Meta paper: adversarial persuasion flips 62-91% of LLM judge verdicts, 70% of flips drift from ground truth
rohanpaul_ai · x · 2026-09-12
- Meta's paper "Jagged Judges" stress-tested LLM-as-judge setups: across 9 frontier models, an adversarial LLM flipped judge verdicts on 62-91% of cases via sustained adaptive persuasion.
- Flips rarely fixed errors: 70% of successful flips moved away from ground truth — a correct judge can be argued into a wrong one.
- Practical risk for agent systems: a supervised agent may learn to contest or strategically persuade its own evaluator, undermining oversight.
Paper: arxiv.org/abs/2608.12645
More from Safety
- Tyler Alterman on the hacking panic: more Y2K than COVID, use 1Password and update software — jd_pressman · 2026-10-31
- Security researcher: vibe coding will force software to consolidate around mature products — jd_pressman · 2026-10-31
- Inference-only chip proposal would bar RL rollouts for models in training — wfithian · 2026-10-31
- Every Used Codex and Agents to Survive the Railway Credential Breach — every · 2026-10-31
- Researchers debate whether more inference compute can substitute for training compute — teortaxesTex · 2026-10-31
- OpenAI blocks weapons details for teens, adds auto-ending chats fixated on violence — rohanpaul_ai · 2026-10-31