AI Reviewers Fooled by Rewrites? 1,200 Papers Tested Reveal SciCore Dual-Branch Fix

Researchers from Virginia Tech, UMD, and MBZUAI rewrote 60 ICLR 2026 papers 20 ways each using GPT-5.5 and Claude Opus 4.8, then evaluated 30 AI reviewer configurations on 1,200 rewrites, finding that prompt-based robustness often yields 'false robustness' with near-random discrimination, while their SciCore dual-branch method—combining full-paper review with a structured Science Core extraction—reduces score drift by 38% while preserving discrimination.

AI peer reviewClaude Opus 4.8GPT-5.5

0 likes · 15 min read

AI Reviewers Fooled by Rewrites? 1,200 Papers Tested Reveal SciCore Dual-Branch Fix

Data Party THU

Oct 27, 2025 · Artificial Intelligence

Why Most LLM Defense Strategies Fail Against Adaptive Attacks

An extensive study reveals that twelve recent large‑language‑model defenses, including prompt‑based, adversarial‑training, filtering, and secret‑knowledge methods, are easily bypassed by a general adaptive attack framework using gradient descent, reinforcement learning, search, and human red‑team techniques, exposing critical robustness gaps.

LLM securityadaptive attacksjailbreak

0 likes · 11 min read

Why Most LLM Defense Strategies Fail Against Adaptive Attacks

robustness evaluation

AI Reviewers Fooled by Rewrites? 1,200 Papers Tested Reveal SciCore Dual-Branch Fix

Why Most LLM Defense Strategies Fail Against Adaptive Attacks