Research map / Vision-language and multimodal / CLIP and zero-shot

PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models

NeurIPS 2026 · Vision-language and multimodal › CLIP and zero-shot · approach: Test-time adaptation of VLMs · 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.

Proposes PriorVLA, which freezes a copy of the pretrained pi0.5 action expert as a read-only Prior Expert and trains a separate Adaptation Expert linked to it via Scene, Motor, and Action Queries to preserve and reuse pretrained priors during VLA fine-tuning.

Problem

Full fine-tuning of pretrained VLA models can shift broad pretrained priors toward narrow downstream training patterns, weakening out-of-distribution generalization and few-shot performance.

Approach

Keep the pretrained action expert frozen as a Prior Expert that is run alongside a trainable Adaptation Expert, and use learnable Expert Queries (Scene, Motor, Action) with masked attention to read scene priors from the VLM and motor priors from the Prior Expert into the Adaptation Expert, which alone produces the action output.

Key results

• On RoboTwin 2.0 standard setting, PriorVLA reaches 77% Easy and 53% Hard average success, +10 and +11 over pi0.5.
• On LIBERO, PriorVLA achieves 99.1% average success across four suites vs. 96.9% for pi0.5 and 97.1% for OpenVLA-OFT.
• With 10 real-world demonstrations per task, PriorVLA reaches 48% ID and 32% OOD success, surpassing pi0.5 by 24 and 22 points.

Surprising findings

• Making the Prior Expert trainable degrades performance relative to keeping it frozen, and a random frozen branch recovers much of the Hard gain (43% vs 49%).

Limitations

• RoboTwin results cover a 13-task subset rather than the full 50-task benchmark, with one model per task.
• OOD real-world factors are applied jointly, so individual factor effects are not disentangled.
• Additional inference computation from executing the frozen Prior Expert during denoising.

Experimental setup

Datasets: RoboTwin 2.0 LIBERO

Benchmarks: RoboTwin 2.0 LIBERO real-world tasks (Franka and AC-One)

Metrics: success rate

Baselines: pi0.5 pi0 Diffusion Policy RDT GR00T-N1.7 OpenVLA-OFT pi0-FAST DreamVLA GR00T-N1 UniVLA F1 DD-VLA

Builds on: pi0.5 pi0 flow matching

Closest papers by meaning

• Align-Then-stEer: Adapting the Vision-Language Action Models through Unified Latent Guidance (ICLR 2026 · similarity 0.93)
• ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
• Beyond Data Scaling: Representation-Centric Pre-training for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
• VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
• FineVLA: Fine-Grained Instruction Alignment for Steerable Vision-Language-Action Policies (NeurIPS 2026 · similarity 0.91)
• RotVLA: Rotational Latent Action for Vision-Language-Action Model (NeurIPS 2026 · similarity 0.91)