Abstract
Vision-language-action (VLA) policies fine-tuned on LIBERO are trained and evaluated on the same ten instruction strings, and rewording the instruction changes base OpenVLA's success by -38.0 points. Test-time prompt ensembling, which improves language models, does not recover this loss in any setting we ran. Paraphrase consensus (the medoid action over K paraphrases) costs base OpenVLA 16.8 points when the training instruction is a member. Without it, the medoid of the paraphrases scores -32.4 points relative to the training instruction, and confidence-based selection scores -28.0. The same medoid over samples of one paraphrase improves on greedy decoding but stays -27.6 points below the training instruction, consistent with aggregation reducing variance without removing the bias the paraphrases share. A wrong-instruction probe measures how much of a policy's success depends on the instruction. Across three policies and four suites, paraphrases cost a policy only where this share is high, and the stronger policies lose a much smaller fraction of it than base OpenVLA. LoRA post-training on paraphrased instructions reduces the one-paraphrase effect to +2.0 points, against -32.0 for a matched control, reproduces across two seeds and on OpenVLA-OFT, transfers to withheld phrasings, and leaves the policy's reliance on the instruction unchanged.
Abstract (translated)
URL
https://arxiv.org/abs/2610.19438
