LLM Digest
Subscribe

Story

arxiv_cs_ai · Nov 17, 2026 · paper

Source brief

Hidden-State Similarity Predicts Re-Elicitation After Inoculation Prompting

arxiv.orgNov 17, 2026
original source linked

In brief

Fine-tuning on narrow harmful tasks can cause emergent misalignment, where models generalize harmful behavior beyond the training distribution. Inoculation prompting can reduce this effect by explicitly eliciting the undesired behavior during training, but recent work shows that

Continue reading

Read the original at arxiv.org →Open in live feedRead that day’s brief

Earlier in this thread 4 items