Extended Data Table 2 Ablations of training and inference of OpenScholar

From: Synthesizing scientific literature with retrieval-augmented language models

  1. Quantitative ablations of OpenScholar components on Scholar-CS. The table reports rubric accuracy and citation F1 for OpenScholar-8B and OpenScholar-GPT-4o, along with variants that remove training, reranking, self-feedback or attribution, as well as retrieval-only variants using OSDS only, Semantic Scholar only or web-only search. Removing reranking or attribution leads to the largest decreases in citation F1, web-only retrieval performs worst overall and combining dense retrieval, Semantic Scholar and web sources yields the strongest factuality and citation support.
Back to article page