More precise AI answers
A new algorithm from ETH Zurich is revolutionizing the use of large language models. With specifically selected data, AI answers become more reliable and the computing effort is significantly reduced.
Large language models fascinate with their knowledge and at the same time irritate with inaccurate or contradictory answers. The reason for this often lies in the uncertainty of the models, which has been difficult to control until now. Researchers at ETH Zurich have now developed the SIFT algorithm (“Selecting Informative data for Fine-Tuning”), a method that addresses precisely this issue. It selects additional data precisely according to whether it reduces uncertainty and improves the quality of the answer.
Information selection through vector analysis
Instead of simply using the closest information, SIFT analyzes the relationship structure of the language information in the multidimensional space of the large models. The decisive factor is the angle between the vectors. Information with complementary content is specifically selected to enable complete and relevant answers. Redundancies and overlaps, as they occur in classic approaches, are systematically avoided.
Big impact even with small models
The targeted enrichment with relevant data makes it possible to drastically reduce the computational effort of large language models. In tests, SIFT-tuning even outperformed powerful AI models with models up to 40 times smaller. At the same time, the system adapts dynamically. During use, the enriched model becomes more and more precise as it continues to train itself during operation.
Evaluating relevance for other areas of application
SIFT also offers a valuable additional benefit. By analysing which data is recognized as particularly relevant, important correlations in specialist areas such as medicine, research or industry can be identified. For example, particularly meaningful diagnostic data can be efficiently filtered out.