Research map / Multi-agent systems and games / Human-AI and goal-conditioned
NetHack is Hard to Hack
Ulyana Piterbarg, Lerrel Pinto, Rob Fergus
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 hierarchical strategy labels and bilevel decoder (ablation) | LSTM policy trained with BC · NLE | mean NLE score: 658 -> 931 | Improved |
| add hierarchical strategy labels (ablation) | LSTM policy trained with BC · NLE | median NLE score: 403 -> 614 | Improved |
| enlarge the baseline decoder without hierarchy (ablation) | LSTM policy trained with BC · NLE | mean NLE score | No clear effect |
| replace LSTM with transformer-LSTM (ablation) | BC policy without hierarchy · NLE | mean NLE score: 658 -> 1318 | Improved |
| add hierarchy to transformer-LSTM (ablation) | BC policy · NLE | mean NLE score: 1318 -> 1151 | Hurt |
| add APPO RL fine-tuning to BC (ablation) | LSTM baseline · NLE | mean NLE score: 658 -> 1204 | Improved |
| add hierarchy (ablation) | APPO + BC LSTM policy · NLE | mean NLE score: 1204 -> 1551 | Improved |
| add hierarchy (ablation) | APPO + BC transformer-LSTM · NLE | mean NLE score: 1326 -> 1346 | No clear effect |
| add APPO fine-tuning to BC (ablation) | transformer-LSTM · NLE | mean NLE score: 1318 -> 1326 | Mixed |
| ablate LSTM encoder to give a flat transformer (ablation) | transformer-LSTM · NLE | mean NLE score | Hurt |
| scale transformer-LSTM from 3 layers (50.8M) to 6 layers (98.9M) (ablation) | transformer-LSTM · NLE | mean NLE score | Hurt |
| scale training data from 10 to 100,000 games (ablation) | LSTM and 6-layer transformer-LSTM with BC · NLE | mean NLE score | Mixed |
Closest papers by meaning
- The NetHack Learning Environment (NeurIPS 2020 · similarity 0.89)
- BlendRL: A Framework for Merging Symbolic and Neural Policy Learning (ICLR 2025 · similarity 0.86)
- Mimicking To Dominate: Imitation Learning Strategies for Success in Multiagent Games (NeurIPS 2024 · similarity 0.84)
- Reason to Play: Behavioral and Brain Alignment Between Frontier LRMs and Human Game Learners (NeurIPS 2026 · similarity 0.84)
- The Algorithm Is Not the Behavior: Learned Priors Override Look-Ahead in a Chess-Playing Neural Network (NeurIPS 2026 · similarity 0.84)
- Neurosymbolic Reinforcement Learning: Playing MiniHack with Probabilistic Logic Shields (AAAI 2025 · 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.
