Ethanfel d29808975f Use sqrt(cfg) scaling for K + store raw prev_eps ...

Full K corrupts at high CFG (output correction = cfg*K = 2.4 at cfg=12).
K/cfg was too weak (0.2 at cfg=12). The paper only tested up to cfg=7.5
where output corrections range 0.5-1.5.

K/sqrt(cfg) keeps output correction = sqrt(cfg)*K growing sub-linearly,
giving 0.69 at cfg=12 — within the paper's working range.

Also store raw (pre-correction) guidance as prev_eps to prevent
correction accumulation through the sliding surface.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

2026-03-04 23:39:16 +01:00

4.2 KiB

Raw Permalink Blame History

View Raw