Success across six settings
Loading chart…
World Action Model
Propose an action. Rethink its consequences. Refine before execution.
A future prediction that is not conditioned on the chunk is identical for every proposal and thus carries no proposal-specific signal, limiting its usefulness for refinement before execution.
ThinkWAM conditions a learned residual on the complete action proposal to correct multi-horizon future latents. The same action head uses these corrected futures in a second pass, supported by affordance-guided memory and near-fine, far-coarse prediction. All prediction stays in latent space.
A second action pass, guided by corrected latent futures.
Use observation–action memory to propose the next action chunk.
Correct near, middle, and far latent futures using the proposed action.
Generate the executable action chunk from the corrected futures.
ThinkWAM uses a proposed action to revise its latent future predictions, then conditions the second action pass on those revised futures. The images used to illustrate latent horizons are conceptual visualizations.
ThinkWAM against strong baselines across simulation and real-world settings.
Loading chart…
Loading chart…
Success rate (%) per suite · bold = best, underline = second best.
3 OOD conditions · 30 trials per method and condition
Galaxea A1Z · NVIDIA RTX 4080 SUPER
Green-object tidying · Galaxea A1Z
What the latent correction changes, and where it helps.
LIBERO and LIBERO-Plus · success rate (%)
† Both action passes are retained; the second pass uses uncorrected futures.
@misc{tang2026thinkwam,
title = {{ThinkWAM}: Rethinking Latent Futures for Action Refinement},
author = {Tang, Haoyun and Deng, Baqiao and Zhou, Enjin and Tian, Songwen
and Xu, Keyao and Jing, Bowen and Wang, Kun and Ma, Ke
and Mei, Zhandong and Li, Huibin},
year = {2026},
eprint = {2611.21258},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2611.21258}
}