Research map / Reinforcement learning / Offline RL
Is Conditional Generative Modeling all you need for Decision Making?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use return-conditioned Decision Diffuser with classifier-free guidance | D4RL offline RL · D4RL average across reported locomotion tasks | normalized average return: 78.9 -> 81.8 | Helped |
| Use Decision Diffuser for long-horizon kitchen tasks | D4RL Kitchen · Kitchen Mixed and Partial | normalized average return: 52.4 -> 65; 50.1 -> 57 | Helped |
| Add classifier-free guidance to the Diffuser setup (ablation) | Hopper tasks · Hopper Med-Expert, Medium, and Med-Replay | normalized average return: 107.6 -> 111.3; 58.5 -> 66.3; 96.8 -> 76.5 | Mixed |
| Use inverse dynamics to infer actions instead of diffusing actions jointly (ablation) | Hopper diffusion policies · Hopper Med-Expert, Medium, and Med-Replay | normalized average return: 111.3 -> 111.8; 66.3 -> 79.3; 76.5 -> 100 | Helped |
| Diffuse the current action conditioned on current state and target return (ablation) | Diffusion policy · Hopper tasks | normalized average return: 105.6; 54.1; 66.5 | Hurt |
| Generate trajectories satisfying combined constraints at test time | Decision Diffuser · Kuka Block Stacking and Rearrangement | success rate: 60.3; 67.2 | Helped |
| Lower sampling temperature to α=0.5 rather than α=0 (ablation) | Decision Diffuser · Hopper-Medium-Expert | normalized average return: 104.3 ± 0.7 -> 111.8 ± 1.6 | Helped |
| Lower sampling temperature to α=0.5 rather than α=1.0 (ablation) | Decision Diffuser · Hopper-Medium-Expert | normalized average return: 107.1 ± 3.5 -> 111.8 ± 1.6 | Helped |
| Remove sampling entropy by setting α=0 (ablation) | Decision Diffuser · Hopper-Medium-Expert | normalized average return: 111.8 ± 1.6 -> 104.3 ± 0.7 | Hurt |
| Use inverse dynamics rather than action diffusion (ablation) | Decision Diffuser · Block Push Position Control | success rate: 87.3 ± 3.1 -> 87.8 ± 2.8 | No clear effect |
| Use inverse dynamics rather than action diffusion (ablation) | Decision Diffuser · Block Push Torque Control | success rate: 71.8 ± 3.4 -> 84.7 ± 2.2 | Helped |
| Increase stochasticity in environment dynamics (ablation) | Decision Diffuser · Block Push Torque Control, p=0 to p=0.15 | success rate: 84.7 ± 2.2 -> 41.3 ± 4.9 | Hurt |
Closest papers by meaning
- Generative Decision Making Under Uncertainty (AAAI 2023 · similarity 0.91)
- On the Modeling Capabilities of Large Language Models for Sequential Decision Making (ICLR 2025 · similarity 0.86)
- Rethinking Decision Transformer via Hierarchical Reinforcement Learning (ICML 2024 · similarity 0.86)
- Constrained Latent Action Policies for Model-Based Offline Reinforcement Learning (NeurIPS 2024 · similarity 0.86)
- DeFlow: Decoupling Behavior-Prior Modeling and Value Maximization for Offline Policy Extraction (NeurIPS 2026 · 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.