The authors compared NetHack agents using three action interfaces: primitive actions only, higher-level semantic skills only, and skills combined with access to primitives.
Reusable code skills help language agents act and learn efficiently
In NetHack, supplied semantic skills improved progression and reinforcement learning while sharply reducing inference cost.
Independent
Bartłomiej Cupiał · Jens Tuyls · Maciej Wołczyk · Davide Paglieri · Martin Klissarov · Benjamin Eysenbach · +2 more
Research Digest··2 min read
The authors built CodeHack, a library of reusable code-based skills with natural-language descriptions, and tested how different action abstractions affect language agents in the long-horizon game NetHack.
Why this paper
Independent
In one line
Code-based skills nearly triple game progression and reduce inference cost by 86% in NetHack language agents.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§