Research map / Reinforcement learning / Offline RL
Uni[MASK]: Unified Inference in Sequential Decision Problems
Micah Carroll, Orr Paradise, Jessy Lin, Raluca Georgescu, Mingfei Sun, David Bignell, Stephanie Milani, Katja Hofmann, Matthew Hausknecht, Anca Dragan, Sam Devlin
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 |
|---|---|---|---|
| Fine-tune a random-masking-trained model on the task (ablation) | UniMASK · Maze2D BC, context 5 | reward: 2.19 -> 2.67 | Improved |
| Fine-tune a random-masking-trained model on the task (ablation) | UniMASK · Maze2D RC, context 10 | reward: 2.31 -> 2.61 | Improved |
| Train with random masking instead of single task, no fine-tuning (ablation) | UniMASK · Maze2D BC, context 5 | reward: 2.66 -> 2.19 | Hurt |
| Train on BC and RC jointly instead of single task (ablation) | UniMASK · Maze2D BC, context 5 | reward: 2.66 -> 2.65 | No clear effect |
| Fine-tune multi-task model (ablation) | UniMASK · Maze2D RC, context 10 | reward: 2.39 -> 2.42 | No clear effect |
| Use GPT backbone (Decision-GPT) instead of BERT-like UniMASK single-task (ablation) | UniMASK · Maze2D RC, context 10 | reward: 2.41 -> 2.73 | Improved |
| Use GPT backbone instead of BERT-like at context 5 (ablation) | UniMASK · Maze2D RC, context 5 | reward: 2.64 -> 2.32 | Hurt |
| Increase context length from 5 to 10 for UniMASK single-task (ablation) | UniMASK · Maze2D BC | reward: 2.66 -> 2.47 | Hurt |
| Increase context length from 5 to 10 for Decision-GPT (ablation) | Decision-GPT · Maze2D BC | reward: 2.66 -> 2.74 | Improved |
| Replace Decision Transformer with UniMASK fine-tuned | UniMASK · Maze2D BC, context 5 | reward: 1.13 -> 2.67 | Improved |
| Replace feedforward NN with UniMASK single-task | UniMASK · Maze2D RC, context 5 | reward: 1.53 -> 2.64 | Improved |
| Random-masking training instead of single-task training (ablation) | UniMASK · Minigrid validation loss | validation loss | Mixed |
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 4 of 4 shared results
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
- UnMaskFork: Test-Time Scaling for Masked Diffusion via Deterministic Action Branching (ICML 2026 · similarity 0.88)
- On the Limits of Test-Time Compute: Sequential Reward Filtering for Better Inference (ICML 2026 · similarity 0.87)
- Accelerated Sampling from Masked Diffusion Models via Entropy Bounded Unmasking (NeurIPS 2025 · similarity 0.86)
- Revealing Behavioral Plasticity in Large Language Models: A Token-Conditional Perspective (ICML 2026 · similarity 0.86)
- DynaAct: Large Language Model Reasoning with Dynamic Action Spaces (NeurIPS 2025 · similarity 0.86)
- Generalizing Verifiable Instruction Following (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.
