Skip to content

Motif

Train NetHack agents from LLM preferences

Visit official site
Category AI Agents

What Motif is

The project is described as a PyTorch system learning reward functions from LLM preferences and training agents on NetHack.

What you can do with it

  • Build annotated pairs from LLM preferences
  • Convert preferences into a reward function
  • Train an agent with reinforcement learning

Pros & cons

✓ What we like

  • Behaviours are described as consistent with human intuition
  • Reinforcement learning trains the resulting agent

! What to watch out for

  • The named game environment is NetHack
  • Reward functions depend on the collected LLM preferences

FAQ

What is the base framework?

Motif is a PyTorch-based project.

How are rewards built?

LLM preferences are turned into a reward function with cross entropy.

Last reviewed: 2026-11-15

More AI office assistant tools

View all →

How we review