SIFTing the Evidence

We tend to give large language models (LLMs) a lot of leeway when it comes to getting details right. If we were to get equally inconsistent answers to questions we asked of other people, we would probably write them off as unreliable and stop asking them questions. Maybe we keep going back to LLMs because we are expecting better responses from them as time goes by, but if that does not happen soon, people are likely to give up on them too.

These reliability problems with LLMs are especially pronounced when it comes to niche areas where the models have seen little or no directly relevant training data. That does not stop them from confidently spitting out an answer that is way off base, but researchers at ETH Zurich believe they have a solution that will keep these models on the right track. Their algorithm, called SIFT, introduces specially-selected enrichment data that is tailored to the specific question that was asked. It has been shown that this approach can reduce the model’s uncertainty, which means it is more likely to give accurate answers.

Put it in context

Unlike older methods that rely on nearest-neighbor searches to find relevant information, SIFT employs a deeper understanding of how information is organized within a language model’s vast knowledge space. LLMs internally represent words and concepts as vectors in a high-dimensional space, where the direction and proximity of vectors correspond to semantic relationships. SIFT leverages these relationships by not only finding closely aligned data points but also selecting information that complements the query from different angles, avoiding redundancy and better covering the full scope of the question.

For example, when answering a two-part question like "How old is Michael Jordan and how many children does he have?", conventional nearest-neighbor retrieval methods would often overload the model with multiple redundant facts about his age, neglecting information about his children. SIFT, on the other hand, evaluates the angles between information vectors to prioritize data that fills distinct aspects of the query, ensuring more balanced and complete answers.

Aside from improving accuracy, SIFT is also able to lower the computing power needed for high-quality responses. By continually measuring uncertainty during response generation, the model dynamically determines how much extra data is necessary to improve an answer’s reliability. This adaptive approach allows smaller, more efficient models to match or even outperform larger, more resource-hungry ones. In tests, the team demonstrated that models fine-tuned with SIFT could outperform the best existing AI models using systems up to 40 times smaller.

The future of LLMs

SIFT’s impact could extend beyond just generating accurate answers. By monitoring which enrichment data the system selects in response to different queries, researchers can gain insights into what information truly matters in specific contexts. This could potentially be valuable in fields like medicine, where identifying the most relevant lab results for a diagnosis could improve future patient outcomes.

The researchers have made the SIFT approach accessible through their Active Fine-Tuning (activeft) library, allowing others to adopt it as a drop-in replacement for older retrieval systems. As LLMs continue to become more important tools across a range of industries, methods like SIFT may prove to be essential in ensuring they live up to their promise of not just sounding smart, but actually being smart.