Sequential fine-tuning leaves models with just 1.2% accuracy on old tasks after 100 updates
burkov · x · 2026-09-10
A study on long-horizon memorization shows sequential supervised fine-tuning overwrites earlier associations, leaving naive models with only 1.2% average accuracy on prior tasks after 100 updates.
Key points:
- Three 100-task datasets were built to isolate the setting: arbitrary symbol associations, LLM-generated fictional facts, and filtered real-world questions
- A search over 90 candidate mechanism compositions (using successive halving) tested whether combining mechanisms that protect different aspects of prior knowledge beats any single mechanism
The problem matters for deployments where models serve as persistent internal memory rather than relying on repeated external retrieval.
More from Research
- DONIC hits 86.79% on bimanual dexterous tracking, crushing prior SOTA's 9.43% — liuziwei7 · 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
- Marin's Open Scaling Law Extrapolates 300× and Predicts Loss Within 0.2% — dlwh · 2026-10-27