Do What You Say: Steering Vision-Language-Action Models Via Runtime Reasoning-Action Alignment Verification
Yilin Wu, Anqi Li, Tucker Hermans, Fabio Ramos, Andrea Bajcsy, Claudia Pérez-D'Arpino
AI summary
Problem
Reasoning Vision-Language-Action models often fail to execute their own intermediate textual plans, creating a reasoning-action faithfulness gap that degrades performance on complex or novel tasks.
Approach
The framework samples multiple candidate action sequences from the model, simulates their outcomes, and uses a pre-trained Vision-Language Model to verify and select the sequence that best matches the model's own textual plan.
Key results
- Up to 15% performance gain on out-of-distribution and compositional tasks
- 8% task success improvement on in-distribution scenarios without retraining
- Preserves long-horizon semantic coherence through runtime verification
- Open-sourced reasoning-annotated LIBERO-100 dataset and extended benchmark
Why it matters
It enables reliable, generalizable robotic control by bridging the gap between high-level reasoning and low-level execution, making reasoning-enabled VLA models more practical for real-world deployment.
Abstract
Reasoning Vision Language Action (VLA) models improve robotic instruction-following by generating step-by- step textual plans before low-level actions, an approach inspired by Chain-of-Thought (CoT) reasoning in language models. Yet even with a correct textual plan, the generated actions can still miss the intended outcomes in the plan, especially in out-of- distribution (OOD) scenarios. We formalize this phenomenon as a lack of embodied CoT faithfulness, and introduce a training- free, runtime policy steering method for reasoning-action align- ment. Given a reasoning VLA’s intermediate textual plan, our framework samples multiple candidate action sequences from the same model, predicts their outcomes via simulation, and uses a pre-trained Vision-Language Model (VLM) to select the sequence whose outcome best aligns with the VLA’s own textual plan. Only executing action sequences that align with the textual reasoning turns our base VLA’s natural action diversity from a source of error into a strength, boosting robustness to semantic and visual OOD perturbations and enabling novel behavior composition without costly re-training. We also contribute a reasoning-annotated extension of LIBERO-100, environment variations tailored for OOD evaluation, and demonstrate up to 15% performance gain over prior work on behavior composi- tion tasks. The overall framework scales with compute (347ms at K = 10 samples) and data diversity. Project Website at: https://yilin-wu98.github.io/steering-reasoning-vla/