[amflare]

> Existing mitigation approaches often degrade performance on open-ended generation and downstream tasks, limiting their practical utility. [...] Unlike continuous reward schemes, our approach assigns a reward of one only when the model's output is entirely factually correct, and zero otherwise.

Someone correct me if I am wrong, as I'm am on the very edge of this space looking in, but does this mean that they are using a "degraded performance with fewer hallucinations" model to fact check the "more powerful yet prone to hallucinations" model?

[mNovak]

My understanding is no, they are collecting a cache of documents from the training set, then after pre-training prompt about those topics. A separate verifier is given both the relevant source documents and generated response, and tasked with checking for conflicts in factuality.

They describe using Qwen 32B as the verifier, and the model under training is Qwen 8B. So in fact the verifier is beefier than the trainee model, though it's unclear if that has to be the case as you scale up.

[svnt]

Also on the edge, but it appears they are relying on the search-augmented identification of conflicts in the generated statement, which is an easier task than constructing an answer to the question. It also encourages abstention because there are no conflicts in “I don’t know” (so “mitigating hallucinations” and “answering more questions correctly” are not necessarily the same thing)