New Paper Jagged Judges Explores LLM Confidence and Epistemic Stability
ShirleyYXWu · x · 2026-08-22
The paper "Jagged Judges: Epistemic Stability Under Silence, Pressure, and Persistence" investigates the stability of LLMs used as judges when facing repeated questioning. It argues that accuracy alone doesn't measure the judge's conviction and proposes methods to evaluate an LLM's true confidence and decision stability under silence, pressure, and persistence.
More from Safety
- Anthropic launches enterprise code security scans on Mythos 5 — AccBalanced · 2026-08-22
- Article: Why America's Data Center Build-out is Facing Resistance — robleclerc · 2026-08-22
- US ban on foreign robots strands startups, requires 65% US parts — davidad · 2026-08-22
- Can agentic coding enable private-by-default software that works with sensitive data? — every · 2026-08-22
- California's $12B science research bond clears committee, Assembly vote due by Aug 31 — anne_churchland · 2026-08-22
- OpenAI Backs Preventative Control Laws, but Industry Practices Lag Behind — sjgadler · 2026-08-22