Teknium: Hermes memory research shows ~15% better generalization than finetuning
Teknium · x · 2026-11-12
Teknium responded to Unify's research on Hermes's memory mechanism, arguing the results show Hermes both memorizes (and uses) memories and generalizes — around 15% above the 28% finetuning baseline. He also proposed partnering to run the same test harness with Opus 5.5, to see whether a stronger model can leverage the harness to learn better than Luna, and said he'll study the paper closely.
Related event: Teknium: Hermes Shows Both Memory and Generalization, ~15% Over Fine-tuning(2 posts)→
More from Models
- LLMs play Among Us against each other; Haiku 5.5 kills then instantly self-reports — Rasmic · 2026-11-12
- Kimi K3.1-Pro abliterated variant hits Hugging Face, claiming 95% refusal removal — zebird0 · 2026-11-12
- Possible eval leak: all frontier models fail a GPQA Diamond question whose official answer seems wrong — qberthet · 2026-11-12
- Engineers Using LLMs Should Understand Tokens, Context Windows and Hallucinations — _jaydeepkarale · 2026-11-12
- Math Benchmark Sheet Shows 90.43% Top Accuracy at Just $1.21 per Run — teortaxesTex · 2026-11-12
- Open model Inkling 2 Small takes top open spot on MathArena's BrokenArXiv — teortaxesTex · 2026-11-12