Research map / Vision-language and multimodal / Robot action and navigation
Vision Language Models are In-Context Value Learners
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 Generative Value Learning (GVL), which prompts frozen VLMs to auto-regressively output task-progress values for shuffled video frames, enabling zero-shot and in-context value estimation for robot tasks without training.
Problem
Learning universal temporal value (task progress) functions for robot videos across many tasks and embodiments requires generalization, accurate state estimation, and temporal consistency; naive VLM prompting over ordered frames yields uninformative monotonic values.
Approach
Treat value estimation as autoregressive prediction of completion percentages over randomly shuffled frames (conditioned on the first frame as anchor), so the VLM must reason about each frame's progress rather than exploit temporal ordering; few-shot in-context examples (including cross-embodiment human videos) further improve it.
Key results
- Zero-shot GVL with Gemini-1.5-Pro achieves average VOC 0.74 on RT-1 and positive VOC on most of 50 OXE datasets, outperforming LIV on language-goal datasets.
- On 250 ALOHA tasks, one-shot in-context examples raise positive VOCs to about 90% with median VOC 0.37 (zero-shot: median 0.12).
- On ALOHA sim success detection, GVL-SD (one-shot) reaches 0.75 accuracy versus 0.62 for SuccessVQA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use autoregressive batch prediction instead of single-frame value prediction (ablation) | GVL on RT-1 · RT-1 | VOC: -0.08 -> 0.74 | Helped |
| add shuffled input frames, anchored by the initial frame (ablation) | GVL · ALOHA success and failure trajectories | VOC-based discrimination | Helped |
| add one same-task in-context trajectory | GVL on ALOHA · ALOHA | Median VOC: 0.12 -> 0.37 | Helped |
| increase the number of in-context examples up to 5 | GVL on ALOHA-13 · 13 ALOHA tasks | Average VOC | Helped |
| use human videos as one cross-embodiment in-context example | GVL on ALOHA · ALOHA | VOC | Helped |
| use examples from a different task rather than the target task | GVL on ALOHA-13 · ALOHA-13 | VOC | Helped |
| use GVL-SD instead of SuccessVQA for success filtering | filtered imitation learning · ALOHA simulation tasks | Policy success rate | Helped |
| filter training data with GVL-SD at varying VOC thresholds | ACT · ALOHA simulation tasks | Policy success rate | Helped |
| set the VOC threshold to 0.75 | GVL-SD filtered ACT · ALOHA simulation tasks | Policy success rate | Hurt |
| use GVL one-shot values for advantage weighting | Diffusion Policy · Real-world ALOHA tasks | Success rate: 7/10 vs 6/10; 7/10 vs 5/10; 9/10 vs 6.5/10; 4/10 vs 6/10; 4.67/10 vs 7/10; 1.5/10 vs 0/10; 7/10 vs 7/10 | Mixed |
| use GPT-4o instead of Gemini-1.5-Pro as the backbone (ablation) | GVL on OXE · OXE | VOC | No clear effect |
Surprising findings
- Removing frame shuffling makes VLMs output monotonically increasing values regardless of trajectory quality, inflating VOC and destroying discriminability.
- Pre-trained VLMs prompted with single frames perform poorly (average VOC -0.08 on RT-1 subset).
Limitations
- Value-Order Correlation is most suitable for aperiodic tasks with a unique frame ordering; wiping or stirring is hard to evaluate.
- VLMs were not fine-tuned for value prediction and multi-view observations were not explored.
- Bimanual ALOHA performance (median VOC 0.12 zero-shot) is notably weaker than on OXE datasets.
Experimental setup
Datasets: Open X-Embodiment (50 datasets, 20 sampled trajectories each) ALOHA bimanual real-world tasks (250 tasks) ALOHA simulation tasks (6 tasks, mixed-quality rollouts)
Benchmarks: Open X-Embodiment VOC evaluation ALOHA real-world VOC evaluation ALOHA simulation success detection real-world ALOHA policy learning
Metrics: Value-Order Correlation accuracy precision recall success rate
Baselines: LIV SuccessVQA SuccessVQA-CoT VLM (Single Frame) ACT (no filtering) Diffusion Policy
Builds on: LIV SuccessVQA Action Chunking Transformer Diffusion Policy advantage-weighted regression
Closest papers by meaning
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
- NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
- Exploring the Limits of Vision-Language-Action Manipulation in Cross-task Generalization (NeurIPS 2025 · similarity 0.89)
- World–Value–Action Model: Implicit Planning for Vision–Language–Action Systems (NeurIPS 2026 · similarity 0.89)
- Video Language Planning (ICLR 2024 · similarity 0.89)
- PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
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.