AI Paper Atlas

Research map / Image generation / Autoregressive and latent generation

MSP: Probabilistically Consistent Multi-Scale Action Generation

Zhixuan Lin, Gengqi Liu, Chao Zheng, Gao Lin, Jindong Yu, Song Gao, Fei Wang

ICML 2026 · Image generation › Autoregressive and latent generation · approach: Next-scale and token autoregressive images

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 changedWhereMeasuredEffect
keep all intermediate scales [1,2,4,8] instead of dropping scale 2 (ablation)MSP · LIBERO-Longsuccess rate (%): 91.6 -> 96.8Improved
use dense scales instead of sparse scales [1,4,16] with long action length (ablation)MSP · LIBERO-Longsuccess rate (%): 79.4 -> 90.2Improved
use downsampling factor 4 instead of 2 (ablation)MSP · LIBERO-Longsuccess rate (%): 89.6 -> 96.8Improved
use downsampling factor 4 instead of 8 (ablation)MSP · LIBERO-Longsuccess rate (%): 93.0 -> 96.8Improved
use action length 32 instead of 16 (ablation)MSP · LIBERO-Longsuccess rate (%): 90.8 -> 96.8Improved
use action length 32 instead of 64 with extra scale (ablation)MSP · LIBERO-Longsuccess rate (%): 90.2 -> 96.8Improved
use MeanFlow instead of diffusion or standard flow matching for scale-wise generation (ablation)MSP · LIBERO-Longsuccess rate (%)Improved
map actions to continuous VAE latent space rather than raw action space (ablation)MSP · LIBERO-Longsuccess rate (%)Improved
add scale-wise autoregressive modeling (ablation)MSP · LIBERO-Longsuccess rate (%)Improved
use MSP instead of QueST for few-shot adaptationMSP · LIBERO-Long few-shot (5 demos)success rate (%): 66.7 -> 76.4Improved
use MSP in place of DSP on LIBERO-GoalMSP · LIBERO-Goalsuccess rate (%): 90.8 -> 93.8No clear effect
use MSP instead of DSP in terms of parametersMSP · LIBERO-Longaction-head parameters: about 60% of DSP parametersImproved

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.

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.

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy