acceptodds
Under review as a conference paper at ICLR 2027

STEERACT: Temporally Gated Single-Feature Steering of Frozen Action Models

11 views

Abstract

Improving pretrained robot action models from a few rollouts requires identifying both what to steer and when to intervene and the sign of the correction. We find that successful and failed executions exhibit distinct temporal patterns in sparse latent space, and that steering one feature per task can improve closed-loop success without updating model weights. Based on this finding, we introduce STEERACT, a reference-guided steering method using a pretrained sparse autoencoder. Offline, STEERACT summarizes sparse-feature activation histories from successful and failed rollouts via temporal descriptors, constructs outcome-conditioned reference banks, and selects the feature with the strongest outcome separation. Online, a support gate assesses whether the observed activation history provides sufficient evidence for reference comparison, while a failure gate checks whether its descriptor is closer to failed than successful references. When both gates are open, a bounded search selects a signed correction predicted to move the current descriptor closer to the success bank. The selected feature direction and reference banks remain fixed; only the intervention coefficient adapts during execution. Using 24 outcome-labeled reference rollouts per task, improves task success by approximately 10% relative to unsteered policies in evaluations spanning the vision–language–action model , the world–action model Fast-WAM, RoboTwin 2.0, DexJoCo, and real-world manipulation. Code and Demo can be found at: https://anonymous-submission-20.github.io/steeract-sub.github.io/.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision–language–action model, the world–action model, inference-time steering.

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_i2dAOpenReview ↗
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
  1. 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?

  2. 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?

  3. 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.

  4. Intervention location. Which exact layers and tokens are used for each backbone, and what determines their selection? Is performance stable across different intervention layers?

  5. 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?

  6. 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?

  7. 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-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.

Feature selectionHammer NailClick MousePick BucketPinch TongsFold GlassesWater PlantMicrowave BiPhoto BiMean
Base2277012452192426.3
Random-12257112449192425.8
Rank-162557614557202027.8
Rank-82757615554202228.0
Rank-1 (Ours)2757619559232730.1

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.

Steering configurationMean success (%)
Base69.0
Top-177.7
Top-279.6
Top-480.0
Top-880.8

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.

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.

MethodRoboTwin 2.0DexJoCo
DSRL-SAC71.158.2
COAST68.465.3
SteerAct (Enc)77.771.9

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.

SettingMean success (%)
Base69.0
Marginal statistics71.1
Multiscale increments70.0
Event-based summaries70.1
Autocorrelation-based matching73.9
Raw-history nearest neighbor72.2
Raw-history DTW71.7
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 settingMean success (%)
Delay 1068.3
Delay 2070.6
Without support gate69.8
Without failure gate67.8
SteerAct71.9

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 task1224 (default)4896
Mean success (%)73.677.777.377.3

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.

BenchmarkEmbodimentBackboneBaseSteerAct (Enc)Gain (pp)
RoboTwin 2.0Grippersπ₀.₅ (VLA)69.077.7+8.7
RoboTwin 2.0GrippersFast-WAM (WAM)69.378.6+9.2
DexJoCoDexterous handsπ₀.₅ (VLA)62.371.9+9.6
DexJoCoDexterous handsFast-WAM (WAM)23.326.8+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.

Reviewer_Gq2bOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SteerAct, a method for improving frozen robot action models through inference-time activation steering. It selects a single SAE feature using successful and failed rollouts, then uses temporal activation patterns to determine when and how strongly to intervene. Experiments on π0.5 and Fast-WAM show improvements across simulation benchmarks and real-world tasks.

Critical strengths
  1. The paper addresses a practical problem: improving pretrained robot policies without further fine-tuning.

  2. The combination of temporal feature selection and adaptive steering is interesting and reasonably motivated.

  3. The evaluation is extensive, covering different backbones, simulation benchmarks, and real-robot tasks, with generally consistent improvements and modest inference overhead.

Critical weaknesses
  1. The method assumes that steering feature activations toward successful references will improve robot behavior. However, features that distinguish successful and failed rollouts may not necessarily be useful for controlling actions. More direct evidence, such as random-feature or shuffled-label controls, would help justify this assumption.

  2. I am not fully convinced that the gains come from modeling temporal patterns. The current ablations compare different descriptors and fixed intervention delays, but simpler alternatives based on current activations or running statistics are missing.

  3. The paper uses 24 reference rollouts per task, requiring at least two successful and two failed trajectories. It is unclear how this is achieved for tasks with very low baseline success rates, and whether additional trials are needed to collect these references.

  4. The ablation on steering coefficients is not entirely fair, as the fixed-coefficient baselines only use positive values, while the proposed method can adapt both sign and magnitude. This makes it difficult to determine where the improvement comes from.

Questions for authors
  1. Can the authors provide evidence that the selected SAE features causally influence task success, rather than merely correlate with it? How does random-feature steering perform?

  2. How does SteerAct compare with gating based only on current activations or simple running statistics? Would shuffling the temporal history affect performance?

  3. How are successful reference rollouts obtained for tasks with very low baseline success rates? Does the actual collection budget exceed 24 trials?

  4. Can the authors separate the contributions of adaptive steering direction, magnitude, and intervention timing?

7 comments

R2-W1 — Does outcome-guided feature selection lead to effective control?

Response. Thank you for raising this question. We agree that distinguishing successful from failed rollouts does not, by itself, establish a feature’s usefulness for control. We therefore evaluated random-feature steering and outcome-bank reversal on eight DexJoCo tasks with Fast-WAM. Hanoi Bi is excluded from this intervention analysis because successful references were unavailable and steering was not enabled.

Controlled comparisons.

  • Random feature: uniformly select one feature from the same 64-feature candidate pool and construct its reference banks using the correct outcome labels. Results are averaged over three independent feature draws per task.
  • Swapped outcome banks: retain the original selected feature and reference descriptors, but exchange the success and failure banks during online control. This reverses their roles in both the failure gate and the correction objective.

All steering variants use one feature and share the same SAE, reference rollouts, intervention site, gating and calibration rules, coefficient-search procedure, and evaluation seeds.

Control-Usefulness Ablation on DexJoCo

Task success rates (%, ↑) with Fast-WAM. Mean denotes the macro-average across the same eight tasks.

VariantHammer NailClick MousePick BucketPinch TongsFold GlassesWater PlantMicrowave BiPhoto BiMean
Base2277012452192426.3
Random feature2257112449192425.8
Swapped outcome banks132579344131219.1
Original (Ours)2757619559232730.1

Results. The original method achieves 30.1% mean success, improving on Base by 3.9 percentage points and random-feature steering by 4.4 points. Reversing the outcome banks while retaining the selected feature reduces mean success by 11.0 points. Differences are computed before rounding the task averages.

These results provide behavioral evidence that both the selected feature and the correct assignment of outcome references contribute to effective steering. The random-feature comparison supports the proposed selection criterion, while the bank-reversal comparison shows that the controller benefits from correctly distinguishing successful and failed references. Both conclusions are based on completed-task success, rather than descriptor improvement alone.

R2-W2 — Temporal-pattern modeling is not sufficiently isolated.

Response. Thank you for suggesting these controls. We agree that our original ablations did not establish whether a single activation measurement or a simple running statistic could suffice. We therefore added a comparison on nine DexJoCo tasks with π₀.₅. We compare current activation (the selected feature’s intensity at the current decision), running mean (its mean intensity over the observed rollout prefix), and entropy-only (the two normalized ordinal-pattern entropies, without the individual pattern frequencies) against our full 18-dimensional temporal descriptor.

We hold the selected feature, reference rollouts, ordinal-pattern support gate, coefficient grid, and evaluation seeds fixed. Each variant uses its own descriptor for outcome matching and for scoring candidate corrections.

DescriptorHammer NailClick MousePick BucketPinch TongsFold GlassesWater PlantHanoi BiMicrowave BiPhoto BiMean (%)
Current activation88779230849521886471.0
Running mean83709128839420816067.8
Entropy-only85739228819222856168.8
Full temporal descriptor90799532839723856371.9

Current activation is a strong control: the full descriptor improves mean success by only 0.9 percentage points, and current activation performs better on three of the nine tasks. The full descriptor improves over the running mean and entropy-only variants by 4.1 and 3.1 percentage points, respectively. These results support using the full descriptor for outcome matching and correction scoring in our setting, but do not establish that temporal modeling accounts for most of the steering gain: the selected feature and the temporal support gate are shared across all variants.

R2-W3 — The 24-reference setting may require additional collection attempts.

Response. SteerAct requires at least two successful and two failed reference rollouts. We start with 24 rollout attempts per task–backbone setting. If the initial batch contains one success, we collect additional rollouts to obtain enough successful references. If none of the 24 attempts succeeds, we stop collection and evaluate the original policy without steering.

Under the default collection protocol, only two settings required additional attempts, both with Fast-WAM on DexJoCo: Fold Glasses (4% base success) required 50 attempts in total, taking 38.02 minutes, and Click Mouse (7% base success) required 31 attempts, taking 21.01 minutes. Both had base success rates below 10%; the decision to collect more was based on the number of successes observed in the initial batch, not on a preset success-rate threshold.

Hanoi Bi with Fast-WAM was the only setting in which none of the initial 24 attempts succeeded. We stopped after 20 minutes without constructing a success bank or applying steering. The reported 0% success rate therefore reflects the unsteered policy. All other settings obtained sufficient references within 24 attempts.

BenchmarkTaskBackboneBase success (%)Total attemptsMean time per attempt (s)Total collection time (min)
RoboTwin 2.0Turn Switchπ₀.₅472410.284.11
RoboTwin 2.0Turn SwitchFast-WAM532413.155.26
DexJoCoFold Glassesπ₀.₅712439.2515.70
DexJoCoFold GlassesFast-WAM450 (+26)45.6338.02
DexJoCoClick MouseFast-WAM731 (+7)40.6621.01
DexJoCoHanoi BiFast-WAM02450.0020.00

Collection costs include all attempts, successful or failed. The per-attempt times for Click Mouse and Hanoi Bi are calculated from their reported totals.

We will clarify that 24 is the initial collection budget, not a guarantee of sufficient references. The cost can be higher when successes are rare, and steering is unavailable if the collection yields no successful references. This limitation also applies to methods such as COAST that build their steering representations from both successful and failed rollouts.

Reviewer_652VOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

STEERACT proposes a temporally gated activation-steering method to improve robot task success without updating the action model’s parameters. Using a shared sparse autoencoder (SAE) and a small set of successful and failed reference rollouts, it selects one feature per task whose temporal activation patterns best distinguish the two outcomes. During execution, a support gate checks whether the current activation history has sufficient reference support, while a failure gate determines whether it more closely resembles failed rollouts. When both gates permit intervention, the method searches over signed coefficients to move the current feature descriptor toward successful references and applies the corresponding hidden-state edit. Experiments with π0.5 and Fast-WAM on RoboTwin 2.0, DexJoCo, and three real-robot tasks show improved task success with modest inference overhead. Ablations further examine the temporal descriptor, gating mechanisms, intervention site, and coefficient adaptation.

Critical strengths
  1. The problem is practically relevant. The paper investigates how to improve existing robot policies using a small amount of additional rollout experience while keeping policy parameters frozen. This is appealing for avoiding retraining and reducing adaptation costs during deployment.
  2. The evaluation is broad, with consistent improvements. The method is evaluated across two backbone models, two simulation benchmarks, and real-robot tasks. In the main experiments, it achieves higher observed success rates than the base policy in 34 of 36 task–backbone combinations, suggesting that the gains are not driven by only a few tasks.
Critical weaknesses
  1. Robustness to diverse execution modes is unclear. A task may admit multiple successful strategies and failure types, yet the paper does not establish whether the selected feature reliably distinguishes outcomes across these modes. With only 24 reference rollouts, feature selection and gating may favor dominant modes, while their reliability on underrepresented strategies and failure types remains unverified.
  2. Feature selection lacks direct controls, and adaptive correction needs stronger validation. The paper does not compare random or lower-ranked SAE features under the same gating and search procedure. Adaptive coefficients achieve 77.7% success, versus 75.8%/75.7% with fixed coefficients (Table 3(d)), but systematic diagnostics linking proxy improvement to more effective interventions are missing.
Questions for authors
  1. The abstract states that code and demos are available at the linked page (anonymous-submission-20.github.io/steeract-sub.github.io), but I could not find the code there. Could the authors check the link or provide the code?
  2. For policies with a reported 0% success rate, how are the required successful reference trajectories obtained, and how many rollout attempts and how much collection time does this require?
  3. see weakness
7 comments

R1-W1 — Robustness to multiple successful strategies and failure modes.

Response. Thank you for highlighting this concern. SteerAct retains individual rollout descriptors, allowing multiple successful and failed activation patterns to inform steering when represented in the reference banks. The support gate withholds intervention when the observed descriptor lacks sufficient reference support.

Using 24 reference rollouts per task, SteerAct (Enc) improves success in 34 of 36 simulated task–backbone settings, with the selected feature and banks fixed across evaluation episodes with varying layouts. In our RoboTwin 2.0 rollout-budget ablation with π₀.₅, increasing the budget from 24 to 48 rollouts does not improve observed mean success. Together, these results suggest that 24 rollouts already capture useful temporal patterns for adaptation in many evaluated tasks.

We agree that these aggregate results do not establish reliability for every execution mode. This reflects a limitation of reference coverage in our few-shot adaptation setting: rare strategies and failure types may remain insufficiently represented. Our results support task-level improvements under a small reference budget, while robustness to underrepresented modes remains unverified.

R1-W2.1 — Does temporal ranking identify more effective steering features?

Response. Thank you for highlighting the need for direct feature-selection controls. We added comparisons with lower-ranked and randomly selected features on nine DexJoCo tasks with Fast-WAM, keeping the steering procedure unchanged.

Feature-Selection Ablation on DexJoCo

Task success rates (%, ↑). Mean denotes the nine-task macro-average, including Hanoi Bi.

Feature selectionHammer NailClick MousePick BucketPinch TongsFold GlassesWater PlantHanoi BiMicrowave BiPhoto BiMean
Rank-1 (Ours)27576195590232726.8
Rank-827576155540202224.9
Rank-1625576145570202024.7
Random-122571124490192422.9
Baseline22770124520192423.3

Controlled procedure. For tasks with sufficient references from both outcomes, each variant selects 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 uniformly samples a candidate, with results averaged over three independent feature draws per task. Each feature receives its own reference banks constructed from the same labeled rollouts. The SAE, intervention site, gating and calibration rules, coefficient search, and evaluation seeds remain unchanged. Settings without sufficient outcome references retain the unsteered policy.

Results. Rank-1 achieves 26.8% mean success, exceeding Rank-8, Rank-16, and Random-1 by 1.9, 2.1, and 3.9 percentage points, respectively. These comparisons directly support the proposed ranking as a useful criterion for selecting steering targets under the same controller. The evidence concerns downstream task success, beyond the offline separation score used to select the feature.

R1-W2.2 — Does adaptive correction improve task success?

Response. To strengthen the validation of adaptive correction, we conducted coefficient ablations on nine DexJoCo tasks with π₀.₅. These comparisons restrict the available coefficients while retaining the selected feature and the rest of the controller, separating the benefits of sign and magnitude selection.

Steering-Coefficient Ablation on DexJoCo

Task success rates (%, ↑). Mean denotes the nine-task macro-average, including Hanoi Bi.

Coefficient selectionHammer NailClick MousePick BucketPinch TongsFold GlassesWater PlantHanoi BiMicrowave BiPhoto BiMean
Base82678322719215775262.3
Fixed +188729133859617815568.7
Fixed +290719535819017835768.8
Adaptive sign, magnitude 190729333849417835969.4
Adaptive sign, magnitude 290769333859620856070.9
Adaptive magnitude, positive sign90799233809620816170.2
Adaptive sign and magnitude (Ours)90799532839723856371.9

Controlled procedure. All variants use the same selected feature, reference banks, temporal descriptor, gate rules, candidate-scoring objective, and evaluation seeds. Only the available coefficient set changes:

  • Fixed positive correction: {0, +1} or {0, +2}.
  • Adaptive sign with fixed nonzero magnitude: {−1, 0, +1} or {−2, 0, +2}.
  • Adaptive magnitude with positive sign: {0, 0.5, …, 10}.
  • Adaptive sign and magnitude: {−10, −9.5, …, 10}.

Zero remains available in every setting, and no correction is applied unless a candidate improves the matching objective. Thus, the fixed-coefficient controls retain the same abstention rule.

Results. Allowing negative corrections at a fixed magnitude improves mean success by 0.8 points for magnitude 1 and 2.1 points for magnitude 2. The full signed grid achieves 71.9%, exceeding these adaptive-sign controls by 2.4 and 1.0 points, respectively, and the positive-only variable-magnitude grid by 1.7 points. It also exceeds Fixed +1 and Fixed +2 by 3.2 and 3.1 points. All differences are computed before rounding.

These results provide task-level evidence for adapting both sign and magnitude under the tested configurations. They do not, by themselves, establish how accurately the predicted descriptor improvement measures the behavioral benefit of an individual intervention; that relationship requires a separate intervention-level diagnostic.

R1-Q1 — Code availability

Response. Thank you for pointing this out. The code is accessible through the “Code” button at the bottom of the project page, which is linked at the end of our abstract. We will make the repository link more prominent on the project page and include a direct link in the final paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.