A Pre-Generation Epistemic Routing Axis in Language Models
Abstract
AbstractWhen answering factual questions, language models may respond correctly, hallucinate, or abstain. We investigate whether this epistemic disposition is encoded in hidden states before any answer token is generated. Using a balanced Wikipedia QA corpus with pre/post article-date splits and rollout-consensus labels, we analyze prompt-only MLP activations from four instruction-tuned 7–9B models. PCA reveals a dominant pre-generation epistemic routing axis that orders outcomes as ABSTAIN, HALLUCINATION, and CORRECT. We then apply activation steering as a bounded causal stress test. Dense residual-write edits can shift many hallucinations toward abstention or correctness, while sparse MLP-intermediate edits localize the steering signal. These results frame the axis as a steering audit: it identifies where fixed edits improve factual behavior and where correct loss or broken outputs call for conditional controllers.