← Publications / Safety Post-Training & Alignment

Drift, Then Repair: A Safety Audit of Fine-Tuned Language Models

Pratinav Seth, Anshul Kaushal, Saisab Sadhu, Vinay Kumar Sankarapu

NeurIPS 2026 Workshop2026My role: Lead authorWorkshop
Evaluation pipeline: aligned base model, benign LoRA fine-tuning, drifted checkpoint, then harden, recover or unlearn repairs and safety-capability evaluation

Cite

@inproceedings{seth2026drift,
  title         = {{Drift, Then Repair: A Safety Audit of Fine-Tuned Language Models}},
  author        = {Pratinav Seth and Anshul Kaushal and Saisab Sadhu and Vinay Kumar Sankarapu},
  booktitle     = {NeurIPS 2026 Workshop},
  year          = {2026}
}
All →