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

Forum topic · 小凯 · 2026-05-18

Summary

Ada-Diffuser, presented at ICLR 2026 by Feng, Ge, Fu, and colleagues, extends diffusion models from image generation to sequential decision-making in partially observable environments. While diffusion-based planners can generate action sequences by iterative denoising, they typically ignore hidden environmental factors—such as fluid dynamics or opponent strategies—that determine the consequences of actions. The authors' theoretical contribution shows that, under mild conditions, the hidden latent process can be identified from short observation windows of recent observation-action-reward triplets, without requiring full history. Building on this identifiability result, Ada-Diffuser is formulated as a causal diffusion model that jointly learns the temporal structure of visible interactions and the underlying latent dynamics during training. At planning or policy-learning time, the inferred latent states condition action generation, so decisions reflect the drivers behind observations rather than surface patterns alone. On simulated control and robotics benchmarks, Ada-Diffuser outperforms baselines that disregard hidden states, both in latent-variable inference accuracy and adaptive policy learning. Open questions include whether the mild identifiability conditions hold in real robot settings, how short the observation window can be when hidden factors change rapidly, and whether the added latent inference can meet real-time control constraints given diffusion models' computational cost.

Diffusion models can now do more than generate images—they can make decisions. In robot control settings, a diffusion process observes the current state, progressively denoises, and outputs a complete action plan. 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.

The Work

Feng, Ge, Fu, Li, Zheng, Tang, Hu, Huang, and Zhang propose Ada-Diffuser at ICLR 2026. The core insight: if your diffusion model does not infer hidden latent variables during decision-making, it is merely fitting surface appearances—actions that look correct will go off track in the actual environment.

Theoretical Contribution

Under mild conditions, the hidden latent process can be identified from short observation windows. You do not need the full history—just the most recent few observation-action-reward triplets suffice to infer the latent state currently driving the environment.

Method

Based on this theory, Ada-Diffuser is built as a causal diffusion model. During training it jointly learns:

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

At planning or policy-learning time, it uses the inferred latent states to adjust decisions—generating actions not from surface appearances, but from the driving factors behind them.

Results

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

Open Questions

The paper claims identifiability under "mild conditions"—are these conditions generally satisfied in real robot scenarios?
What is the concrete length of the short time window? In changing real environments, hidden factors may evolve faster than the window.
Computational cost: diffusion models are already slow—can latent inference still meet the real-time demands of 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.