Context-Anchored Generation (CAG): Fixing Hallucinations at the Decoding Layer

Context-Anchored Generation (CAG) addresses LLM hallucination not as a knowledge gap but as a control problem at the decoding layer. It maintains a semantic anchor from the prompt and tracks token-level drift using cosine similarity, intervening during generation via two modes—constraint and expansion—rather than filtering outputs post-hoc. CAG requires no retraining, is model-agnostic, adds negligible overhead, and also mitigates repetition loops, long-context degradation, and topic switching.