Chain-of-Verification (CoVe)

The agent first produces an initial, unverified draft response to the user's query using its standard generation capabilities. This baseline serves as the raw material to be scrutinized, not the final output. The model is prompted to generate freely, capturing all potentially relevant facts, entities, and claims that will later be subjected to rigorous verification.

The agent analyzes its own baseline response and systematically decomposes it into a set of atomic, independent fact-checking questions. Each question targets a single verifiable claim. This factoring step is critical because it prevents the model from simply re-reading its own output and reinforcing confabulations. Instead, it forces a fresh, targeted retrieval for each discrete fact.

Each factored question is answered independently, crucially without providing the original baseline context. This isolation prevents the model from being biased by its initial phrasing. The agent answers each question based solely on its parametric knowledge or retrieved context for that specific query. This step acts as a cross-examination, where each factual pillar is tested in isolation.

The agent compares the answers from the independent verification step against the claims in the original baseline response. Inconsistencies are identified and flagged. The model then generates a final, corrected output that synthesizes the verified facts and explicitly omits or rectifies any claims that were contradicted or unsupported during the verification execution phase.

By decoupling the generation of claims from their verification, CoVe directly addresses the core mechanism of hallucination where a model's autoregressive nature causes it to commit to and elaborate on an initial error. The technique has been shown to significantly reduce factual error rates in tasks like biography generation and long-form question answering, often matching or exceeding the performance of more computationally expensive multi-model debate approaches.