Research map / Video / Action recognition
LARY: A Latent Action Representation Yielding Benchmark
Dujun Nie, Fengjiao Chen, Jun Kuang, Qi Lv, Xiaoyu Li, Xuezhi Cao
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 |
|---|---|---|---|
| use general latent-feature LAM (LAPA-DINOv2) instead of embodied UniVLA | LAPA framework · LARYBench classification | Avg accuracy: 18.82 -> 57.41 | Improved |
| use latent-based V-JEPA 2.1 instead of pixel-based Wan2.2 | encoder · LARYBench classification | Avg accuracy: 66.58 -> 76.29 | Improved |
| replace SigLIP2 or MAGVIT2 encoder with DINOv3 in LAPA | LAPA General LAM · LARYBench classification | Avg accuracy: 42.09 -> 62.73 | Improved |
| use DINOv3 encoder directly instead of LAPA-DINOv3 | General LAM · LARYBench absolute regression CALVIN | normalized MSE: 0.50 -> 0.18 | Improved |
| use best General LAM (LAPA-DINOv3) in place of V-JEPA 2.1 for delta regression | delta regression · LARYBench delta regression | Avg normalized MSE: 0.50 -> 0.57 | No clear effect |
| increase codebook size 8 to 64 (ablation) | LAPA-DINOv3 · Composite Human | accuracy: 71.31 -> 70.15 | No clear effect |
| increase codebook size 64 to 256 (ablation) | LAPA-DINOv3 · Composite Human | codebook utilization: 100.0 -> 89.5 | Hurt |
| increase sequence length 16 to 49 (ablation) | LAPA-DINOv3 · codebook utilization | utilization %: 1.6 -> 100.0 | Improved |
| increase sequence length 49 to 64 (ablation) | LAPA-DINOv3 · AgiBotWorld-Beta regression | MSE: 0.83 -> 0.72 | Mixed |
| increase latent dimension 32 to 256 (ablation) | LAPA-DINOv3 · Composite Human | accuracy: 60.94 -> 70.15 | Improved |
| increase latent dimension 256 to 1024 (ablation) | LAPA-DINOv3 · Composite Human | accuracy: 70.15 -> 72.55 | Mixed |
| increase sampling stride 5 to 30 | FLUX.2-dev · VLABench regression | MSE: 0.04 -> 0.62 | Hurt |
Closest papers by meaning
- Learning a Unified Latent Action Space from Videos with Action-centric Cycle Consistency (CVPR 2026 · similarity 0.91)
- From Pixels to Tokens: A Systematic Study of Latent Action Supervision for Vision-Language-Action Models (ICML 2026 · similarity 0.91)
- RotVLA: Rotational Latent Action for Vision-Language-Action Model (NeurIPS 2026 · similarity 0.91)
- Latent Action Pretraining from Videos (ICLR 2025 · similarity 0.91)
- DiT4DiT: Jointly Modeling Video Dynamics and Actions for Generalizable Robot Control (NeurIPS 2026 · similarity 0.90)
- Action Images: End-to-End Policy Learning via Multiview Video Generation (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.
