Inference-Time Context Augmentation Can Act as Undirected Perturbation
October 23, 2026
An evaluation of five context augmentation methods across seven LLM sizes reveals that accuracy gains are often inconsistent. The study finds that inserting filler tokens or elaborations can act as undirected perturbations, correcting some predictions while simultaneously breaking others.
HOW THIS AFFECTS YOU
●
builderBe cautious when using automated context augmentation to reduce hallucinations in production workflows.
●
researcherYou should account for instability when evaluating context-driven reasoning improvements.