Literature
Display
← All papers

SteerAct: Unlocking Action Models through Few-Shot Temporal Steering

Kaichen Zhou · Zhiyan Li · Chongkai Gao · Yuhan Wang · Wenlong Huang · Maggie Wang · Fangneng Zhan · Manling Li · Qianqian Wang · Mac Schwager · Zhuang Liu · Li Fei-Fei · Mengyu Wang · Paul Pu Liang

Preprint posted · arXiv (preprint) ↗

Robotics9 min text · CondensedOriginal paper ↗
Contents

Overview

This edition covers all 23 pages and Appendices A–F. It preserves the six figures, seven complete tables, annotated Equations 5/6, full offline/online algorithm, all other equations and original project presentations/paired rollouts. Kaichen Zhou and Zhiyan Li contributed equally; Mengyu Wang and Paul Pu Liang share supervision. The 24-rollout adaptation budget is separate from prior task fine-tuning and shared SAE training.

Complete original Figure 1: Steering one feature at the right time. Using just 24 rollouts per task, SteerAct improves robot behavior by steering a single feature activation, without retraining the action model.
As the robot acts, SteerAct compares the patterns of how this feature activation changes with patterns from those existing rollouts to decide when to steer and how to steer. (a) Unsteered and steered real-robot executions illustrate how SteerAct improves task completion. (b) SteerAct improves task success across all 18 evaluated RoboTwin 2.0 and DexJoCo tasks with π 0.5 \pi_{0.5} . (c) On DexJoCo with π 0.5 \pi_{0.5} , SteerAct achieves the highest mean success among the evaluated methods.
Figure 1: Steering one feature at the right time. Using just 24 rollouts per task, SteerAct improves robot behavior by steering a single feature activation, without retraining the action model. As the robot acts, SteerAct compares the patterns of how this feature activation changes with patterns from those existing rollouts to decide when to steer and how to steer. (a) Unsteered and steered real-robot executions illustrate how SteerAct improves task completion. (b) SteerAct improves task success across all 18 evaluated RoboTwin 2.0 and DexJoCo tasks with π0.5\pi_{0.5}. (c) On DexJoCo with π0.5\pi_{0.5}, SteerAct achieves the highest mean success among the evaluated methods.

At a glance

SteerAct improves a frozen, task-fine-tuned action policy by editing one selected sparse-feature direction. Ordinal activation-history descriptors choose a task-specific feature offline, then support/failure gates and a signed coefficient search control interventions online. Main simulation gains over Base range from about 3.4 to 9.6 percentage points across four benchmark/backbone settings; physical success rises 31.1→52.2%. These are empirical improvements on selected tasks, not a guarantee that one feature always helps. Model weights stay fixed during steering, while the shared SAE and base task policies have been trained beforehand.

1 Introduction

Successful and failed rollouts can differ in how a feature rises/falls even when their mean activations resemble each other. A common temporal descriptor lets selection and intervention use that outcome-related structure without aligning rollouts by timestep. Hidden-state access and a suitable learned dictionary are required. The method avoids training a separate RL steering controller, rather than avoiding all task-specific preparation.

Complete original Figure 2: Temporal evidence guides when and how to steer. (a) The selected feature’s activation traces compare a successful reference, an unsteered rollout, and rollouts with adaptive steering. (b) The temporal-support score measures whether the observed activation history of current rollout provides sufficient evidence for comparison with patterns from successful and failed reference rollouts. (c) The steering coefficient α t \alpha_{t} determines the direction and strength of the activation edit; α t = 0 \alpha_{t}=0 indicates no intervention at that decision.
Figure 2: Temporal evidence guides when and how to steer. (a) The selected feature’s activation traces compare a successful reference, an unsteered rollout, and rollouts with adaptive steering. (b) The temporal-support score measures whether the observed activation history of current rollout provides sufficient evidence for comparison with patterns from successful and failed reference rollouts. (c) The steering coefficient αt\alpha_{t} determines the direction and strength of the activation edit; αt=0\alpha_{t}=0 indicates no intervention at that decision.

2 Related Work

Related work separates candidate action verification, sampler/noise guidance, representation steering and sparse-feature interpretation. SteerAct intervenes directly at one configured residual stream and uses observed activation history to choose when and how much to edit.

2.1 Inference-time Steering

Value-guided candidate selection and latent-noise controllers change generation through sampling, ranking or learned guidance. This method instead reuses a frozen hidden-state direction. It needs internal activations and a pretrained SAE, with its own offline preparation and online gate/search overhead.

2.2 Representation-level Steering and Sparse Autoencoders

Sparse autoencoders provide decoder columns as candidate intervention directions. Prior robot steering uses interpretable features or success/failure subspaces; SteerAct adds ordinal-history selection and online reference matching. Outcome separation is associational until an edit is evaluated, and a chosen feature need not have a clean semantic label. Source comparisons: Cunningham et al. (2024) on sparse features, Swann et al. (2026) on robot feature steering, and Miao et al. (2026) on COAST.

3 Methodology

Offline, complete reference-rollout descriptors form success and failure banks and determine one feature. Online, the current prefix is assessed against those fixed banks, then a local coefficient search edits the original hidden states. Both the SAE and action model remain frozen throughout deployment.

3.1 Problem Setting

The SAE maps hidden states to sparse activations with fixed decoder directions. The selected direction d is added as α(t)d to every targeted token at a robot decision. Robot decisions are distinct from diffusion/flow denoising iterations. Encoder-branch and action-expert interventions target different model locations; the latter applies enabled edits at every denoising step.

Equations 1–2
𝐳=E⁡(𝐡),𝐡^=Wdec​𝐳+𝐛.\mathbf{z}=E(\mathbf{h}),\qquad\widehat{\mathbf{h}}=W_{\mathrm{dec}}\mathbf{z}+\mathbf{b}. (1)
𝐡t,j′=𝐡t,j+αt​𝐝.\mathbf{h}^{\prime}_{t,j}=\mathbf{h}_{t,j}+\alpha_{t}\mathbf{d}. (2)

3.2 Temporal Descriptor

Feature intensity averages rectified SAE activations over measured tokens. At spacings one and two, all available triples receive one of six ordinal patterns, with ties ranked by temporal order. Six frequencies, normalized entropy, tie fraction and availability per spacing form an eighteen-dimensional descriptor. It discards absolute timing and much magnitude/order information, and therefore cannot uniquely identify execution phase. The monotone log preprocessing leaves these order patterns unchanged in exact arithmetic.

Equations 3–4
yf,e,t=1Ne,t​∑j=1Ne,tmax⁡{[E⁡(𝐡e,t,j)]f,0}.y_{f,e,t}=\frac{1}{N_{e,t}}\sum_{j=1}^{N_{e,t}}\max\left\{\left[E(\mathbf{h}_{e,t,j})\right]_{f},0\right\}. (3)
𝐯f,e,t=ϕ(yf,e,1:t)∈ℝ18.\mathbf{v}_{f,e,t}=\phi(y_{f,e,1:t})\in\mathbb{R}^{18}. (4)
Complete original Figure 3: SteerAct overview. Offline, an SAE fitted to frozen-policy
activations exposes candidate directions. Outcome-labeled rollouts determine a
feature and its success/failure banks of ordinal dynamics. Online, supported
failure-like prefixes enter a signed coefficient search; other prefixes retain α t = 0 \alpha_{t}=0 . The selected decoder direction modifies one configured residual
stream, after which the frozen policy generates actions.
Figure 3: SteerAct overview. Offline, an SAE fitted to frozen-policy activations exposes candidate directions. Outcome-labeled rollouts determine a feature and its success/failure banks of ordinal dynamics. Online, supported failure-like prefixes enter a signed coefficient search; other prefixes retain αt=0\alpha_{t}=0. The selected decoder direction modifies one configured residual stream, after which the frozen policy generates actions.

3.3 Offline Reference Construction and Feature Selection

For each candidate, the score is twice the mean cross-outcome distance minus within-success and within-failure mean distances, excluding self-pairs. Highest score selects the feature, decoder direction and fixed banks. This requires at least two examples of each outcome. Complete reference histories are compared with potentially short online prefixes; the support gate addresses evidence adequacy empirically. Missing numeric candidate-filter/intervention-site details limit exact reconstruction from the manuscript alone.

Complete original annotated Equation 5 with all colored callouts and arrows.
Original equation notation and annotations
Equation 5
f∗=arg⁡maxf∈𝒞⁡(2​d¯fS​F−d¯fS​S−d¯fF​F). (5)

Original callouts: Selected feature; Candidate feature; Candidate feature set; Success–failure distance; Within-success distance; Within-failure distance.

3.4 Online Reference-Guided Steering

The support gate samples plausible ordinal-frequency descriptors from a Dirichlet model and checks nearby pooled references. If supported, a scaled nearest-bank distance check detects failure-like prefixes; hysteresis controls both gates. When enabled, a bounded signed grid chooses the coefficient whose predicted last-intensity edit minimizes success-bank distance. This is local feature-statistic matching, not a model of future reward or task success. Re-encoding after an actual edit replaces the latest stored intensity, so subsequent history includes interventions. Many candidate values can share identical ordinal patterns; coefficient tie-breaking is not numerically specified.

Complete original annotated Equation 6 with all colored callouts and arrows.
Original equation notation and annotations
Equation 6
αt=arg⁡minα∈𝒜t​DS​(ϕ​(y1,…,yt−1,max⁡{0,yt+α​ρt​𝐰f∗⊤​𝐝})). (6)

Original callouts: Steering coefficient; Candidate coefficients; Success-bank distance; Temporal descriptor; Past intensities; Predicted current intensity.

4 Experiments

Evaluation spans two backbones, two selected nine-task simulation subsets, three physical tasks and eight-task timing/gating controls. All compared methods within a setting share the same fine-tuned policy checkpoint. The 24 outcome-labeled references and held-out evaluation rollouts serve different roles.

4.1 Experimental Setup

Simulation fine-tunes each backbone on 500 scripted successful demonstrations per task before freezing it. Steering then uses 24 separate outcome-labeled policy rollouts, and evaluation uses 200 held-out rollouts per task/method. The shared TopK SAE is trained on pooled hidden activations from those online rollout sets. Real π0.5 uses thirty task demonstrations and thirty evaluation trials per task. Baselines include one positive-only COAST variant; DSRL-SAC is reported only for π0.5. The latency table is per-decision inference on an RTX 6000, separate from offline fitting/selection costs. The noise-controller baseline follows Wagenmaker et al. (2025).

Complete original Table 1: SteerAct improves success across backbones
and manipulation benchmarks. Among the evaluated methods, SteerAct (Enc) achieves the highest mean success in all four benchmark–backbone settings with modest inference overhead.
Success rates are reported in percent ( ↑ \uparrow ); each Mean averages nine tasks.
Time reports per-decision inference latency (ms, ↓ \downarrow ), shared across datasets.
Enc/Dec denote encoder-branch/action-expert steering.
Gray highlights SteerAct (Enc); bold / underline mark the best/second-best distinct success rates within each backbone, including ties.
Table 1: SteerAct improves success across backbones and manipulation benchmarks. Among the evaluated methods, SteerAct (Enc) achieves the highest mean success in all four benchmark–backbone settings with modest inference overhead. Success rates are reported in percent (↑\uparrow); each Mean averages nine tasks. Time reports per-decision inference latency (ms, ↓\downarrow), shared across datasets. Enc/Dec denote encoder-branch/action-expert steering. Gray highlights SteerAct (Enc); bold/underline mark the best/second-best distinct success rates within each backbone, including ties.
Accessible table, original values and citation links
RoboTwin 2.0 DexJoCo
Method Bottles in Dustbin Stack Three Blocks Rank Blocks by Size Dual Shoes Dump Bin into Big Bin Open Microwave Turn Switch Mouse Pad Diverse Bottles Mean Hammer Nail Click Mouse Pick Bucket Pinch Tongs Fold Glasses Water Plant Hanoi Bi Microwave Bi Photo Bi Mean Time ↓\downarrow (ms/decision)
Backbone: π0.5\pi_{0.5}
Base 74 72 43 73 89 84 47 62 77 69.0 82 67 83 22 71 92 15 77 52 62.3 207.9
DSRL-SAC 80 78 40 57 49 100 65 71 100 71.1 80 60 70 18 74 96 10 69 47 58.2 3582.7
DrVLA 62 84 60 86 72 91 48 68 72 71.4 85 76 87 26 68 94 15 84 55 65.6 221.6
COAST 62 74 49 75 88 86 40 66 76 68.4 87 75 79 33 69 93 19 74 59 65.3 214.4
SteerAct (Dec) 79 83 39 81 92 87 44 67 66 70.9 79 74 75 21 87 96 17 91 59 66.6 278.1
SteerAct (Enc) 88 83 58 82 94 93 52 69 80 77.7 90 79 95 32 83 97 23 85 63 71.9 223.8
Backbone: Fast-WAM
Base 72 77 63 71 91 31 53 87 79 69.3 22 7 70 12 4 52 0 19 24 23.3 372.6
DrVLA 78 72 76 73 94 52 63 74 84 74.0 25 3 73 17 2 63 0 25 22 25.6 419.2
COAST 64 72 74 77 84 52 59 78 66 69.6 21 7 66 15 5 55 0 21 25 23.9 387.5
SteerAct (Dec) 84 79 75 81 91 54 60 90 80 77.1 24 5 63 21 5 57 0 25 17 24.1 423.9
SteerAct (Enc) 84 78 72 85 94 43 73 95 83 78.6 27 5 76 19 5 59 0 23 27 26.8 415.7

4.2 Evaluation Across Benchmarks, Embodiments, and Backbones

Encoder steering yields RoboTwin means 69.0→77.7% for π0.5 and 69.3→78.6% for Fast-WAM, plus DexJoCo 62.3→71.9% and 23.3→26.8%. The reported gains use unrounded task means. It improves 34/36 pairs, with Fast-WAM Click Mouse dropping 7→5% and Hanoi remaining zero. π0.5 encoder latency adds 15.9 ms (7.6%); Fast-WAM adds 43.1 ms (11.6%). Pooled two-proportion tests support aggregate gains under the reported trial model, without independent training-seed uncertainty or guarantees for every task. Fast-WAM’s dexterous performance remains much lower than π0.5’s.

4.3 Real-Robot Evaluation

On I2RT YAM with two D405 and one D435i cameras, encoder success is 16/30 Folding Cloth, 19/30 Cup Ping-Pong and 12/30 Push Cube, versus Base 10/30, 12/30 and 6/30. Across ninety trials this is 47/90 versus 28/90, a 21.1-point gain (reported pooled p=.004). The 12.2-point margin over DrVLA’s 36/90 has reported p=.10, so it is not statistically established by that test. Decoder steering averages 33.3%, close to Base’s 31.1%. Selected before/after examples illustrate behavior rather than expanding the trial denominator.

Complete original Table 2: SteerAct improves success on all three
real-robot tasks. With π 0.5 \pi_{0.5} frozen, SteerAct (Enc) achieves the
highest success on every evaluated task, reaching
52.2% average success compared with 31.1% for Base
and 40.0% for the strongest baseline, DrVLA.
Values are success rates (%, ↑ \uparrow ).
Enc/Dec denote encoder-branch/action-expert steering.
Gray highlights SteerAct (Enc); bold / underline mark the best/second-best
distinct results in each row.
Table 2: SteerAct improves success on all three real-robot tasks. With π0.5\pi_{0.5} frozen, SteerAct (Enc) achieves the highest success on every evaluated task, reaching 52.2% average success compared with 31.1% for Base and 40.0% for the strongest baseline, DrVLA. Values are success rates (%, ↑\uparrow). Enc/Dec denote encoder-branch/action-expert steering. Gray highlights SteerAct (Enc); bold/underline mark the best/second-best distinct results in each row.
Accessible table, original values and citation links
Task Base COAST DrVLA SteerAct Dec SteerAct Enc
Folding Cloth 33.3% 26.7% 40.0% 36.7% 53.3%
Cup Ping-Pong 40.0% 36.7% 56.7% 43.3% 63.3%
Push Cube 20.0% 0% 23.3% 20.0% 40.0%
Average 31.1% 21.1% 40.0% 33.3% 52.2%
Complete original Figure 4: Single-feature steering improves task completion
across manipulation settings. Representative unsteered and steered rollouts on RoboTwin 2.0, DexJoCo, and real robots illustrate how SteerAct corrects execution errors and enables successful task completion while keeping the underlying action model frozen.
Figure 4: Single-feature steering improves task completion across manipulation settings. Representative unsteered and steered rollouts on RoboTwin 2.0, DexJoCo, and real robots illustrate how SteerAct corrects execution errors and enables successful task completion while keeping the underlying action model frozen.

4.4 Ablation Studies

The ordinal descriptor scores 77.7%, versus 73.9% for the strongest tested alternative, a combined autocorrelation/spectral descriptor. Larger reference banks are not monotonically better: twelve total rollouts give 73.6%, default twenty-four 77.7%, and forty-eight/ninety-six 77.3%. Successful-only counts in another ablation exclude failed references and must not be read as total budgets. Top-two/four features reach 79.6/80.0%, so one feature is efficient rather than the best tested count. Fixed coefficients 1/2 retain gates but score 75.8/75.7%. On VLABench, full gates give 36.13%, versus frozen 23.63%; removing both gives 22.38%. These controls test the combined temporal/gating design, not an isolated causal role of behavioral phase.

Complete original Table 3: Ablations on RoboTwin 2.0 with π 0.5 \pi_{0.5} .
Values are mean success rates (%) across nine tasks.
Gray marks the default single-feature Enc variant
(24 total source rollouts).
Panels (b)–(d) use Enc; (c) counts successful source rollouts only.
Full results are in Appendix Table 5 .
Table 3: Ablations on RoboTwin 2.0 with π0.5\pi_{0.5}. Values are mean success rates (%) across nine tasks. Gray marks the default single-feature Enc variant (24 total source rollouts). Panels (b)–(d) use Enc; (c) counts successful source rollouts only. Full results are in Appendix Table 5.
Accessible table, original values and citation links

(a) Representation and intervention site

Method Base Marginals Increments Events Autocorr. Raw NN Raw DTW Dec Enc
Success 69.0 71.1 70.0 70.1 73.9 72.2 71.7 70.9 77.7

(b) Total rollouts

Rollouts Success
12 73.6
48 77.3
96 77.3

(c) Successful rollouts

Rollouts Success
8 74.4
24 77.6
48 77.9

(d) Steering choices

Setting Success
Top-22 features 79.6
Top-44 features 80.0
Fixed α=1\alpha=1 75.8
Fixed α=2\alpha=2 75.7
Complete original Table 4: Mean success rates (%, ↑ \uparrow ) across eight
VLABench tasks, with 200 evaluation episodes per task and setting ( π 0.5 \pi_{0.5} with SteerAct(Enc)).
Full results are in Appendix Table 6 .
Table 4: Mean success rates (%, ↑\uparrow) across eight VLABench tasks, with 200 evaluation episodes per task and setting (π0.5\pi_{0.5} with SteerAct(Enc)). Full results are in Appendix Table 6.
Accessible table, original values and citation links
Setting Frozen Delay 10 Delay 20 Without support gate Without failure gate Without gates Ours
Mean 23.63 24.63 24.88 23.13 29.88 22.38 36.13

5 Conclusion

A single sparse feature can improve frozen-policy execution when reference matching guides edits. Results remain backbone-, task- and intervention-site-dependent. The method supplies empirical adaptation rather than a safety guarantee or a solution to every zero-success task.

Original presentations and paired robot demonstrations

The primary project supplies two presentations, three real-robot failure/success pairs, eight simulation pairs and a separate mechanism pair: all 26 original movies remain unchanged, with the two original presentation posters and generated first-frame previews elsewhere. The interactive mechanism display maps video time to decisions differently for steered/unsteered clips (.37/.999 seconds per decision); that visualization timing is separate from Table 1’s measured inference latency. Selected movies do not add independent quantitative trials.

Steer Animation
Steer Method
pingpong transport failure
pingpong transport success
folding cloth failure
folding cloth success
push cube failure
push cube success
robotwin put bottles failure
robotwin put bottles success
robotwin pi05 rank blocks baseline
robotwin pi05 rank blocks steering
robotwin turn switch failure
robotwin turn switch success
robotwin place mouse pad failure
robotwin place mouse pad success
dexjoco pi05 pick bucket baseline
dexjoco pi05 pick bucket steering
dexjoco fold glasses failure
dexjoco fold glasses success
dexjoco fastwam water plant baseline
dexjoco fastwam water plant steering
dexjoco bimanual photograph failure
dexjoco bimanual photograph success
mechanism steering
mechanism baseline

Appendix

The complete appendix includes model context, algorithm and mathematical details, training/control protocols, full ablation tables, selected rollout filmstrips and limitations. Main figure/table/equation numbering continues through the appendix.

Appendix A Foundation Action Models

This work adapts already learned/fine-tuned representations rather than introducing a new action architecture. The full detailed ablation Tables 5 and 6 retain every task, reference budget, descriptor variant, coefficient setting, gate removal and tie in the reported means.

Complete original Table 5: Ablations on RoboTwin 2.0 with π 0.5 \pi_{0.5} :
(a) temporal descriptors, matching rules, feature counts,
steering strengths, and intervention sites;
(b) total source rollouts; (c) successful source rollouts. (d) feature count and coefficient adaptation.
Entries are success rates (%, ↑ \uparrow ); Mean averages over the nine displayed tasks.
Matching controls use Enc and retain the ordinal-based support gate.
Gray highlights the full method. Bold / underline mark the best/second-best distinct results per column within each group, including ties.
Table 5: Ablations on RoboTwin 2.0 with π0.5\pi_{0.5}: (a) temporal descriptors, matching rules, feature counts, steering strengths, and intervention sites; (b) total source rollouts; (c) successful source rollouts. (d) feature count and coefficient adaptation. Entries are success rates (%, ↑\uparrow); Mean averages over the nine displayed tasks. Matching controls use Enc and retain the ordinal-based support gate. Gray highlights the full method. Bold/underline mark the best/second-best distinct results per column within each group, including ties.
Accessible table, original values and citation links
Method / Setting Put Bottles in Dustbin Stack Three Blocks Rank Blocks by Size Place Dual Shoes Dump Bin into Big Bin Open Microwave Turn Switch Place Mouse Pad Pick Diverse Bottles Mean
(a) Representation and steering ablations
Base 74 72 43 73 89 84 47 62 77 69.0
Marginals 76 76 48 77 88 77 49 67 82 71.1
Increments 78 76 39 78 89 83 43 61 83 70.0
Events 76 78 45 77 89 82 50 54 80 70.1
Autocorrelation 82 83 49 83 93 86 49 56 84 73.9
Raw NN 75 81 52 76 87 89 52 66 72 72.2
Raw DTW 78 84 49 62 92 96 50 66 68 71.7
SteerAct (Dec) 79 83 39 81 92 87 44 67 66 70.9
SteerAct (Enc) 88 83 58 82 94 93 52 69 80 77.7
(b) Total source rollouts — SteerAct (Enc)
12 rollouts 77 79 53 79 93 95 44 64 78 73.6
48 rollouts 79 84 63 84 94 96 47 70 79 77.3
96 rollouts 74 87 57 79 94 94 51 74 86 77.3
(c) Successful source rollouts — SteerAct (Enc)
8 successful rollouts 77 79 60 75 90 88 54 66 81 74.4
24 successful rollouts 79 86 64 77 94 94 49 71 84 77.6
48 successful rollouts 76 87 59 80 94 96 52 70 87 77.9
(d) Feature count and coefficient adaptation — SteerAct (Enc)
Top-22 features 90 87 59 84 94 96 51 73 82 79.6
Top-44 features 90 89 62 81 95 96 54 71 82 80.0
Fixed α=1\alpha=1 86 81 66 77 90 93 44 66 79 75.8
Fixed α=2\alpha=2 86 88 63 68 86 97 49 59 85 75.7
Complete original Table 6: Timing and gating comparisons on VLABench.
Success rates (%, ↑ \uparrow ) use 200 evaluation episodes
per task and reported setting; Mean is the task macro-average.
Gray highlights maturity-based steering. Bold / underline indicate the best/second-best
distinct reported results per column, including ties.
Table 6: Timing and gating comparisons on VLABench. Success rates (%, ↑\uparrow) use 200 evaluation episodes per task and reported setting; Mean is the task macro-average. Gray highlights maturity-based steering. Bold/underline indicate the best/second-best distinct reported results per column, including ties.
Accessible table, original values and citation links
Method Painting Book Drinking Chemistry Tube Poker Mahjong Toy Fruit Mean
Frozen 17 46 15 47 18 6 7 33 23.63
Delay 10 17 51 22 47 13 13 5 29 24.63
Delay 20 17 44 19 52 16 11 12 28 24.88
Without support gate 15 49 23 49 12 8 5 24 23.13
Without failure gate 13 48 27 53 22 18 25 33 29.88
Without gates 17 43 24 42 22 9 3 19 22.38
SteerAct 34 50 39 60 22 14 26 44 36.13
Complete original Algorithm 1, including Require, stage headings, all line numbers and re-encoding step.
Algorithm 1 SteerAct: Offline Selection and Online Steering
Original selectable algorithm
1: Frozen action model, pretrained SAE, outcome-labeled rollouts, candidate features 𝒞\mathcal{C}, and fixed control settings.
2: Offline: select a feature and its references
3: For each f∈𝒞f\in\mathcal{C}, construct banks ℬfS,ℬfF\mathcal{B}_{f}^{S},\mathcal{B}_{f}^{F} from descriptors in Eq. (3) and Eq. (4).
4: Select f∗f^{*} and its decoder direction 𝐝\mathbf{d} using Eq. (5).
5: Retain ℬS,ℬF\mathcal{B}_{S},\mathcal{B}_{F} for f∗f^{*}; calibrate the support radius and coordinate scales (Appendix B).
6: Online: steer a new rollout
7: Initialize empty history, α0←0\alpha_{0}\leftarrow 0, and closed gates with zeroed counters.
8: for each robot decision tt until termination do
9:   Run the model to the intervention site; obtain 𝐙t\mathbf{Z}_{t} and the unedited intensity yty_{t} using Eqs. (1) and (3).
10:   Append yty_{t}; compute 𝐯t=ϕ(y1:t)\mathbf{v}_{t}=\phi(y_{1:t}) using Eq. (4), retaining the pattern counts.
11:   Compute mtm_{t} using Sec. 3.4; update the temporal support gate using mt≥τmm_{t}\geq\tau_{m}.
12:   If support is open, update the failure gate using Sec. 3.4; otherwise reset it and its counters.
13:   αt←0\alpha_{t}\leftarrow 0.
14:   if both gates are open then
15:    αt←BoundedSearch(y1:t,𝐙t,αt−1)\alpha_{t}\leftarrow\textsc{BoundedSearch}(y_{1:t},\mathbf{Z}_{t},\alpha_{t-1}), evaluating candidates with Eq. (6).
16:   end if
17:   if αt≠0\alpha_{t}\neq 0 then
18:    Edit each targeted token using Eq. (2).
19:    Re-encode the edited states and replace yty_{t} with the measured intensity from Eq. (12).
20:   end if
21:   Retain the applied αt\alpha_{t} for the next decision.
22:   Resume the model from the resulting hidden states and execute its action output.
23: end for

Appendix B Additional Details of Reference-Guided Steering

Algorithm 1 preserves offline bank construction/selection/calibration and every online gate, coefficient, re-encoding and execution step. Two reference rollouts per outcome are a minimum for the within-bank score, not a guarantee that such examples are easy to collect for a failing policy.

B.1 Details of the Support Gate

Each spacing gets six ordinal counts plus 1/2 pseudo-counts. Five hundred sampled probability vectors supply approximate support uncertainty; adjacent overlapping triples violate the independent-count model, as the appendix acknowledges. Tie/availability values stay observed rather than sampled. Offline radius calibration uses three times a specified quantile of pooled k-th-neighbor distances, excluding self-neighbors. The quantile, neighbor counts, minimum triple counts and hysteresis lengths are described as fixed but their numeric values are absent from the four-row hyperparameter table. Support is a representation-coverage check, not a calibrated probability of successful or physically safe execution.

B.2 Details of the Failure Gate

Outcome matching uses fixed coordinate scales combining equally weighted success/failure population variances and their squared mean difference, with a 1e-8 floor. It averages equal effective nearest-neighbor counts from both banks. Only the support gate uses unscaled distances and sampled descriptors; feature selection also stays raw Euclidean. The main text describes outcome neighbors using k(sup), while the appendix distinguishes a separate requested k; the latter supplies the detailed rule without a reported numeric setting.

Equations 7–8
sq=max⁡{10−8,σS,q2+σF,q22+(μS,q−μF,q)24}.s_{q}=\max\!\left\{10^{-8},\sqrt{\frac{\sigma_{S,q}^{2}+\sigma_{F,q}^{2}}{2}+\frac{(\mu_{S,q}-\mu_{F,q})^{2}}{4}}\right\}. (7)
dist⁡(𝐯,𝐮)=[∑q=118(vq−uqsq)2]1/2.\operatorname{dist}(\mathbf{v},\mathbf{u})=\left[\sum_{q=1}^{18}\left(\frac{v_{q}-u_{q}}{s_{q}}\right)^{2}\right]^{1/2}. (8)

B.3 Details of the Steering Coefficient

Candidate response is α·ρ·wᵀd under an affine selected preactivation, unchanged active set and identical measured/edited token sets. A TopK SAE is not globally affine, and threshold/rank changes can invalidate this local approximation. Nonfinite or tiny estimated response triggers abstention. Search changes only the latest raw intensity and compares the resulting ordinal descriptor with success references, then actual edited states are re-encoded. It does not predict a future trajectory. The algorithm’s measured-intensity cross-reference points to the approximate-response equation; the stated re-encoding operation is retained. Log(1+y/s) preprocessing is monotone and does not add ordinal information.

Equations 9–12
ρt=1Nt∑j=1Nt{[E(𝐡t,j)]f∗>0},\rho_{t}=\frac{1}{N_{t}}\sum_{j=1}^{N_{t}}\mathbf{1}\!\left\{\left[E(\mathbf{h}_{t,j})\right]_{f^{*}}>0\right\}, (9)
[E⁡(𝐡)]f∗=𝐰f∗⊤​𝐡+bf∗,\left[E(\mathbf{h})\right]_{f^{*}}=\mathbf{w}_{f^{*}}^{\top}\mathbf{h}+b_{f^{*}}, (10)
[E⁡(𝐡t,j′​(α))]f∗≈[E⁡(𝐡t,j)]f∗+α​𝐰f∗⊤​𝐝.\left[E\!\left(\mathbf{h}^{\prime}_{t,j}(\alpha)\right)\right]_{f^{*}}\approx\left[E(\mathbf{h}_{t,j})\right]_{f^{*}}+\alpha\mathbf{w}_{f^{*}}^{\top}\mathbf{d}. (11)
y^t​(α)≈yt+α​ρt​𝐰f∗⊤​𝐝.\widehat{y}_{t}(\alpha)\approx y_{t}+\alpha\rho_{t}\mathbf{w}_{f^{*}}^{\top}\mathbf{d}. (12)

Appendix C Experimental Protocol and Hyperparameters

The full protocol distinguishes 500 successful scripted demonstrations for simulation fine-tuning, thirty demonstrations for real fine-tuning, twenty-four outcome-labeled online references per task, pooled SAE training and separate 200/30-trial evaluations. The SAE expands to eight times hidden width and retains the hundred largest activations per token. Table 7 reports M=500, support threshold .65, coefficient bound 10 and grid spacing .5. Exact layer/token sites, some candidate controls, gate thresholds/counts, response minimum and coefficient-tie handling are not fully specified in the manuscript. Frozen steering should not be read as adaptation from only twenty-four episodes starting with no task training.

Complete original Table 7: Shared SteerAct control hyperparameters.
Table 7: Shared SteerAct control hyperparameters.
Accessible table, original values and citation links
Parameter Symbol Value
Support samples MM 500
Support-gate threshold τm\tau_{m} 0.65
Maximum coefficient magnitude AmaxA_{\max} 10
Coefficient grid spacing δα\delta_{\alpha} 0.5

Appendix D Representation and Intervention-Site Ablations

Marginals, increments, threshold events and autocorrelation/spectral summaries are custom control descriptors. Raw nearest-neighbor/DTW variants compare histories directly but retain the ordinal support gate, so those rows do not remove all ordinal machinery. Enc/Dec compare intervention locations. Complete definitions remain in these notes and full task results remain in the retained tables; no alternative is described as an exact reproduction of a separate published system.

Appendix E Qualitative Paired Simulation Rollouts

The two full filmstrips retain ten selected matched-initial-scene comparisons, all unsteered/steered rows, instructions and temporal frames. They intentionally show failure-to-success cases. They illustrate possible corrections and do not estimate their frequency independently of the quantitative tables.

Appendix F Limitations.

Timing evidence comes from gate and fixed-delay ablations, without an experiment isolating behavioral phase from instantaneous state. Hidden-state access, SAE feature quality and underlying policy competence constrain transfer. Some tasks still regress or remain at zero, and multiple features outperform the single-feature default in the reported ablation.

Complete original Figure 5: Simulated rollouts on RoboTwin 2.0.
Each panel compares the unsteered frozen backbone (top) with
SteerAct (bottom) from a matched initial scene; time advances
from left to right.
Figure 5: Simulated rollouts on RoboTwin 2.0. Each panel compares the unsteered frozen backbone (top) with SteerAct (bottom) from a matched initial scene; time advances from left to right.
Complete original Figure 6: Paired simulated rollouts on RoboTwin 2.0 and DexJoCo (continued). The unsteered and steered rows begin from matched initial scenes, and frames advance from left to right.
Figure 6: Paired simulated rollouts on RoboTwin 2.0 and DexJoCo (continued). The unsteered and steered rows begin from matched initial scenes, and frames advance from left to right.
Concept demonstration

Open image ↗