2026-04-06
- Coconut oracles
- Training LRM
- For all latent reasoning, GPT-2-large works much better than -small, 2.4% -> 45%
- Trained LRM: https://huggingface.co/syvb/gpt-2-latent-reasoning
- some other research on coconut interp
- model they used: https://huggingface.co/bcywinski/codi_llama1b-answer_only
- Training LRM