Paraphrase the Instruction, Not the Architecture: GSR Fixes VLA Brittleness Without Data Brute Force

Ask a vision-language-action model to pick up a mug when you rephrase "grasp the cup" as "take the vessel," and success rates can collapse — even though the model still "knows" the task internally. A paper submitted to arXiv on August 3, 2026, argues the failure is architectural: joint encoding of dynamic visuals and text introduces feature shifts that a brittle action policy cannot absorb.

The authors propose Grounded Semantic Re-binding (GSR), an intervention that decouples task semantics from visual observations and retrains a fresh action expert on fused representations. On the LIBERO-Para benchmark, they report success-rate improvements of up to 44.6 percent using only canonical demonstrations.

Probing the brittleness

Vision-language-action (VLA) models retain correct task identity internally even when paraphrases fail. The breakdown happens downstream: joint encoding shifts features in ways the action policy cannot translate into correct control commands.

The GSR intervention

GSR extracts task semantics and visual features independently, fuses them explicitly, and trains a re-initialized action expert on canonical demonstrations only. On LIBERO-Para, GSR reports a PRIDE score of 70.4, outperforming Xiaomi-Robotics-0 on instruction-generation capabilities in the authors' tests. ParaVLA, a 0.33B-parameter decoupled model, shows near-perfect robustness to instruction rewording.

Design over data brute force

Robust semantic grounding may be achievable through architecture rather than brute-force data scaling — critical for deployment economics in warehouses and homes where phrasing varies constantly.

### Sources

• arXiv — Grounded Semantic Re-Binding for Robust Instruction Generalization in Vision-Language-Action Models (August 3, 2026)