- Rating
- 3
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes SteerAct, a method for improving frozen robot action models through task-specific interventions on a single sparse autoencoder (SAE) feature. Using successful and failed reference rollouts, it selects the feature whose temporal activation patterns best separate the two outcomes. During execution, support and failure gates determine whether to intervene, and a bounded search chooses a signed intervention strength predicted to move the current activation descriptor closer to successful references. The policy and SAE remain frozen during adaptation and deployment.
Critical strengths
- The method offers a lightweight way to adapt frozen robot policies using only 24 outcome-labeled reference rollouts per task, without training an auxiliary controller.
- The evaluation spans two model backbones, two simulation benchmarks, and three real robot tasks. The preferred encoder variant improves success in 34 of 36 simulated task–backbone settings.
- The gating ablations provide useful evidence that selective intervention matters: the full method outperforms fixed-delay variants, while removing both gates reduces performance below the frozen policy on VLABench.
Critical weaknesses
- The central contribution is not sufficiently isolated. The experiments demonstrate overall gains but do not establish how much comes from the proposed feature-selection criterion.
- The intervention mechanism is insufficiently understood. The analysis overlooks feature interactions and the broader effects of editing a single feature.
- The advantage over alternative steering approaches remains unclear. Existing comparisons do not isolate the benefit of intervening in selected SAE directions.
- Several design choices lack sufficient justification. This makes the method’s robustness and ease of application to other models difficult to assess.
- The scope of generalization is limited. Results support task-specific adaptation, while transfer beyond the reference-task settings remains unestablished.
Questions for authors
-
Feature selection. Does steering the highest-scoring feature outperform steering random or lower-scoring features under the same procedure? Is the selection score associated with the resulting improvement in task success?
-
Feature interactions. To what extent does changing the selected feature affect other features? For the multi-feature variants, are these interactions considered when choosing intervention strengths?
-
Alternative intervention spaces. Would direct noise-space search achieve similar gains under comparable data and inference budgets? The DSRL-SAC comparison does not isolate the intervention space because its training procedure also differs.
-
Intervention location. Which exact layers and tokens are used for each backbone, and what determines their selection? Is performance stable across different intervention layers?
-
Token aggregation. Could averaging across tokens hide important differences between concentrated and widely distributed feature activity? Would retaining token-level variation improve feature selection or steering?
-
Generalization. With the selected feature and reference banks fixed, does the method remain effective under changes in objects, layouts, instructions, or disturbances? Which of these variations are covered by the reported evaluation?
-
Reference collection. Can the required successful and failed references be obtained within 24 attempts when the base policy rarely succeeds or rarely fails? Are any additional collection attempts included in the reported budget?
9 comments
R3-W3 — The advantage of selected SAE directions over alternative intervention spaces is unclear.
Response. Thank you for distinguishing complete-method performance from the effect of intervention space. Our experiments compare SteerAct with both noise-space adaptation through DSRL-SAC and representation-level steering through COAST.
Alternative steering methods — π₀.₅. Mean success rates (%, ↑) across nine tasks per benchmark.
| Method | RoboTwin 2.0 | DexJoCo |
|---|---|---|
| DSRL-SAC | 71.1 | 58.2 |
| COAST | 68.4 | 65.3 |
| SteerAct (Enc) | 77.7 | 71.9 |
On DexJoCo, SteerAct exceeds DSRL-SAC and COAST by 13.7 and 6.6 percentage points, respectively. These comparisons demonstrate the effectiveness of the complete temporal-steering method under the reported evaluation settings.
We additionally isolate feature choice within the SAE intervention space: on eight DexJoCo tasks with Fast-WAM, the highest-ranked feature achieves 30.1% mean success, compared with 25.8% for a randomly selected feature under the same controller. This supports selecting an informative SAE direction rather than applying the controller to an arbitrary feature.
However, neither result establishes that SAE-space intervention is intrinsically superior to noise-space intervention. DSRL-SAC also differs in how its controller is learned, and direct noise-space search under matched reference-data and inference budgets has not been evaluated. Such a comparison must match the scoring information and measured computation, rather than candidate count alone. We will therefore scope our claim to the demonstrated effectiveness of temporal steering and feature selection within the SAE interface.
R3-W4 — Several design choices lack sufficient justification.
Response. Thank you for this concern. Our completed ablations examine the temporal representation, intervention branch, gating procedure, and source-rollout budget. The following comparisons clarify which choices are supported by the experiments.
(a) Temporal representation and intervention site — RoboTwin 2.0 with π₀.₅. Mean success rates (%, ↑) across nine tasks.
| Setting | Mean success (%) |
|---|---|
| Base | 69.0 |
| Marginal statistics | 71.1 |
| Multiscale increments | 70.0 |
| Event-based summaries | 70.1 |
| Autocorrelation-based matching | 73.9 |
| Raw-history nearest neighbor | 72.2 |
| Raw-history DTW | 71.7 |
| SteerAct (Dec) | 70.9 |
| SteerAct (Enc; full descriptor) | 77.7 |
All matching controls use Enc and retain the ordinal-based support gate. They therefore test the representation used for outcome matching and correction scoring within a common support-gating procedure. The full descriptor exceeds autocorrelation, raw-history nearest neighbors, and DTW by 3.8, 5.5, and 6.0 percentage points, respectively. The separate Enc–Dec comparison supports Enc as the default intervention branch in this setting.
(b) Timing and gating — DexJoCo with π₀.₅. Mean success rates (%, ↑) across nine tasks, using 200 evaluation episodes per task with matched layout seeds.
| Timing / gating setting | Mean success (%) |
|---|---|
| Delay 10 | 68.3 |
| Delay 20 | 70.6 |
| Without support gate | 69.8 |
| Without failure gate | 67.8 |
| SteerAct | 71.9 |
These comparisons retain the same selected feature, reference banks, temporal descriptor, and full signed coefficient-search grid. Delay 10 and Delay 20 replace the support gate with a fixed waiting period while retaining the failure gate. Each gate-removal control retains the other gate and the minimum history needed to compute the descriptor. All settings retain the same no-improvement fallback.
The full controller exceeds the fixed-delay alternatives by 3.6 and 1.3 points. Removing the support or failure gate reduces success by 2.1 and 4.1 points, respectively, computed before rounding. These results support both reference-dependent timing and outcome-selective intervention.
(c) Source-rollout budget — RoboTwin 2.0 with π₀.₅. Mean success rates (%, ↑) across nine tasks.
| Source rollouts per task | 12 | 24 (default) | 48 | 96 |
|---|---|---|---|---|
| Mean success (%) | 73.6 | 77.7 | 77.3 | 77.3 |
The default 24-rollout setting performs competitively with the larger tested budgets; increasing the budget does not improve observed mean success in this study. This supports the practical usefulness of a small reference collection, without implying that 24 rollouts cover every execution mode.
Together, these ablations provide empirical justification for the tested design. They do not establish insensitivity to arbitrary layers, token selections, or hyperparameters. We will distinguish the demonstrated branch-level comparison from the additional layer- and token-level sensitivity questions and clarify the configurations used for each backbone.
R3-W5 — Generalization beyond reference-task settings remains unestablished.
Response. Thank you for highlighting this distinction. Our experiments test task-specific adaptation to new executions with varying layouts, and the applicability of the same framework across model families and manipulation embodiments.
For each task where steering is enabled, the selected feature and reference banks remain fixed throughout 200 evaluation episodes, while the layout varies between episodes. The controller therefore uses the collected references to guide subsequent executions without rebuilding the banks for each layout.
Results across backbones and embodiments. Mean success rates (%, ↑) across nine tasks per benchmark. Gains are computed before rounding and reported in percentage points.
| Benchmark | Embodiment | Backbone | Base | SteerAct (Enc) | Gain (pp) |
|---|---|---|---|---|---|
| RoboTwin 2.0 | Grippers | π₀.₅ (VLA) | 69.0 | 77.7 | +8.7 |
| RoboTwin 2.0 | Grippers | Fast-WAM (WAM) | 69.3 | 78.6 | +9.2 |
| DexJoCo | Dexterous hands | π₀.₅ (VLA) | 62.3 | 71.9 | +9.6 |
| DexJoCo | Dexterous hands | Fast-WAM (WAM) | 23.3 | 26.8 | +3.4 |
These aggregates include Hanoi Bi, for which Fast-WAM remains unsteered because successful references were unavailable.
SteerAct improves observed success in 34 of 36 task–backbone combinations and improves the macro-average in all four benchmark–backbone settings. This supports the framework’s applicability across VLA and WAM models and across gripper-based and dexterous-hand manipulation. Fixed references across varying evaluation layouts also provide evidence of within-task reuse beyond the collected reference executions.
Cross-task transfer is a separate claim. We have not evaluated transferring one task’s selected feature and banks unchanged to an unseen task. Nor do the reported experiments separately establish robustness to unseen object identities, instruction shifts, or external disturbances. We will make the demonstrated scope explicit: task-specific few-shot adaptation across varying layouts, models, and embodiments, rather than unrestricted transfer of a fixed steering feature and reference bank.
R3-W1 — The contribution of feature selection is not sufficiently isolated.
Response. Thank you for highlighting this concern. We added controlled comparisons that change the selected SAE feature while keeping the steering procedure unchanged. This directly tests whether our temporal ranking identifies features that are more useful for improving task success.
Feature-selection ablation — DexJoCo with Fast-WAM. Task success rates (%, ↑) on eight tasks. Hanoi Bi is excluded because successful references were unavailable and no steering was performed. Mean denotes the eight-task macro-average.
Controlled setting. Each variant steers exactly one feature from the same 64-feature candidate pool. Rank-1, Rank-8, and Rank-16 select the corresponding feature according to the offline outcome-separation score. Random-1 samples uniformly from the pool, with results averaged over three independent feature draws per task. Each feature receives its own reference banks from the same labeled rollouts. The SAE, intervention site, gating and calibration rules, coefficient search, and evaluation seeds remain unchanged.
Results. Rank-1 exceeds Rank-8, Rank-16, and Random-1 by 2.1, 2.4, and 4.4 percentage points, respectively, computed before rounding. It also improves over Base by 3.9 points, whereas random-feature steering does not improve the observed macro-average. These results provide direct behavioral evidence that selecting a highly ranked feature matters under the same controller, beyond demonstrating separation between offline reference descriptors. They support the proposed ranking criterion without assuming that every outcome-discriminative feature is equally useful for control.
R3-W2 — Feature interactions and the intervention mechanism are insufficiently understood.
Response. Thank you for raising this distinction. In SteerAct, single-feature steering means editing the hidden state along one selected SAE decoder direction. It does not mean that all other feature activations remain unchanged. The edit can affect other SAE features and subsequent network computations.
For the multi-feature variants, the implementation averages the selected decoder directions and searches for one shared coefficient using their aggregate activation history. The local response estimate evaluates how this combined direction affects the selected features, and measured post-edit activations provide feedback for subsequent decisions. The controller does not optimize an independent coefficient for each feature or explicitly constrain changes to unselected features.
Feature-count ablation — RoboTwin 2.0 with π₀.₅. Mean success rates (%, ↑) across nine tasks. Top-k uses the k highest-ranked features.
A single selected direction already provides substantial improvement over Base, while combining additional selected directions further improves observed success. These results establish the behavioral effectiveness of the tested interventions; they do not establish feature independence or explain how changes to nonselected features contribute to success.
To examine that mechanism directly, we plan to compare the full SAE activation vectors immediately before and after intervention and quantify changes in selected and unselected features. This diagnostic is not yet available. We will make the distinction between controlling one direction and changing only one feature explicit in the manuscript.