Sber and Skoltech Teach AI to Recognize Its Own Hallucinations
Researchers at Sber’s Center for Practical Artificial Intelligence and Skoltech have developed the TOHA method, which detects unreliable answers from large language models that work with external sources.
The idea is to look for errors not in the answer itself, but in the model’s “inner workings” – the topological structure of its attention maps. This approach does not require training separate classifiers and uses only a minimal amount of labeled data. It is also more computationally efficient than competing methods such as SelfCheckGPT, which requires multiple generations of an answer to verify it.
The method was tested on question-answering and text-summarization tasks. Across several test sets, TOHA delivered results on par with current state-of-the-art methods. The technology has already been integrated into the open-source SIRIN library, making it available for enterprise assistants that work with knowledge bases. The researchers presented the technology at the ACL 2026 conference.
