Vision-language-action (VLA) models let users command robots in natural language, but the mapping from phrasing to behavior is often brittle: semantically similar instructions can produce very different actions, and some capabilities resist prompting altogether. The authors propose a framework that interactively searches for language sequences that improve closed-loop task performance, then distills these sequences into a test-time language feedback policy (LFP).

The key safeguard is a learned improvement head that predicts whether steering will actually help. By conformalizing this head, the system gains harmlessness guarantees: on out-of-distribution scenarios, it avoids interventions that would decrease performance relative to the original instruction. The approach works on arbitrary frozen pre-trained VLAs, requiring neither access to the original training distribution nor fine-tuning of the underlying model.

In experiments, the conformalized LFP improved base VLA performance by 24.7% in simulation and 65.0% in hardware on seen environments. Under visual and semantic perturbations with calibration samples, it produced recovery behaviors not observed with open-loop prompting, while maintaining strong harmlessness guarantees. The paper was accepted to CoRL 2026.