Research map / Vision-language and multimodal / Robot action and navigation
From Pixels to Tokens: A Systematic Study of Latent Action Supervision for Vision-Language-Action Models
Yihan Lin, Haoyang Li, Yang Li, Haitao Shen, Yihan Zhao, Chao Shao, Jing Zhang
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 |
|---|---|---|---|
| add image-based latent action supervision (LA-Direct) | unified VLA baseline (Qwen3-VL-2B) · LIBERO-Long | success rate (%): 85.8 -> 96.6 | Improved |
| add action-based latent action tokens (LA-Tok) | unified VLA baseline · RoboTwin 2.0 (4-task subset) | avg success rate (%): 60.5 -> 78.0 | Improved |
| use image-based instead of action-based latent actions for long-horizon tasks | VLA baseline · LIBERO-Long | success rate (%): +8.4% to +10.8% vs +6.8% (LA-Tok) | Improved |
| explicit direct decoding (LA-Direct) instead of explicit conditional decoding (LA-Cond) (ablation) | image-based latent actions · LIBERO-Long | success rate (%): 94.2 -> 96.6 | Improved |
| use LA-Cond instead of LA-Direct (ablation) | image-based latent actions · RoboTwin 2.0 (4-task subset) | avg success rate (%): 71.8 -> 73.8 | Improved |
| explicit supervision (LA-Direct) instead of implicit alignment (LA-Align) (ablation) | image-based latent actions · LIBERO-Long | success rate (%): 94.8 -> 96.6 | Improved |
| discrete token supervision instead of continuous regression (ablation) | LA-Direct · LIBERO (4-suite average) | success rate (%): 94.4 -> 97.1 | Improved |
| discrete token supervision instead of continuous regression (ablation) | LA-Tok · LIBERO (4-suite average) | success rate (%): 93.3 -> 95.5 | Improved |
| align at early VLM layers instead of layer 17 (ablation) | LA-Align · LIBERO (4-suite average) | success rate (%): 95.0/97.0 -> 91.6 | Hurt |
| align at final layer instead of mid-to-late layer (ablation) | LA-Align · LIBERO (4-suite average) | success rate (%): 97.0 -> 95.0 | Hurt |
| lengthen action placeholders to match latent variants (ablation) | unified VLA baseline · LIBERO | success rate (%): +0.9% (PH-Cond), -1.2% (PH-Direct) | No clear effect |
| add latent action supervision (LA-Cond) under joint multi-task training | unified VLA baseline · RoboTwin 2.0, 10 tasks joint training | avg success rate (%): 53.60 -> 74.50 | Improved |
Closest papers by meaning
- Latent Action Pretraining from Videos (ICLR 2025 · similarity 0.93)
- RotVLA: Rotational Latent Action for Vision-Language-Action Model (NeurIPS 2026 · similarity 0.92)
- Not All Features Are Created Equal: A Mechanistic Study of 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)
- ATI-VLA: Action-Centric Predictive Vision–Language–Action Models via Actionable Alignment Then Adaptive Injection (NeurIPS 2026 · similarity 0.90)
- NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
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.
