held fixed.

  • Prompt Construction, Sample Sizes and Diversity. For Fun. 1/3, donor and recipient describe the same object, view, size, background, and lighting, varying only the specified object color. For Fun. 2, object identity and composition remain fixed while object/background colors are swapped (e.g., red mug on green background ↔\leftrightarrow green mug on red background), enabling evaluation of attribute ownership and background leakage. For Fun. 4, object identity, pose, and view remain fixed while lighting and shadow conditions vary. Fig. 5 evaluates 10 distinct donor–recipient prompt pairs for each function, covering 19 object categories and different color, object–background, and shadow variations. Representative prompt templates, with repeated constraints abbreviated, are:
Fun. 1/3: Studio photo of one {object_color} {object},
          flat {background_color} background, fixed view, ...
Fun. 2:   A single {object_color} {object}, centered,
          plain {background_color} background, studio lighting, ...
Fun. 4:   A single {object}, fixed view,
          {lighting_and_shadow_condition}, ...
  • How the cases are evaluated. The generator remains frozen (Sec. 5.1). FCP first selects function-relevant token positions (Sec. 4.3; Fig. 3); DML learns head weights with those positions fixed (Sec. 4.4; Eq. 4); the retained Top-KK configuration is then evaluated on full generations against the corresponding regional and head controls (Sec. 4.5; Fig. 5). Fig. 5(a) measures target/reference regions, and Fig. 5(b) compares learned configurations with head controls. The metrics are defined per function rather than treating any color change as success.

  • Reliability of intervention effects. Fig. 5 evaluates all four functions using function-specific measurements, reference regions, and head controls, examining both the intended visual changes and their spatial selectivity. Fig. 8(c,d) further provides individual evaluation results and 95% confidence intervals under additional evaluation settings. Together, these complementary analyses support the reliability of the observed functional intervention effects.

Added experimental details to Sec. 5.1 of the revised manuscript.

Question 1: What is ablated in Fig. 9(b)? Are all components necessary?

Response:

Fig. 9(b) examines the contribution of individual DML objective terms (Eq. 4). No-control removes control-position supervision; No-margin removes the margin penalty; Source-only uses donor-token targets alone. All variants retain the FCP-selected positions, activation interventions, and DML mask learning (Secs. 4.2–4.4). All variants preserve positive functional effects, with Source-only achieving the highest displayed score. This demonstrates the robustness of the localization framework to simplified training objectives.

Question 3: The attention-head masks appear dense. Do functional interventions persist under stronger sparsity?

Response:

  • Sparsity is already incorporated into VFL. The regularizer RkR_k in Eq. (4) penalizes total mask weight and encourages near-binary head selection (Sec. 4.4). Thus, the learned masks in Fig. 8(a) are already obtained under sparsity constraints. Our objective is to identify functionally meaningful head configurations, balancing sparsity with causal intervention effectiveness rather than maximizing sparsity alone.

  • Distributed computation motivates joint head localization. Transformer heads can interact through compositional circuits[1], while [2] identified a behavioral circuit comprising multiple interacting head groups. This motivates VFL's focus on joint head configurations rather than isolated heads (Sec. 1). Consistently, Fig. 8(e) shows non-uniform contributions from different head groups, supporting the importance of their collective functional effects. The resulting mask density therefore reflects a regularized, function-oriented localization objective.

References

[1] Elhage, N., et al. A Mathematical Framework for Transformer Circuits. Transformer Circuits, 2021, §“Two-Layer Attention-Only Transformers” → §“Three Kinds of Composition”.

[2] Wang, K. R., Variengien, A., Conmy, A., Shlegeris, B., and Steinhardt, J. Interpretability in the Wild: A Circuit for Indirect Object Identification in GPT-2 Small. ICLR 2023, §3.1–§3.4, particularly §3.2 and §3.4.

Minor concerns: Small figures; duplicated and incomplete line.

Response:

Thank you. We have enlarged the relevant figure labels, removed the duplicated text, and completed the sentence specifying Fun. 2's operational token set. These revisions are reflected in Figs. 1, 2, 3, 5, 6, 8 and Lines 222–224, 266.

Reviewer_6dq2OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes Visual Function Localization (VFL), a mechanistic interpretability framework for understanding how autoregressive image generation models perform visual functions. The paper identifies four visual functions: attribute transmission, object-attribute ownership, object-internal propagation, and local rendering refinement. VFL combines Function-Guided Causal Probing (FCP) to identify relevant image token positions with Differentiable Mask Learning (DML) to localize function-related attention heads. By transferring internal activations from donor to recipient generations, the framework evaluates the effects of localized interventions on final generated images. Experiments on Janus-Pro and Emu3 demonstrate that the proposed approach can induce function-specific visual changes while largely preserving unrelated image content, suggesting that visual functions are supported by distributed attention heads.

Critical strengths
  • This paper provides valuable insights into the generation mechanisms of autoregressive image models, revealing how tokens at different positions contribute to information transmission and visual generation.
  • The proposed analysis framework is simple and efficient, as it requires no modification to the pretrained model parameters.
  • The paper is well-organized and easy to follow, with clear explanations of the methodology and experimental findings.
Critical weaknesses
  • The four visual functions are largely defined based on human prior knowledge and assumptions, which may limit the framework's ability to discover other important visual functions.
  • The experiments are mainly conducted on simple objects with clean backgrounds. While this setting facilitates analysis and interpretation, it may introduce a significant gap between the experimental setup and more realistic, complex image generation scenarios.
  • It would be valuable to further investigate how the insights gained from this analysis could be leveraged to improve image generation performance. Such experiments would provide a better understanding of the practical significance of the findings.
  • Although the framework is evaluated on Janus-Pro and Emu3, it remains unclear whether the identified visual functions and their corresponding token positions and attention heads exhibit consistent patterns across different autoregressive image generation models. A more systematic cross-model analysis would strengthen the generalizability of the conclusions.
Questions for authors

I believe the findings of this paper provide valuable insights into the generation mechanisms of autoregressive image models. I would appreciate the authors' responses to the concerns raised in the weaknesses section. Depending on the responses and additional evidence provided, I would be willing to raise my score.

5 comments

Weakness 1: The four visual functions rely on human prior knowledge, potentially limiting the discovery of other important functions.

Response:

Hypothesis-driven functional analysis is an established paradigm in mechanistic interpretability, rather than a limitation specific to VFL. Prior work typically begins with a well-defined behavior and then identifies the internal mechanisms responsible for it. For example, [1] formalize circuit discovery around a predefined task, dataset, and behavioral metric, while [2] investigate predefined visual attributes through controlled causal interventions.

Importantly, the four visual functions in VFL are representative analysis targets, not an exhaustive taxonomy of visual computation. The framework separates what functional behavior is investigated from how its supporting internal mechanisms are localized. Consequently, additional visual functions can be incorporated by defining corresponding behavioral readouts and candidate token positions, while retaining the same FCP–DML localization framework.

Thus, VFL contributes a generalizable methodology for investigating visual functions, rather than restricting visual computation to four predefined categories.

References

[1] Conmy, A., et al. Towards Automated Circuit Discovery for Mechanistic Interpretability. NeurIPS 2023, §2.1.

[2] Basu, S., et al. Localizing and Editing Knowledge in Text-to-Image Generative Models. ICLR 2024, §3.2.

Weakness 2: The experiments mainly use simple objects with clean backgrounds, potentially leaving a gap to realistic and complex image-generation scenarios.

Response:

Thank you for this valuable suggestion. We extend VFL to naturalistic scenes and provide quantitative evidence of function-specific interventions beyond the original controlled setting.

  • Controlled scenarios and prior work. Following established visual mechanistic interpretability studies (Basu et al., ICLR/ICML 2024), which use attribute-specific interventions and controlled prompt pairs to isolate causal effects, VFL designs function-specific scenarios, intervention positions, and evaluation metrics (Secs. 3–4). Such controlled settings help distinguish intended functional effects from unrelated visual changes.

  • Scope of Fun. 1 and Fun. 4. Their original evaluations are deliberately designed around controlled single-object scenes. Fun. 1 (Attribute transmission) probes early-prefix interventions, whose effects can propagate throughout complex scenes, making attribute-specific changes difficult to isolate. Fun. 4 (Local rendering refinement) isolates ground-shadow changes, which are strongly coupled with illumination, object geometry, and surrounding surfaces in complex scenes, making it difficult to preserve the causal contrast required for shadow-specific evaluation. We therefore retain these controlled evaluations and extend Fun. 2/3, whose object-specific regions allow clearer causal comparisons.

  • Naturalistic-scene experiments. We evaluate Fun. 2 (Object-attribute ownership) and Fun. 3 (Object-internal propagation) using previously learned head masks and scene-adapted intervention positions, compared against layer-matched random-head controls.

    Following the regional color-progress measurement in Eq. (2), we quantify donor-directed changes within each fixed region RR:

    \left(\bar c_R(y)-\bar c_R(y^{\mathrm{rec}})\right)^\top \left(\bar c_R(y^{\mathrm{don}})-\bar c_R(y^{\mathrm{rec}})\right) }{ \left\|\bar c_R(y^{\mathrm{don}})-\bar c_R(y^{\mathrm{rec}})\right\|_2^2+\epsilon },$$ where $\bar c_R$ is the mean RGB color in region $R$. **Fun. 2 — Object-attribute ownership.** We construct a two-object scene with a **red apple and blue mug** (recipient) and a **green apple and red mug** (donor). The intended effect is to transfer the apple's green attribute while preserving the mug's blue color. We measure **Target progress** $T=P_{\mathrm{apple}}$, **Distractor leakage** $L=P_{\mathrm{mug}}$, and **Binding score** $B=T-\max(0,L)$. | Intervention | Target ↑ | Leakage ↓ | Binding ↑ | |:--|--:|--:|--:| | **Learned** | **0.2533** | **0.0349** | **0.2184** | | Random | 0.0279 | 0.1565 | -0.1286 | **Learned achieves greater target-color transfer with substantially lower distractor leakage**, supporting selective object-attribute ownership in a multi-object scene. **Fun. 3 — Object-internal propagation.** We examine an indoor scene containing a blue ball with a red-ball donor. The learned intervention introduces a visible donor-directed red region within the ball. We measure color progress in the **Future** and **Onset** regions, corresponding to later- and earlier-generated regions of the same object, respectively. The propagation specificity is: $$S_3=P_{\mathrm{Future}}(y)-\max(0,P_{\mathrm{Onset}}(y)).$$ | Intervention | Future ↑ | Onset | Specificity ↑ | |:--|--:|--:|--:| | **Learned** | **0.7014** | 0.2626 | **0.4388** | | Random | 0.3547 | 0.2408 | 0.1139 | **Learned produces nearly twice the Future-region color progress (0.7014 vs. 0.3547) with comparable Onset changes**, yielding substantially higher specificity (**0.4388 vs. 0.1139**). This supports preferential attribute propagation toward later-generated object regions. *Added to Sec. 5.5 of the revised manuscript.*

Weakness 3: How can the discovered mechanisms improve image generation, and what is their practical significance?

Response:

Prior research has established a direct connection between mechanistic understanding and practical improvements in image generation. [1] demonstrate that identifying causal components enables DiffQuickFix, which efficiently edits visual concepts such as objects, styles, and factual associations. Extending this direction, Basu et al. introduce LocoEdit[2], showing that mechanistically localized layers support targeted editing across multiple diffusion architectures, including SD-XL. These findings demonstrate that localizing the internal mechanisms underlying visual attributes can translate into more precise and efficient control over generation behavior.

More recent work further demonstrates that mechanistic insights can directly guide inference-time generation. Tinaz et al. identify interpretable concepts in diffusion models and use targeted activation interventions to manipulate image composition and style[3]. Their results reveal that effective control depends on when and where interventions are applied during generation. This provides further evidence that understanding the functional organization of a generative model can enable more selective and interpretable control, beyond conventional parameter-level optimization.

VFL extends this established connection between mechanistic localization and generation control to autoregressive image generators. Unlike diffusion models, autoregressive generators construct images through sequential token predictions, making both the intervention positions and their supporting attention-head configurations essential to functional control. VFL addresses these aspects jointly through FCP and DML. Importantly, our existing experiments already demonstrate this practical connection: Figs. 4–5 show function-directed visual changes with measurements of non-target effects, while Fig. 7 demonstrates the composition of independently localized color and shadow functions within a single generation, further evaluated against alternative intervention conditions in Fig. 9(c).

Thus, VFL provides an interpretable, parameter-free inference-time mechanism for selective and compositional visual control, establishing a foundation for mechanism-guided improvements in autoregressive image generation.

Weakness 4: It remains unclear whether the identified visual functions and their corresponding token positions and attention heads exhibit consistent patterns across different autoregressive image generation models.

Response:

Cross-model mechanistic analysis is valuable for identifying common functional patterns across different generators. For example, Basu et al. [4] investigate the localization of similar visual attributes across multiple text-to-image models, revealing both common functional targets and differences in their supporting internal components. This motivates examining the commonalities in visual functions and their supporting computations across different autoregressive image generators.

Our existing cross-model evaluation provides evidence of such functional commonalities (Sec. 5.3). We apply the same FCP–DML framework to frozen Janus-Pro and Emu3, examining the four visual functions defined in Sec. 3. Figs. 4–5 show function-specific interventions on Janus-Pro, while Fig. 6 presents the learned head masks and representative interventions on Emu3. Importantly, both generators exhibit visual changes consistent with the same four functional hypotheses, indicating commonalities in the visual effects produced by the corresponding interventions.

These findings highlight two complementary aspects of VFL: the four visual functions provide common targets for investigating autoregressive image generation, while FCP and DML enable the corresponding function-relevant image-token positions and attention-head weights to be identified within each model. Together, the Janus-Pro and Emu3 results support both the shared functional observations and the cross-model applicability of VFL's localization framework.

References

[1] Basu, S., et al. Localizing and Editing Knowledge in Text-to-Image Generative Models. ICLR 2024, §5.3.

[2] Basu, S., et al. On Mechanistic Knowledge Localization in Text-to-Image Generative Models. ICML 2024, §6.2.

[3] Tinaz, B., Fabian, Z., and Soltanolkotabi, M. Emergence and Evolution of Interpretable Concepts in Diffusion Models. NeurIPS 2025, §4.3.

[4] Basu, S., et al. On Mechanistic Knowledge Localization in Text-to-Image Generative Models. ICML 2024, §§5.1–5.2.

Reviewer_ita7OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Visual Function Localization (VFL), a framework for investigating internal visual functions in autoregressive image generation models. It combines Function-Guided Causal Probing (FCP) to identify relevant image-token positions with Differentiable Mask Learning (DML) to learn intervention weights for attention heads. Experiments on Janus-Pro and Emu3 demonstrate visual changes associated with four proposed functions: attribute transmission, object-attribute ownership, object-internal propagation, and local rendering refinement.

Critical strengths
  1. Interesting research direction. The paper explores mechanistic interpretability in autoregressive image generation, an important topic that has received relatively limited attention compared with language and diffusion models.
  2. Intervention-based evaluation. Rather than relying solely on attention visualization, the authors directly manipulate internal activations and examine the resulting images. The inclusion of token-set and attention-head controls provides useful evidence for evaluating the intervention effects.
Critical weaknesses
  1. Insufficient evidence for functional localization.

The proposed intervention demonstrates that modifying certain activations can induce desired visual changes. However, this does not necessarily establish that these components are responsible for the corresponding visual functions during normal generation. The identified heads may simply provide effective pathways for injecting donor information. Additional necessity-based experiments, such as selectively ablating the identified heads during normal generation, would help establish stronger causal evidence for the claimed functional localization.

  1. Limited evidence of head-level specificity.

In Figure 5(b), the learned head masks achieve nearly identical results to the matched head controls for Functions 1 and 3. This raises questions about whether the identified heads are genuinely function-specific or whether other head combinations could produce similar effects. Together with the mask instability observed in Figure 8(b), these results weaken the claim that VFL consistently identifies stable, function-specific internal components.

  1. Limited generalization.

The experiments focus primarily on single-object scenes with simple backgrounds, and the first three functions are mainly evaluated through color changes. More diverse visual attributes, multi-object scenes, and complex backgrounds would help establish whether the proposed framework generalizes beyond these controlled settings. The cross-model evaluation on Emu3 would also benefit from more comprehensive quantitative results.

Questions for authors
  1. How can the authors distinguish the identified functional heads from general pathways that are effective for donor activation injection? Could necessity-based interventions provide stronger evidence?

  2. How do the authors explain the instability of learned masks across training subsets in Figure 8(b)? Does this affect the interpretation and transferability of the localized functions?

  3. Can the authors provide additional quantitative results on intervention quality and attribute consistency, including an analysis of intervention strength and more diverse visual examples?

4 comments

Weakness 1 / Question 1: Can donor-to-recipient interventions reliably localize visual functions, and is necessity-based evidence required?

Response:

  • Published localization methods use controlled activation interventions. Zhang and Nanda[1] describe activation patching, also called causal tracing or interchange intervention, as a standard technique for locating behavior-relevant components. Meng et al.[2] restore internal activations to investigate factual predictions, and Conmy et al.[3] use activation replacement in automated circuit discovery. In image generation, Basu et al.[4] use causal interventions to localize visual knowledge, while Basu et al.[5] study the effect of changing conditioning information at selected layers. These studies provide directly relevant precedents for interpreting controlled intervention effects as evidence of functional localization.

  • VFL follows this practice with function-specific comparisons. FCP (Sec. 4.3; Fig. 3) compares candidate image-token positions and alternative token sets. For Fun. 2 (Object-attribute ownership), Eq. (3) compares the selected positions against equal-size spatial controls; Eq. (2) separately evaluates object and background color changes. DML (Sec. 4.4) then learns the attention-head weights, and Sec. 4.5; Figs. 4–5 evaluate the resulting changes in the final decoded image, including head controls. These steps make the evidence more specific than the observation that an arbitrary donor intervention changes the image.

  • Necessity-based interventions provide complementary evidence. Ablating attention heads during normal generation examines their necessity under an ablation intervention, whereas VFL evaluates function-specific image-level effects through controlled donor-to-recipient activation interventions. Both approaches contribute to mechanistic understanding, and the latter follows the established intervention-based localization methodology adopted in prior work.

We hope this clarifies why the FCP and DML results provide intervention-based evidence for function-relevant token positions and attention-head weights, in line with established circuit and visual-localization studies.

References

[1] Zhang, F., and Nanda, N. Towards Best Practices of Activation Patching in Language Models: Metrics and Methods. ICLR 2024, §1.

[2] Meng, K., et al. Locating and Editing Factual Associations in GPT. NeurIPS 2022, §2.1.

[3] Conmy, A., et al. Towards Automated Circuit Discovery for Mechanistic Interpretability. NeurIPS 2023, §2.3.

[4] Basu, S., et al. Localizing and Editing Knowledge in Text-to-Image Generative Models. ICLR 2024, §3.2.

[5] Basu, S., et al. On Mechanistic Knowledge Localization in Text-to-Image Generative Models. ICML 2024, §5.1.

Weakness 2 / Question 2: How do head-mask specificity and stability affect the interpretation of VFL's functional localization?

Response:

Thank you for the thoughtful questions regarding head-level specificity and mask stability.

  • Non-identifiability of mechanistic explanations. Prior research suggests that the same model behavior need not admit a unique mechanistic explanation. Méloux et al.[1] systematically investigate this question and demonstrate that multiple distinct circuits can explain the same behavior, even when evaluated using mechanistic interpretability criteria. Elhage et al.[2] also describe interactions between Transformer attention heads, illustrating the complexity of attributing computations to individual components. These findings offer an important perspective: identifying a functionally relevant set of components does not necessarily imply that it is the only configuration capable of supporting that function. Consequently, differences between localized configurations need not, by themselves, invalidate their functional interpretation.

  • Interpreting the VFL results. This perspective helps explain why head-level specificity and mask stability should be examined in terms of the functional behavior being localized. Fig. 5(b) compares the learned head masks with alternative head configurations, while Fig. 8(b) examines differences between masks learned from training subsets A/B/C. These observations are relevant to the possibility that a visual function may admit more than one effective head configuration, rather than requiring a single uniquely identifiable set. Meanwhile, Figs. 8(c,d) evaluate the functional effects of the learned configurations across evaluation cases and intervention strengths. Taken together, these analyses support interpreting VFL's localization results through the functional effects of the identified configurations, while recognizing that the underlying head selections may not be unique.

References

[1] Méloux, M., et al. Everything, Everywhere, All at Once: Is Mechanistic Interpretability Identifiable? ICLR 2025, §§3 and 5.1.

[2] Elhage, N., et al. A Mathematical Framework for Transformer Circuits. Transformer Circuits, 2021, §“Three Kinds of Composition”.

Weakness 3 / Question 3:
Generalization beyond simple scenes, intervention quality, attribute consistency, and intervention-strength analysis.

Response:

Thank you for suggesting broader visual scenarios and quantitative evidence. We would like to clarify the rationale for our controlled experimental design and provide additional results on naturalistic scenes.

  • Controlled scenarios and prior work. Following established visual mechanistic interpretability studies [1, 2], which use attribute-specific interventions and controlled prompt pairs to isolate causal effects, VFL designs function-specific scenarios, intervention positions, and evaluation metrics (Secs. 3–4). Such controlled settings help distinguish the intended functional effect from unrelated visual changes [3].

  • Scope of Fun. 1 and Fun. 4. Their original evaluations are deliberately designed around controlled single-object scenes. Fun. 1 (Attribute transmission) probes early-prefix interventions, whose effects can propagate throughout complex scenes, making attribute-specific changes difficult to isolate. Fun. 4 (Local rendering refinement) isolates ground-shadow changes, which are strongly coupled with illumination, object geometry, and surrounding surfaces in complex scenes, making it difficult to preserve the causal contrast required for shadow-specific evaluation. We therefore retain their controlled evaluations and extend Fun. 2/3, whose object-specific regions allow clearer causal comparisons in complex compositions.

  • Naturalistic-scene validation of Fun. 2 and Fun. 3. We extend Fun. 2 (Object-attribute ownership) and Fun. 3 (Object-internal propagation) to naturalistic scenes using previously learned head masks and scene-adapted intervention positions, compared against layer-matched random-head controls.

    Following the regional color-progress measurement in Eq. (2), we quantify donor-directed changes in each region RR as:

    \left(\bar c_R(y)-\bar c_R(y^{\mathrm{rec}})\right)^\top \left(\bar c_R(y^{\mathrm{don}})-\bar c_R(y^{\mathrm{rec}})\right) }{ \left\|\bar c_R(y^{\mathrm{don}})-\bar c_R(y^{\mathrm{rec}})\right\|_2^2+\epsilon },$$ where $\bar c_R$ denotes the mean RGB color in the fixed region $R$. **Fun. 2 — Object-attribute ownership.** We construct a two-object scene with a **red apple and blue mug** (recipient) and a **green apple and red mug** (donor). The intended effect is to transfer the apple's green attribute while preserving the mug's blue color. We measure **Target progress** $T=P_{\mathrm{apple}}$, **Distractor leakage** $L=P_{\mathrm{mug}}$, and **Binding score** $B=T-\max(0,L)$. | Intervention | Target ↑ | Leakage ↓ | Binding ↑ | |:--|--:|--:|--:| | **Learned** | **0.2533** | **0.0349** | **0.2184** | | Random | 0.0279 | 0.1565 | -0.1286 | **Learned achieves greater target-color transfer with substantially lower distractor leakage**, supporting selective object-attribute ownership in a multi-object scene. **Fun. 3 — Object-internal propagation.** We examine an indoor scene containing a blue ball with a red-ball donor. The learned intervention introduces a visible donor-directed red region within the ball. We measure color progress in the **Future** and **Onset** regions, corresponding to later- and earlier-generated regions of the same object, respectively. The propagation specificity is: $$S_3=P_{\mathrm{Future}}(y)-\max(0,P_{\mathrm{Onset}}(y)).$$ | Intervention | Future ↑ | Onset | Specificity ↑ | |:--|--:|--:|--:| | **Learned** | **0.7014** | 0.2626 | **0.4388** | | Random | 0.3547 | 0.2408 | 0.1139 | **Learned produces nearly twice the Future-region color progress (0.7014 vs. 0.3547) with comparable Onset changes**, yielding substantially higher specificity (**0.4388 vs. 0.1139**). This supports preferential attribute propagation toward later-generated regions within the object.
  • Existing quantitative evidence and intervention-strength analysis. These naturalistic examples complement the controlled evaluations already reported in the submission. Fig. 5(a) distinguishes target and reference regions for the four functions, while Fig. 8(d) examines the response of Fun. 3 to varying intervention strengths. Fig. 6 presents the supplementary evaluation on frozen Emu3, where representative interventions exhibit visual changes consistent with all four functional hypotheses.

Added to Sec. 5.5 of the revised manuscript.

References

[1] Basu, S., et al. Localizing and Editing Knowledge in Text-to-Image Generative Models. ICLR 2024, §3.2–4.

[2] Basu, S., et al. On Mechanistic Knowledge Localization in Text-to-Image Generative Models. ICML 2024, §5.1.

[3] Zhang, F., and Nanda, N. Towards Best Practices of Activation Patching in Language Models: Metrics and Methods. ICLR 2024.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer,

Thank you again for your thoughtful review. We have carefully addressed your concerns in our rebuttal and provided additional experimental evidence. We would greatly appreciate any further feedback you may have on our responses.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.