Reference-based hidden-state auditing detects LLM bias shifts with minimal compute
NASK-PIB · hf · 2026-09-10
A NASK-PIB study proposes a reference-based auditing method that measures representational bias shift across model variants via relative representations of hidden states, detecting internal bias changes with minimal compute.
More from Research
- Why winning training recipes at proxy scale can lose at target scale — gerardsans · 2026-09-10
- Buckmaster Admits No Math Contribution, Calls for Fields Medal to Go to Martínez-Zoroa — stevenstrogatz · 2026-09-10
- Two Years of Math vs 88 Hours of AI: Spanish Mathematicians Say 'Without Our Idea, OpenAI Wouldn't Have Solved It' — stevenstrogatz · 2026-09-10
- 3Blue1Brown's visual explainer of how LLMs work resurfaces — blaizedsouza · 2026-09-10
- AI-Designed Drug Rentosertib Shifts Six Aging Clocks in Phase IIa Trial — AIFlow_ML · 2026-09-10
- Bug Hunt Bench grades frontier models on 105 real bugs; DeepSeek-V4.1-Flash lands 24/105 for $1.80 — PawelHuryn · 2026-09-10