Loading activity…
@vaidehi_patil_
PhD student @unccs advised by @mohitban47 | @Google PhD Fellow | Undergrad @IITBombay | Prev. Intern @GoogleDeepMind @Apple @AmazonScience @AdobeResearch
AI
Its creators say benign fine-tuning can trigger harmful agent actions, including leaking private information through email tools.
AI
The researchers say LADE uses first-token probabilities to detect harmful prompts without accessing a model’s internal hidden states.
AI
A researcher says her October 6 poster covers belief steering via theory of mind, with a privacy-focused workshop talk planned for October 9.
AI
The authors introducing EvoHarnessBench say even today’s self-evolving methods cannot reliably adapt to new capabilities while retaining earlier competence.
