← Back to feed

Rephrase Before You Act: Characterizing and Mitigating Language Sensitivity in Vision-Language-Action Models

L5 · ResearcherResearcharXiv· 10/7/2026

Addresses critical robustness issues in VLA models through systematic analysis and zero-shot mitigation techniques.

AI Summary

Research paper identifies systematic language sensitivity in vision-language-action models where minor phrasing changes cause performance swings up to 61 points, and proposes rule-based rephrasing method that improves success rates by 16-27%.

Excerpt

Vision-language-action models (VLAs) are strikingly sensitive to instruction phrasing and do not inherit the language robustness of the vision-language models they are built on. A one-word edit can move success by tens of points: $π_{0.5}$ turns on a LIBERO stove 100% of the time for "switch on the stove" and 2% for "switch on the hot plate", and a $π_0$ checkpoint finetuned with rephrase augmentation still shows swings of up to 61 points. We characterize this sensitivity with statistically test

Read Original
0 upvotes · 0 downvotes · 1 min read

Related Articles

L5 · ResearcherResearcharXiv
Decentralized SGD under Heavy-Tailed Noise: Optimal Convergence Rates and the Role of Gradient Clipping

Researchers prove that clipped decentralized SGD achieves optimal convergence rates under heavy-tailed noise conditions while maintaining linear speed-up across agents.

L5 · ResearcherResearcharXiv
Distilling Graph Geometry: Knowledge Gap from GNNs to MLPs

Researchers propose G²MLP, a graph geometry-aware MLP distillation method that addresses spectral failure modes in GNN-to-MLP knowledge transfer using Ollivier-Ricci curvature.

L5 · ResearcherResearcharXiv
Decoupling Exploration from Optimization in RLVR

Researchers propose Exploration-Distillation (ExpDis), a method that decouples exploration from optimization in RLVR training, improving mathematical reasoning performance across seven benchmarks.

L5 · ResearcherResearcharXiv
EngramEdit: Decoupled Knowledge Updates in LLMs through Conditional Memory

EngramEdit enables precise LLM knowledge updates via conditional memory, achieving 3x baseline accuracy in chain-of-thought reasoning while preserving unrelated knowledge.

L5 · ResearcherResearch@emollick.bsky.social
Ethan Mollick (@emollick.bsky.social): Some early first-hand accounts of the experience of encountering a narrow superhuman intelligence as mathematicians grapple with the hundreds of big AI proofs released by OpenAI. Problems solved in…

Mathematicians are struggling to understand 372 AI-generated proofs from OpenAI, including a proof of the long-unsolved Unique Games Conjecture, which appear superhuman but incomprehensible to humans.