# Embedding Search Found More Trial Evidence in a 7,354-Trial Test

> A study comparing two search methods found that embedding-based retrieval identified 92% of relevant evidence, versus 45% for label searches.

Oossa · 2026-10-28 · https://oossa.com/en/embedding-search-found-more-trial-evidence-in-a-7-354-trial-test

A new arXiv paper compares label-based searches with a two-stage method that first finds text similar in meaning, then uses a language model to filter results for relevance. In a corpus of 7,354 randomized trials, the researchers say the method found 92% of the evidence a systematic review would include, compared with 45% for searches using curated labels. The paper notes that costs rise with search depth.

## The facts

- Corpus: 7,354 randomized trials.
- Evidence found: 92% with embedding retrieval and language-model filtering, versus 45% with label search.

## Why it matters

For teams preparing systematic reviews, the approach could help surface evidence that does not fit preset labels, though the paper says deeper searches cost more.

## Sources & references

1. [Nothing Left Behind: Finding Evidence from a Large Randomized Evaluation Corpus](https://arxiv.org/abs/2610.36910) – arXiv, 2026-10-28

Last updated: 2026-10-28