NTH

Say the Same, Act Differently: Diagnosing and Steering Text-Action Divergence in Reasoning VLAs

AuthorsZihao Feng, Qingzhao Zhang, Chunyu Xia, Bo Yu, Z. Morley Mao, Ramesh Govindan

AffiliationsUniversity of Southern California University of Arizona General Motors

November 6, 2026 2 min read
Watch on YouTube
The one-line take

A VLA can give exactly the same explanation while its actions become unsafe, so explanations alone are not enough to monitor embodied AI.

Key results

1.87%
Alpamayo action-direction overlap with text gradients

Mean TNAS diagnostic overlap in the driving model.

5.28%
InstructVLA action-direction overlap with text gradients

Mean TNAS diagnostic overlap in the manipulation model.

52.6%
Full-image silent collision rate

New collisions with an unchanged rationale across 154 obstacle-annotated driving scenes.

35.1%
Localized-patch silent collision rate

New collisions with an unchanged rationale under a forward-camera patch.

89.5%
Alpamayo exact rationale preservation

Exact preservation at the strongest TNAS guidance setting.

85.0%
InstructVLA exact response preservation

Exact preservation alongside a 1.73-unit action-chunk shift.

What the paper found

This paper shows that a reasoning vision-language-action model can keep its explanation word for word while changing its physical action. The authors introduce the text-neutral action-subspace, or TNAS: they identify a locally action-sensitive hidden-state direction, then project out gradients that affect text. That direction overlaps the active text-gradient span by just 1.87% in NVIDIA’s Alpamayo driving model and 5.28% in InstructVLA for manipulation. By aligning projected-gradient-descent image perturbations with TNAS, the attacks steer actions while reducing text drift: teacher-forced drift falls by 68.4% in Alpamayo and 40.5% in InstructVLA compared with text-preserving attacks. On 154 obstacle-annotated driving scenes, full-image attacks cause a new collision with an unchanged rationale in 52.6% of scenes; a localized forward-camera patch does so in 35.1%. At the strongest guidance, Alpamayo’s rationale remains exactly unchanged in 89.5% of examples, while InstructVLA preserves its exact response in 85.0% alongside a 1.73-unit action-chunk shift. A Qwen3-14B rationale-only reader predicts whether Alpamayo stops with 84.3% accuracy, yet the attack results show why readable explanations alone cannot certify safe behavior: monitors must also verify the predicted action.

Original abstract

Reasoning vision-language-action (VLA) models pair physical control with natural-language rationales that supervisors can read to monitor embodied safety. Prior attacks on VLA reasoning corrupt the rationale itself and leave visible evidence in the text. We show that visual perturbations can instead alter the physical action while keeping the generated rationale word-for-word identical. At the shared text/action interface of Alpamayo for driving and InstructVLA for manipulation, our text-neutral action-subspace (TNAS) diagnostic shows that the locally most action-sensitive direction is nearly orthogonal to the active text-loss gradients, with 1.87% and 5.28% norm overlap. At equal hidden-state perturbation norms across both models, interventions along this text-neutral residual shift actions 4.64x to 47.8x more than text-gradient perturbations, whereas text-gradient perturbations induce 21.2x to 277x more text drift. Steering visual perturbations along this direction reaches the same decoupling from the image. Across 154 obstacle-annotated driving scenes in Alpamayo, TNAS-guided attacks induce a new collision under a word-for-word identical rationale in 52.6% of scenes under full-image perturbations and 35.1% under a localized forward-camera patch, compared with 4.55% for unguided action attacks. Relative to a text-preserving baseline, TNAS guidance raises exact rationale preservation from 83.0% to 89.5% and cuts teacher-forced text drift by 68.4% at a comparable collision rate. In InstructVLA, TNAS guidance cuts text drift by 40.5% and preserves the exact response on 85.0% of examples alongside a 1.73-unit action-chunk shift. An unchanged rationale therefore cannot certify a safe action, so monitors for reasoning VLAs must verify the predicted action channel directly.

Read the original paper

More in Embodied AI

Browse all 69 papers →
02Embodied Ai

Kimodo-Policy: Learning Kinematic Action Models for Humanoid Loco-Manipulation

Yunheng Wang, Jiaxi Zhang, Wei Lu, Yuetong Fang, Taowen Wang, Yixiao Feng, Shuning Zhang, Junzhe Xu, Yichi Wang, Zizhao Yuan, Kailun Huang, Lusong Li, Mingqian Zhou, Zecui Zeng, Renjing Xu

Kimodo-Policy turns a pretrained whole-body motion generator into a control policy for humanoid robots that walk and manipulate objects.

Read analysis