AI RESEARCH
How Language Models Process Out-of-Distribution Inputs: A Two-Pathway Framework
arXiv CS.LG
•
ArXi:2605.00269v1 Announce Type: cross Recent white-box OOD detection methods for LLMs -- including CED, RAUQ, and WildGuard confidence scores -- appear effective, but we show they are structurally confounded by sequence length (|r| >= 0.61) and collapse to near-chance under length-matched evaluation. Even raw attention entropy (mean H(alpha) across heads and layers), a natural baseline we include for completeness, shows the same confound. The confound stems from attention's Theta(log T) dependence on input length.