RUUZEN
AI

Sber AI Lab Develops SIRIN Library for AI System Control

Researchers from Sber AI Lab have described SIRIN, a library for detecting hallucinations and other errors in LLM, RAG, and multi-agent systems.

18 Nov, 10:35Read in Russian
habr.com

SIRIN was developed by Maxim Makarenko, Head of the Applied Research Team at Sber AI Lab (Sber), and Senior Researcher Yulia Belikova. The library combines methods for assessing uncertainty, analyzing internal model representations, and verification using judge models.

The system can work with both open and closed generative models and detect errors in entire responses as well as individual statements. It is also proposed for use before generation: in RAG systems, SIRIN checks whether the retrieved documents are sufficient for a response. If the data is insufficient, the search can be repeated or generation can be abandoned.

The authors note that model confidence does not always indicate correctness: an LLM can convincingly formulate a conclusion that is not present in the source data. Training a probe requires a minimum of 300–500 examples, while a fine-tuned judge model requires at least 1,000.

In multi-agent systems, it is necessary to verify not only the final text but also the entire chain of actions: agents' messages, tool calls, their results, and intermediate decisions. An error at an early stage may persist even if subsequent agents correctly processed the incorrect initial premise.