Ada-Diffuser: Latent-Aware Diffusion Models for Decision-Making (ICLR 2026)

Forum topic · 小凯 · 2026-05-18

Summary

A Chinese tech forum post discusses Ada-Diffuser, a paper by Feng, Ge, Fu, Li, Zheng, Tang, Hu, Huang, and Zhang accepted at ICLR 2026. The method extends diffusion models from image generation to decision-making in robot control, addressing the problem that environment dynamics are often only partially observable—hidden factors like wind direction, fluid viscosity, or opponent strategies shape action outcomes. The paper's theoretical contribution shows that, under mild conditions, the hidden latent process can be identified from short observation windows of recent observation-action-reward triples, without full history. Ada-Diffuser is built as a causal diffusion model that jointly learns the temporal structure of visible interactions and the underlying latent dynamics, using inferred latent states to condition planning and policy learning. On simulated control and robotic benchmarks, it outperforms baselines that ignore latent states, both in latent inference accuracy and adaptive policy learning. The author raises open questions: whether the mild identifiability conditions hold in real robotics, the required window length versus how fast hidden factors change, and whether added latent inference can meet real-time constraints.

Diffusion models can now do more than generate images—they are being applied to decision-making in robot control, where a diffusion process denoises step by step to produce a full action plan conditioned on the current state. But there is a catch: environment dynamics are often only partially observable. You cannot see the wind direction, fluid viscosity, or an opponent's strategy—yet these hidden factors determine what your actions will produce.

At ICLR 2026, Feng, Ge, Fu, Li, Zheng, Tang, Hu, Huang, and Zhang propose Ada-Diffuser. The core idea: if a diffusion model makes decisions without inferring hidden latent variables, it is only fitting surface appearances—actions that look correct can go off track in the actual environment.

Theory

Under mild conditions, the hidden latent process can be identified from short time-window observations. There is no need to inspect the full history—just the most recent observation-action-reward triples suffice to infer the latent state currently driving the environment.

Method

Ada-Diffuser is built as a causal diffusion model. During training it learns two things simultaneously:

the temporal structure of observed interactions (the visible part), and
the underlying latent dynamics (the invisible part).

During planning or policy learning, it uses the inferred latent states to adjust decisions—generating actions based not on surface phenomena, but on the factors driving them.

Results

On simulated control and robotic benchmarks, Ada-Diffuser outperforms baselines that ignore latent states, in both latent-variable inference accuracy and adaptive policy learning.

Open questions (from the forum author)

Are the "mild conditions" for identifiability commonly satisfied in real robot scenarios?
What is the concrete length of the short time window? In changing real environments, hidden factors may shift faster than the window length.
Computational cost: diffusion models are already slow—can the added latent inference still meet real-time requirements for robot control?

References

1. Feng, F., Ge, S., Fu, M., et al. (2026). *Ada-Diffuser: Latent-Aware Adaptive Diffusion for Decision-Making*. arXiv:2605.16054 [cs.LG]. 2. Janner, M., Du, Y., Tenenbaum, J. B., & Levine, S. (2022). *Planning with Diffusion for Flexible Behavior Synthesis*. ICML. 3. Ajay, A., et al. (2023). *Compositional Foundation Models for Hierarchical Planning*. NeurIPS.