cs.RONov 18, 2026

Meaning Reaches Object Choice but Not Completion: Where Is a VLA Vulnerable to Landmark Synonyms?

Authors: Ryoichi Kurihara, Kota Sunamoto

Organizations: MAMEZO Co., Ltd., Tokyo, Japan

Abstract

VLA policies are fragile to rewording. Success rates alone cannot say whether a policy fails to choose the object or to complete the task, nor where in the policy to intervene. We study the released GR00T N1.7 on two LIBERO-Spatial tasks differing only in the landmark word telling which of two bowls to take. Synonyms absent from the training instructions send the policy to the bowl they name, unlike made-up words, yet complete 67.6% of one task's closed-loop episodes, against the original word's 100%. Before the connector between the vision-language model and the action generator, replacing only the landmark tokens' vectors with the other landmark's (pasting) sends nearly every episode to the other bowl. After the connector, pasting at the landmark, the remaining instruction, or the image tokens alone rarely moves the choice, but any two do. Hiding the landmark tokens from the action generator after the connector still lowers completion for the synonyms that fall short. Finetuning on diversified landmark words repairs that one task's completion for synonyms absent from finetuning; finetuning on the original instructions does not. Training only the connector on the diversified words suffices, as does copying the connector of a policy fully finetuned on them into one finetuned on the original instructions. An open-loop score on starting frames misses that weights outside the connector also carry part of the repair. For now, on this evidence, robustness to rewording should be judged in closed loop, with actions executed, intervening on both sides of the connector, scoring choice and completion separately.

Figures & tables

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Oct 14, 2026

Paraphrase Consensus Is Not Self-Consistency: Instruction Reliance and Robustness in VLA Policies

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.
Oct 7, 2026

Do Vision-Language-Action Models Understand Instructions? A Mechanistic Interpretability Study on Language Grounding

Vision-Language-Action models are designed to generalise across environments and task descriptions, raising the question of whether their action generation actually depends on the language instruction, or whether they largely rely on visual cues and superficial correlations. Robustness to variance in the visual and linguistic observation space is critical for real-world deployment, yet VLAs lack explicit grounding modules and instead rely on the intrinsic language grounding capabilities of their Vision-Language model backbones. For this reason, we conduct a controlled mechanistic interpretability study on the language grounding capabilities of two state-of-the-art Vision-Language-Action models, π0.5π_{0.5} and GR00T N1.7, by applying activation and attribution patching to the residual stream of the action generation modules. We systematically corrupt the task instruction of input samples of the LIBERO benchmark following five strategies: synonym replacement, semantic scaling, directional corruption, random object substitution, and empty string. Our experiments find that both models are comparatively insensitive to abstract rephrasing and to referencing non-existent objects, but react strongly to empty task descriptions and, especially, to directional language. During action generation, this sensitivity is concentrated in different loci for each model: mainly in the early, periodic cross-attention layers for GR00T N1.7, versus distributed across the earliest and selected later layers for π0.5π_{0.5}. For GR00T N1.7, directional perturbations drive some of the largest causal effects while leaving the internal representational geometry comparatively unchanged, a dissociation we do not observe clearly for π0.5π_{0.5}. Finally, the reliability of attribution patching is model-dependent: it closely tracks activation patching for GR00T N1.7 but not for π0.5π_{0.5}.
Oct 7, 2026

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

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: π0.5π_{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 π0π_0 checkpoint finetuned with rephrase augmentation still shows swings of up to 61 points. We characterize this sensitivity with statistically tested single-edit swings and an oracle phrase search, which shows that phrasing alone nearly closes the 21-point gap between in-distribution and out-of-distribution tasks. We then reduce it without modifying the policy. Because the sensitivity is systematic, it can be expressed as explicit rules: we score many phrasings of a few training tasks, have a large language model distill the evidence into ten to twenty rephrasing rules, and at deployment rewrite each incoming instruction once under these rules. The rules improve the frozen π0π_0 by 16 to 27% relative on twelve held-out tasks across adversarial, VLM-generated, and human-generated phrasings, with gains concentrated on out-of-distribution tasks. The pipeline replicates on π0.5π_{0.5} and LIBERO, lifting in-finetune success from 93.6% to 97.8%. The method requires no retraining and no per-step verification, and applies zero-shot to unseen tasks and instructions. Project website: https://sttawm.github.io/rephrase-before-you-act