acceptodds
Under review as a conference paper at ICLR 2027

SPARSE TRANSCODERS FOR MECHANISTIC ANALYSIS OF VLA POLICIES

Abstract

We adapt sparse time-conditioned transcoders (STCs) to the action expert of the flow-matching vision-language-action policy π0.5, giving a sparse intervention in- terface for systematically studying continuous action prediction. We train one STC for each of the 18 action-expert MLP blocks; the resulting sparse codes activate only 5.02% of the dictionary on average and support feature discovery, circuit tracing, open-loop ablation, and steering. In a discovery-selected black- bowl / bottom-drawer case study, a rare feature firing on only 0.166% of probed observations yields a traced 16-node upstream circuit whose ablation produces about a 2× larger action-chunk change than ablating the target alone, and about a 2× larger effect than matched same-size random circuits in the evaluated cir- cuit setting. Target ablation is zero on negative controls where the feature is in- active, while behavior-grounded steering increases the Cartesian-speed score on 100% of evaluated examples for target-feature steering and 85% for circuit steer- ing; an agreement-filtered contrastive direction changes the gripper coordinate by about 1.6 at α = 2, compared with about 0.05 for a matched random direc- tion. Separately, we run a visual-counterfactual audit using non-time-conditioned TopK transcoders on the PaliGemma language-prefix MLPs. There, 83.72% of a capped set of visually selective features reproduce their selectivity on held-out frames, but small sparse patches have limited and direction-dependent action ef- fects, and attribution-based action restoration does not improve closed-loop suc- cess over occlusion-only or random-restoration controls. Together, the two stud- ies show that sparse transcoders can support feature and circuit interventions in VLA policies while separating feature selectivity, open-loop action effects, steer- ing, closed-loop behavior, and evidence that a circuit carries its attributed visual content

Primary area:
interpretability and explainable AI
Keywords:
Mechanistic interpretability, sparse transcoders, circuit tracing, flow matching, vision-language-action policies, robotic manipulation, feature attribution, LIBERO

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_vafSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The author proposed a sparse representation-based analysis forπ0.5 VLA policy by adapting timestep-conditioned sparse transcoders (STCs). The learned representations are used for feature discovery, attribution-based circuit tracing, open-loop ablation and steering. In a selected case study, ablating a traced 16-node circuit changes predicted actions more than ablating the target feature alone or matched random circuits. Separately, the author used non-time-conditioned TopK transcoders to examine visual features in the policy’s PaliGemma language-prefix MLPs. Although many selected features retain visual selectivity on held-out frames, sparse interventions have limited action effects, and attribution-based restoration does not improve closed-loop success over occlusion-only or random-restoration controls.

Critical strengths
  1. The author presented a valuable explanation adaptation to VLA. Applying STCs across action-expert MLPs is meaningful engineering and empirical extension of sparse circuit analysis to a flow-matching policy. The training setup, loss functions, intervention protocols and results are informative.
  2. The author did not equate visually selective features with open-loop action effects from behavioral recovery, which is commendable in the field of explanability. In particular, they report that attribution-based closed-loop restoration fails to outperform the occlusion-only and random-restoration controls, while full-MLP restoration succeeds. This negative result makes the paper more informative and credible than a report limited to successful examples.
Critical weaknesses
  1. Limited technical novelty. Sparse-transcoder, timestep-conditioning, circuit-tracing, and counterfactual intervention are existing techniques. The paper adapts these techniques to a flow-matching VLA policy. The methodological advance appears incremental. Meanwhile the empirical results do not yet clearly compensate for the incremental methodological advance.
  2. Limited validation. The main action-expert circuit result centers on a single discovery-selected feature and its traced 16-node circuit. Although the ablations outperform matched random controls on evaluated open-loop examples, the paper does not show whether applying the same feature-selection, tracing, and intervention procedure to other candidate features would yield similarly coherent and intervention-sensitive circuits. Nor is it clear how consistently such results would hold across tasks and independently held-out episodes. Consequently, the evidence supports a promising case study more strongly than a generalizable account of action-expert computation.
  3. Limited experimental support for behavioral claims. The action-expert ablation and steering experiments show changes in predicted actions, but do not establish whether the identified features and circuits causally affect closed-loop task performance. In the separate PaliGemma counterfactual experiments, attribution-based restoration does not improve closed-loop success over occlusion-only or random-restoration controls, despite the strong full-MLP positive control. Together, these results leave the behavioral significance of the sparse explanations uncertain. The visual-audit result should not, however, be interpreted as a direct failure of the action-expert circuit.
Questions for authors
  1. The paper presents the F9970 circuit as a case study but also describes the tracing procedure as a general workflow. Beyond this selected feature, what evidence suggests that the same feature-selection, tracing, and intervention procedure would identify similarly intervention-sensitive circuits for other action-expert features?
  2. The introduction lists four contributions and acknowledges prior work on sparse-transcoder circuit tracing and timestep conditioning. To be more specific, what is the technical advance in adapting these methods to the flow-matching action expert, beyond applying the existing pipeline to a new model? Which results should instead be viewed as empirical findings of this case study and the separate visual audit?
  3. Please report the computational cost of training the STCs, including hardware, GPU-hours, peak memory, and other necessary costs. It would also be useful to report the runtime of feature discovery and circuit tracing, distinguishing one-time costs from costs that scale with the number of analyzed features or episodes.
  4. Circuit tracing uses STCs in replacement mode, but the reported MLP reconstruction losses do not establish end-to-end fidelity. Please report action-chunk differences from the original policy across tasks and flow timesteps. Closed-loop success under replacement, compared with the original policy, would also be informative.
  5. Presentation. Figure 1 and several other figures are difficult to read at their current resolution. Please improve their legibility and check if they are vector graphs. Please also check the notation throughout, such as “π” at the bottom of Figure 1 is wrongly written as “Pi”
  6. The appendix provides several diagnostics of the failed closed-loop restoration. Do they identify the main reason attribution-based sparse restoration fails to improve success, or do feature selection and intervention coverage remain unresolved explanations?
Reviewer_sjYoOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper explores the use of sparse transcoders for interpretability of vision-language-action (VLA) models through two complementary studies. First, the authors use sparse time-conditioned transcoders (STCs) to the π0.5\pi_{0.5} action expert for feature discovery, circuit tracing, ablation, and steering, demonstrating that the identified sparse features and circuits can influence predicted actions along intended behavioural axes. Second, the authors use non-time-conditioned TopK transcoders on the PaliGemma language-prefix MLPs to conduct a visual-counterfactual audit. They find that many visually selective features retain their selectivity on held-out frames, while attribution-based restoration alters open-loop action predictions but does not reliably improve closed-loop task success. Overall, the paper demonstrates the potential of sparse transcoders for interpreting and manipulating VLA.

Critical strengths

The paper includes reasonably designed control experiments, such as random single-feature controls, which help assess the reliability of the reported findings.

Critical weaknesses
  1. The writing should be improved.

1), the paper lacks sufficient architectural context. Section 2.1 introduces STCs without providing an overview of the relevant π0.5\pi_{0.5} architecture, particularly where the target MLPs are located and how the STCs interact with them. This makes the proposed method difficult to understand.

2), several symbols and labels are not clearly defined. For example, on page 3, line 152, in xℓ,r∈RB×T×dmodelx_{\ell,r}\in\mathbb{R}^{B\times T\times d_{\mathrm{model}}}, while ℓ\ell presumably denotes the layer index, the meaning of rr remains unclear. Similarly, Table 1 uses labels such as L12:τ=1.0:F7584L12:\tau=1.0:F7584 without explaining the naming convention. In Figure 1, labels such as preactipreact_i and fencif_{enc} i are neither defined nor clearly connected to the equations in Section 2.1.

3), several experimental setups are insufficiently explained. For instance, Table 2 does not clarify the difference between “Target ablation, positives” (row 1) and “Target ablation, matched run” (row 6).

  1. The reported effects in Table 2 and Table 3 are relatively small in magnitude, particularly the steering results in Table 3 (on the order of 10−310^{-3} to 10−410^{-4}). It is questionable whether such small differences are sufficient to support the authors' conclusions.
Questions for authors
  1. Given that the experiments are primarily based on selected case studies, how can the effectiveness and generalizability of the proposed method be validated on a larger scale?

  2. Do the authors have any insights into why these action-level effects do not translate into meaningful behavioral improvements? What are the potential limitations of the proposed approach in explaining closed-loop behavior?

Reviewer_2RUrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates mechanistic interpretability in VLA policies using sparse transcoders. The authors adapt sparse time-conditioned transcoders to the action expert of π0.5\pi_{0.5} and analyze internal features through circuit tracing, ablation, and steering.

Critical strengths
  1. The paper extends time-conditioned sparse transcoders to the action expert of a flow-matching VLA policy, providing a useful framework for feature- and circuit-level analysis of continuous action predictions.
  2. The STCs achieve similar training, validation, and test reconstruction losses across 18 action-expert MLP blocks while maintaining an average feature activation rate of 5.02%.
Critical weaknesses
  1. Section 2.1 evaluates STC fidelity mainly through MLP reconstruction loss, while circuit tracing and ablation in Sections 2.3 and 2.4 rely on replacement mode. An additional comparison of action predictions before and after STC replacement, without intervention, would help assess how reconstruction errors affect the subsequent causal analysis.
  2. In Section 2.4, F9970 is already inactive on the negative controls, making the zero ablation effect expected. This provides limited evidence of semantic specificity. Controls with similar activation levels but different semantic contexts would offer stronger evidence.
  3. Section 2.5 defines the speed score using the mean L2L_2 norm of translational action components. This metric measures movement magnitude without accounting for direction. Evaluating velocity projected onto the drawer-closing direction would better establish whether steering produces the intended motion.
  4. The use of Section 1.1 is unusual, and the authors should carefully check the paper's formatting. In addition, the AI-generated Fig.1 is difficult to read and would benefit from a clearer presentation.
Questions for authors

Please see Weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.