lf.wtf
Levi Foster

AI Research

Lab write-ups from my work on editing language models directly: reading what a small model already holds, then grafting in parts that use it. Results are on problems that were sealed before any building started.

Write-ups

  1. Circuit grafts

    Small rule circuits took SmolLM2-135M from 21.3% to 83.1% on held-out multiplication. Five math skills went up by 37 to 62 points, with no original weight changed and no gradient training.

    SmolLM2-135M Read the write-up