Spotting Distillation Traces Through Model Self-Reporting
RyanGreenblatt · x · 2026-07-19
This reply discusses a research/writing outcome regarding distillation, focusing on inferring its presence through model self-reporting.
The author mentions:
- K3 frequently claims to be claude-opus-4-5-20251101, whereas Claude itself does not.
- However, this could also just come from other pre-training data, so it can't be taken as direct evidence of distillation.
- It didn't mention Mythos/Fable, which the author considers a "slightly reverse" piece of evidence, though still very limited.
Overall, this is a research-oriented discussion centered around "how to determine distillation sources from model outputs."
Related event: Kimi K3’s Habit of Calling Itself Claude Draws Scrutiny(3 posts)→
More from Research
- PSNR below 15 dB in papers likely means the metric was misused, researcher warns — kwangmoo_yi · 2026-10-27
- Ortet's first health eval: 34 open-weight models, 50 tasks, generic rankings fall short — thegautamkamath · 2026-10-27
- Scaling-fit discrepancy: output-head FLOPs are ~65% of compute at the smallest rung, ~3% at 535B — dlwh · 2026-10-27
- Marin's dlwh Explains the Scaling Law Fit: Estimating the Loss Floor Early — dlwh · 2026-10-27
- Marin's Open Scaling Law Extrapolates 300× and Predicts Loss Within 0.2% — dlwh · 2026-10-27
- Marin Shares Forecast-vs-Measured Plots: Paloma Hits 0.4% RMS Error — dlwh · 2026-10-27