The Unequal Influence of Bad Advice: Using Training Data Attribution to Modulate Emergent Misalignment

The Unequal Influence of Bad Advice: Using Training Data Attribution to Modulate Emergent Misalignment. It centres on Fine-tuning, and also names AI Safety and Benchmarks. Reported by arXiv. Bharat Hunt files it under AI Research and AI Models — the section covering papers, benchmarks, evaluations, interpretability and safety results.

Written by Bharat Hunt from the headline and the coverage below. The original reporting is the source of truth.