Search-R2: Actor-Refiner Collaboration for Enhanced Search-Integrated Reasoning скачать в хорошем качестве

Search-R2: Actor-Refiner Collaboration for Enhanced Search-Integrated Reasoning

Introducing the Search-R2 framework to innovatively improve the search integration reasoning ability of language models. Existing models have experienced trust allocation problems, which lead to minor errors in the search process leading to the failure of the entire reasoning. To solve this, the researchers propose a collaborative structure between Actor, which creates the initial path, and Meta-Refiner, which accurately locates and fixes the point where the error occurred. In particular, we introduced a 'cut-and-regenerate' mechanism that cuts out and regenerates only the wrong part, and a hybrid compensation model that simultaneously evaluates the correct answer rate of the results and the density of search information. This method has proven to overwhelm existing models in various Q&A benchmarks by reducing unnecessary repetitive searches and increasing logical consistency. As a result, the study provides a new path that allows models to perform complex multi-step reasoning by making more efficient and sophisticated use of external knowledge. https://arxiv.org/pdf/2602.03647