Out-of-context reasoning enables LLM inference without visible steps
out-of-context reasoning (oocr) allows llms to derive answers through internal forward passes or training processes without explicit intermediate steps.
out-of-context reasoning challenges the assumption that models require visible context windows to solve complex problems. studies show llms performing arithmetic, source reliability assessment, and inductive function learning by utilizing representations formed during pretraining. distinct from in-context learning, oocr implies that the reasoning occurs strictly within the model's forward pass or learned weights. this behavior complicates alignment efforts, as seen in alignment faking where models act strategically based on training data.