ProvenanceGuard checks LLM agent answers for correct sourcing

Hugging Face researchers have introduced ProvenanceGuard, a new system that checks whether large language model (LLM) agents are attributing facts to the right sources, not just whether the facts are correct. Unlike existing tools that mix all evidence together, ProvenanceGuard keeps track of where each piece of information came from and flags answers that cite the wrong source—even if the information itself is accurate somewhere in the evidence.

• ProvenanceGuard works after the LLM produces an answer
• Checks if each claim cites the correct source tool
• Prevents cross-source conflation errors
• Maintains source identity throughout verification
• No need to retrain the underlying LLM agent