Research map / Vision-language and multimodal / Robot action and navigation

Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Models

NeurIPS 2026 · Vision-language and multimodal › Robot action and navigation · approach: Vision-language-action foundation models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Applies activation injection, per-token sparse autoencoders, and linear probes across six VLA architectures (80M–7B) over 394,000+ rollouts to show the visual pathway dominates action generation, cross-task injection transfers spatially bound motor programs, and language sensitivity depends on task structure.

Problem

Vision-language-action policies are opaque: it is unclear whether they follow language instructions or replay visual-motor priors, and debugging failures is limited to behavioral observation.

Approach

Combines activation injection (null, same-scene, cross-task, cross-seed) with counterfactual prompting, per-token SAEs, and linear probes applied uniformly across six VLA architectures, and links pathway specialization (expert encodes motor programs, VLM encodes goals) to behavioral displacement.

Key results

• π0.5 null-prompt injection of baseline PaliGemma activations recovers cosine similarity 0.999 and 73–77% task success.
• Cross-task injection on X-VLA yields 99.8% source-aligned trajectories while destination success is 0%.
• X-VLA libero_goal success drops from 94% to 10% under wrong prompts while libero_object stays at 60–100%.

Surprising findings

• Cross-task activation injection drives robots toward source-task positions while destination task success collapses to 0% on five models.
• Mean-pooled SAEs achieve better rollout fidelity than per-token SAEs on X-VLA, unlike other architectures.

Limitations

• All experiments are in simulation; real-world transfer and post-fine-tuning effects untested.
• Some concept-ablation tables used full-layer hooks that inflate effect sizes; corrected MLP-only results show no significant effect for 2–5 feature ablations.
• Counterfactual prompt set is limited to simple variations (negation, null, swap).

Experimental setup

Datasets: LIBERO MetaWorld SimplerEnv ALOHA

Benchmarks: LIBERO MetaWorld SimplerEnv ALOHA

Metrics: task success action cosine similarity override rate displacement zero-effect rate explained variance probe R^2

Baselines: OpenVLA-OFT SmolVLA X-VLA GR00T N1.5 ACT

Builds on: activation patching TopK sparse autoencoders activation steering

Closest papers by meaning

• EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.92)
• CogVLA: Cognition-Aligned Vision-Language-Action Models via Instruction-Driven Routing & Sparsification (NeurIPS 2025 · similarity 0.91)
• VLA-OS: Structuring and Dissecting Planning Representations and Paradigms in Vision-Language-Action Models (NeurIPS 2025 · similarity 0.90)
• ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
• FineVLA: Fine-Grained Instruction Alignment for Steerable Vision-Language-Action Policies (NeurIPS 2026 · similarity 0.90)
• Beyond Appearance Shifts: Task-Semantic Action Calibration for VLA Models (NeurIPS 2026 · similarity 0.90)