SteerAct Steers VLA Inner Features at Test Time, Boosting Robot Success

Researchers from Stanford, MIT and Harvard introduced SteerAct, which steers inner features of frozen VLA models at test time—using just 24 rollouts and a single feature per task—to raise real-robot success from 31% to 52% without retraining.

2026-10-21 ~ 2026-10-21 · 2 related posts

• SteerAct lifts real-robot success from 31.1% to 52.2% by steering one feature — liuzhuang1234 · 2026-10-21
• SteerAct steers test-time features to boost robot manipulation without retraining VLA models — yining_hong · 2026-10-21