🧠 How to keep language models honest?

Large language models are great at almost everything, except one thing: they excel at fabricating convincing nonsense, even when connected to knowledge bases via RAG. Traditional methods for tackling hallucinations usually demand either tons of labeled data or heavy computations involving repeated answer generation.

Researchers from Skoltech and Sber's Center for Applied AI have found a much more elegant solution — a new method called TOHA (TOpology-based HAllucination detector).

🔹 Instead of putting the model through heavy extra workloads, the approach looks "under the hood" to analyze the topological structure of its attention maps using the MTop-Div metric. If the response graph diverges from the context, the algorithm instantly flags the discrepancy.
🔹 The method requires no training of additional classifiers and uses minimal labeled data.
🔹 In terms of accuracy, it easily competes with resource-heavy alternatives like SelfCheckGPT.

The tool is already integrated into Sber's open-source SIRIN library, ready to protect corporate knowledge bases from AI fabrications.

The research has been accepted to the main track of the A* ACL 2026 conference. Dive into the details and author comments in our full release.