Research map / Reinforcement learning / Visual and representation RL
A Mixture Of Surprises for Unsupervised Reinforcement Learning
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 separate skill distributions for surprise maximization and minimization | CIC-style unsupervised pretraining · URLB, 12 downstream tasks | IQM: 6% | Helped |
| Use separate maximization and minimization skill distributions | CIC · URLB, 12 downstream tasks | Optimality Gap: 20% | Helped |
| Use MOSS instead of NegativeCIC | NegativeCIC · ViZDoom Defend the Line | score: 462±11 -> 510±22 | Helped |
| Use MOSS instead of NegativeCIC | NegativeCIC · ViZDoom Take Cover | score: 51793±2936 -> 63497±2384 | Helped |
| Add stochasticity by injecting Gaussian action noise with Bernoulli probability 0.3 (ablation) | MOSS · Walker | performance loss | No clear effect |
| Add stochasticity by injecting Gaussian action noise with Bernoulli probability 0.3 (ablation) | CIC · Walker | performance | Hurt |
| Use the same skill distribution for both objectives instead of disjoint skill distributions (ablation) | MOSS · Walker | performance | Hurt |
| Replace the fixed objective schedule with adaptive objective switching (ablation) | MOSS · Quadruped Jump | score: 674±11 -> 687±36 | Helped |
| Replace the fixed objective schedule with adaptive objective switching (ablation) | MOSS · Quadruped Run | score: 485±6 -> 512±20 | Helped |
| Replace the fixed objective schedule with adaptive objective switching (ablation) | MOSS · Quadruped Stand | score: 911±11 -> 869±26 | Hurt |
| Replace the fixed objective schedule with adaptive objective switching (ablation) | MOSS · Quadruped Walk | score: 635±36 -> 758±37 | Helped |
| Increase maximization share from 50% to 60% (ablation) | MOSS · Walker Run | score: 531±20 -> 599±11 | Helped |
Closest papers by meaning
- Towards Principled Unsupervised Multi-Agent Reinforcement Learning (NeurIPS 2025 · similarity 0.91)
- Unsupervised Reinforcement Learning in Multiple Environments (AAAI 2022 · similarity 0.89)
- Wasserstein Unsupervised Reinforcement Learning (AAAI 2022 · similarity 0.89)
- Constrained Ensemble Exploration for Unsupervised Skill Discovery (ICML 2024 · similarity 0.87)
- SEMDICE: Off-policy State Entropy Maximization via Stationary Distribution Correction Estimation (ICLR 2025 · similarity 0.87)
- Convergence Theorems for Entropy-Regularized and Distributional Reinforcement Learning (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.