Research map / Reinforcement learning / Visual and representation RL
Decision Stacks: Flexible Reinforcement Learning via Modular Generative Models
Siyan Zhao, Aditya Grover
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 changed | Where | Measured | Effect |
|---|---|---|---|
| Add reward modeling as input to the diffusion action model (ablation) | Decision Stacks (diffusion obs) · Hopper-medium-v2 POMDP | normalized score: 43.7 -> 58.2 | Improved |
| Add reward modeling as input to the transformer action model (ablation) | Decision Stacks (diffusion obs) · Hopper-medium-v2 POMDP | normalized score: 43.6 -> 57.7 | Improved |
| Add reward modeling as input to the MLP action model (ablation) | Decision Stacks (diffusion obs) · Hopper-medium-v2 POMDP | normalized score: 38.1 -> 45.6 | Improved |
| Add reward modeling to transformer action model with transformer observation model (ablation) | Decision Stacks (transformer obs) · Hopper-medium-v2 POMDP | normalized score: 45.1 -> 58.0 | Improved |
| Replace MLP action model with diffusion action model (ablation) | Decision Stacks (diffusion obs) · Hopper-medium-v2 POMDP | normalized score: 45.6 -> 58.2 | Improved |
| Use hybrid diffusion obs / transformer reward / diffusion action instead of pure transformer (ablation) | Decision Stacks · Hopper-medium-v2 POMDP | normalized score: 53.0 -> 58.2 | Improved |
| Use Decision Stacks modular action model instead of Decision Diffuser MLP inverse dynamics | Decision Diffuser · Maze2D Single Goal large | normalized score: 111.8 -> 171.6 | Improved |
| Use Decision Stacks instead of Diffuser without handcoded controller | Diffuser · Maze2D Multi Goal average | normalized score: 70.0 -> 123.4 | Improved |
| Use Decision Stacks instead of Decision Diffuser on Maze2D single goal umaze | Decision Diffuser · Maze2D Single Goal umaze | normalized score: 113.8 -> 111.3 | No clear effect |
| Use Decision Stacks instead of Decision Diffuser in MDP locomotion | Decision Diffuser · D4RL locomotion MDP average | normalized score: 82.2 -> 81.1 | No clear effect |
| Use Decision Stacks instead of Diffuser in POMDP locomotion | Diffuser · D4RL locomotion POMDP average | normalized score: 64.2 -> 74.3 | Improved |
| Use Decision Stacks instead of Diffuser in MDP locomotion | Diffuser · D4RL locomotion MDP average | normalized score: 75.3 -> 81.1 | Improved |
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.
- Decision Transformer — Decision Transformer: Reinforcement Learning via Sequence Modeling (NeurIPS 2021) · 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.
Closest papers by meaning
- Offline Reinforcement Learning via High-Fidelity Generative Behavior Modeling (ICLR 2023 · similarity 0.89)
- Constrained Latent Action Policies for Model-Based Offline Reinforcement Learning (NeurIPS 2024 · similarity 0.86)
- Diffusion Subgoal Planning for Long-Horizon Offline Goal-Conditioned Reinforcement Learning (NeurIPS 2026 · similarity 0.86)
- Consistency Models as a Rich and Efficient Policy Class for Reinforcement Learning (ICLR 2024 · similarity 0.86)
- Compositional Conservatism: A Transductive Approach in Offline Reinforcement Learning (ICLR 2024 · similarity 0.86)
- Think-RM: Enabling Long-Horizon Reasoning in Generative Reward Models (NeurIPS 2025 · similarity 0.86)
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.
