Research map / Reinforcement learning / Offline RL
Katakomba: Tools and Benchmarks for Data-Driven NetHack
Vladislav Kurenkov, Alexander Nikulin, Denis Tarasov, Sergey Kolesnikov
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 |
|---|---|---|---|
| Use HDF5 memory-mapped storage instead of TTYRec | NetHack dataset loading · NetHack data loading, batch_size=256, seq_len=64 | loading time: 255ms -> 18ms | Improved |
| Use HDF5 in-memory storage instead of TTYRec | NetHack dataset loading · NetHack data loading, batch_size=256, seq_len=64 | loading time: 255ms -> 14ms | Improved |
| Use compressed HDF5 with on-read decompression instead of TTYRec | NetHack dataset loading · NetHack data loading, batch_size=256, seq_len=64 | loading time: 255ms -> 2.42s | Hurt |
| Set CQL alpha to 0.0001 rather than 0.0005 (ablation) | CQL · NetHack, final unnormalized game score | Return: 396.50 $\pm$ 39.35 -> 526.72 $\pm$ 71.37 | Improved |
| Set CQL alpha to 0.0001 rather than 0.001 (ablation) | CQL · NetHack, final unnormalized game score | Return: 395.56 $\pm$ 139.92 -> 526.72 $\pm$ 71.37 | Improved |
| Set CQL alpha to 0.0001 rather than 0.05 (ablation) | CQL · NetHack, final unnormalized game score | Return: 226.55 $\pm$ 196.45 -> 526.72 $\pm$ 71.37 | Improved |
| Set CQL alpha to 0.0001 rather than 0.01 (ablation) | CQL · NetHack, final unnormalized game score | Return: 32.42 $\pm$ 43.69 -> 526.72 $\pm$ 71.37 | Improved |
| Set CQL alpha to 0.0001 rather than 0.1 (ablation) | CQL · NetHack, final unnormalized game score | Return: 14.04 $\pm$ 10.29 -> 526.72 $\pm$ 71.37 | Improved |
| Set CQL alpha to 0.0001 rather than 0.5 (ablation) | CQL · NetHack, final unnormalized game score | Return: 0.00 $\pm$ 0.00 -> 526.72 $\pm$ 71.37 | Improved |
| Set CQL alpha to 0.0001 rather than 1.0 (ablation) | CQL · NetHack, final unnormalized game score | Return: 0.59 $\pm$ 0.45 -> 526.72 $\pm$ 71.37 | Improved |
| Use a policy-constraint-free method rather than methods with policy constraints | Recurrent offline RL on NetHack · Katakomba benchmark, normalized score | normalized score | Hurt |
| Use offline RL algorithms instead of Behavioral Cloning | Katakomba benchmark · NetHack, normalized score and death level | normalized score | No clear effect |
Closest papers by meaning
- Simple Ingredients for Offline Reinforcement Learning (ICML 2024 · similarity 0.86)
- A Clean Slate for Offline Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
- OGBench: Benchmarking Offline Goal-Conditioned RL (ICLR 2025 · similarity 0.85)
- Scaling Offline Evaluation of Reinforcement Learning Agents through Abstraction (AAAI 2024 · similarity 0.85)
- Dispelling the Mirage of Progress in Offline MARL through Standardised Baselines and Evaluation (NeurIPS 2024 · similarity 0.84)
- Iteratively Refined Behavior Regularization for Offline Reinforcement Learning (NeurIPS 2024 · similarity 0.84)
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.
