Research map / Vision-language and multimodal / Robot action and navigation
Actions as Language: Fine-Tuning VLMs into VLAs Without Catastrophic Forgetting
Asher James Hancock, Xindi Wu, Lihan Zha, Olga Russakovsky, Anirudha Majumdar
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| represent actions as natural language and fine-tune with LoRA only | VLM2VLA (Gemma-3-12B-IT) · 12 VQA benchmarks | VQA score retention: MME 1182.3 -> 1391.7; MMMU 46.0 -> 42.7 | Improved |
| fine-tune VLM into VLA with tokenized actions and full fine-tuning (OpenVLA/ECoT) | Prismatic 7B · VQA benchmarks | VQA score: MMStar 38.8 -> 0 (OpenVLA) | Hurt |
| replace language actions with least-likely action tokens (VLM2VLA-AT) (ablation) | VLM2VLA · VQA benchmarks | VQA score: ChartQA 58.3 -> 41.8; MMMU 42.7 -> 45.9 | No clear effect |
| represent actions as language instead of least-likely tokens (ablation) | VLM2VLA · Real robot OOD, Pick Up the Item Above Ash Ketchum | success rate (%): 30% -> 60% | Improved |
| represent actions as language instead of least-likely tokens (ablation) | VLM2VLA · Real robot, simple ID tasks | success rate | No clear effect |
| represent actions as language instead of least-likely tokens (ablation) | VLM2VLA · Real robot multilingual and compositional tasks | success rate | Improved |
| co-train with robot and web data (pi0.5) | pi0.5 · VQA benchmarks | VQA score: RealWorldQA 2.7 | Hurt |
| use OpenVLA trained on Open-X-Embodiment versus VLM2VLA | OpenVLA · Real robot ID pick-up and pick-and-place | success rate | Improved |
Closest papers by meaning
- FineVLA: Fine-Grained Instruction Alignment for Steerable Vision-Language-Action Policies (NeurIPS 2026 · similarity 0.93)
- Knowledge Insulating Vision-Language-Action Models: Train Fast, Run Fast, Generalize Better (NeurIPS 2025 · similarity 0.92)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
- LLaRA: Supercharging Robot Learning Data for Vision-Language Policy (ICLR 2025 · similarity 0.92)
- Beyond Data Scaling: Representation-Centric Pre-training for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
- NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
