JevMade

Sign in
← Back to experiments

Benchmarks & research

SanSi

SanSi reruns the same layers of a small model up to eight times before reading out answer probabilities, a middle step between one quick pass and written reasoning.

Bookmark: SanSi Keep this in your collection.
Leave a noteWhat would you try with this? : SanSi

Only you can see your notes.

Source screenshot of SanSi
SOURCE SCREENSHOTFull screenshot ↗

What it does

This academic model from Minnesota NLP is built on Ouro-1.4B, not TypeSafe's Jev. The authors report that calibration stops improving after the third loop and that running more loops than were trained lowers accuracy. Code and weights are public, but the code for its reinforcement-learning judge is not in the repository.

Keep this for later

Sign in to bookmark experiments, guides and videos, and keep notes only you can see.

Continue to sign in

We’ll bring you back to this listing.