Research map / Vision-language and multimodal / Robot action and navigation
ALAM: Algebraically Consistent Latent Transitions for Vision-Language-Action Models
Zuojin Tang, Haoyun Liu, Xinyuan Chang, Changjie Wu, Dongjie Huo, Yandan Yang, Bin Liu, Zhejia Cai, Feng Xiong, Mu Xu, jiachen luo, De Ma and 2 more
Read the paperCheck your idea against 137,217 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 ALAM, a latent action model for VLA pretraining from action-free videos whose latent transitions are algebraically consistent, so the latent codes better support policy generation.
Problem
VLA models are limited by scarce action-labeled robot data, while action-free videos are abundant; reconstruction-trained latent action codes may predict future observations but lack the structure needed for policy generation.
Approach
Enforces algebraic consistency on latent action transitions learned from action-free videos, so latent codes compose structurally rather than only supporting future-observation reconstruction.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add ALAM latent transitions co-generated with actions | pi0 · MetaWorld MT50 | avg success rate (%): 47.9 -> 85.0 | Improved |
| add ALAM latent transitions co-generated with actions | pi0 · LIBERO | avg success rate (%): 94.2 -> 98.1 | Improved |
| add ALAM versus JALA on LIBERO-Long | pi0 · LIBERO-Long | success rate (%): 96.0 -> 94.4 | Hurt |
| add algebraic composition and reversal losses to latent pretraining (ablation) | ALAM encoder · MetaWorld MT50 | avg success rate (%): 70.7 -> 85.0 | Improved |
| add reversal loss alone (ablation) | ALAM encoder · MetaWorld MT50 | avg success rate (%): 71.6 -> 85.0 | Improved |
| add additivity loss alone (ablation) | ALAM encoder · MetaWorld MT50 | avg success rate (%): 74.1 -> 85.0 | Improved |
| keep VQ bottleneck at pretraining (ablation) | ALAM encoder · MetaWorld MT50 | avg success rate (%): 70.8 -> 85.0 | Improved |
| replace unstructured LAM with ALAM latents under joint flow matching (ablation) | pi0 + latent stream · MetaWorld MT50 | avg success rate (%): 65.0 -> 85.0 | Improved |
| replace unstructured LAM with ALAM latents under joint diffusion (ablation) | pi0 + latent stream · MetaWorld MT50 | avg success rate (%): 38.3 -> 65.6 | Improved |
| add unstructured LAM latents with joint diffusion (ablation) | pi0 · MetaWorld MT50 | avg success rate (%): 47.9 -> 38.3 | Hurt |
| switch from two-stage to joint flow matching (ablation) | ALAM · MetaWorld MT50 | avg success rate (%): 68.9 -> 85.0 | Improved |
| freeze, block or shuffle the latent stream at test time (ablation) | pi0 + ALAM · MetaWorld MT50 | avg success rate (%): 85.0 -> 75.3-76.4 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Evo-1 — Evo-1: Lightweight Vision-Language-Action Model with Preserved Semantic Alignment (CVPR 2026) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Experimental setup
Builds on: latent action models
Closest papers by meaning
- Learning a Unified Latent Action Space from Videos with Action-centric Cycle Consistency (CVPR 2026 · similarity 0.94)
- RotVLA: Rotational Latent Action for Vision-Language-Action Model (NeurIPS 2026 · similarity 0.92)
- A Latent World-Action Model with Jointly Aligned Reasoning (NeurIPS 2026 · similarity 0.92)
- Latent Action Pretraining from Videos (ICLR 2025 · similarity 0.91)
- Beyond World-Frame Action Heads: Motion-Centric Action Frames for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
- DiT4DiT: Jointly Modeling Video Dynamics and Actions for Generalizable Robot Control (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.
