How Language Models Process Out-of-Distribution Inputs: A Two-Pathway Framework
For researchers and practitioners needing reliable OOD detection in LLMs, this work identifies a critical confound and offers a principled framework to disentangle genuine signals from length artifacts.
The paper reveals that existing OOD detection methods for LLMs are confounded by sequence length, collapsing to near-chance under length-matched evaluation. It proposes a two-pathway framework where embeddings capture topic shifts and processing trajectory detects covert-intent inputs, achieving 0.721 avg AUROC across tasks and 0.850 for jailbreak detection.
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. To identify genuine OOD signals after deconfounding, we propose a two-pathway framework: embeddings capture what text is about (effective for topic shifts), while the processing trajectory -- hidden-state evolution across layers -- captures how the model processes input. The relative power of each pathway varies along a vocabulary-transparency spectrum: embedding methods excel on vocabulary-distinctive OOD, while trajectory features detect covert-intent inputs that share vocabulary with normal text (0.721 avg AUROC; Jailbreak: 0.850). Three evidence lines support this framework: (1) a crossover between k-NN and trajectory scoring across 6 tasks, where each pathway wins on different OOD types; (2) a per-layer analysis showing that layer-0 k-NN signal is almost entirely a length artifact (Jailbreak: 0.759 raw -> 0.389 matched) -- processing constructs genuine OOD signal from near-chance embeddings; and (3) circuit attribution showing adversarial tasks engage attention circuits more than semantic tasks (p = 0.022; Jailbreak patching p < 0.001), with partial cross-model replication. Code release upon publication.