RAG, Fine-Tuning & ContextSam Martin, Fabien Roger
Martin and Roger publish Classifier Context Rot
Published Classifier Context Rot — frontier models missed dangerous behavior that came after 800,000 tokens of harmless activity up to 30× more often
Why it matters
Empirical evidence that long-context limits also break down safety monitoring performance
Source — Primary sources first; secondary reports are labelled “Press”
Paperarxiv.orghttps://arxiv.org/abs/2605.12366Added in Week 35, 2026 · See that week