Researchers identify "Forking": train loss falls while validation loss turns up
ZimingLiu11 · x · 2026-10-26
Ziming Liu and MetaCircleAI report a newly named phenomenon called "Forking," caused by mis-generalizing a trick found by AutoResearch. Forking appears in over-encoding methods (akin to DeepSeek v4.1's Engram memory trick): under data replay, train loss keeps falling while validation loss turns upward, with the two curves splitting like a fork. The authors argue Forking deserves attention on the same level as grokking and double descent.
More from Research
- DONIC hits 86.79% on bimanual dexterous tracking, crushing prior SOTA's 9.43% — liuziwei7 · 2026-10-27
- Why don't more LLMs ship 10M-token context? A researcher asks about attention tradeoffs — agarwl_ · 2026-10-27
- 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