Tag
The paper proposes EvoTree, a framework for automatically generating scientific evolution trees from citation graphs, achieving state-of-the-art performance on a new annotated benchmark across AI subfields.
Introduces Deep Research Pretraining (DRP), an offline framework that generates search-open-write trajectories from citation and hyperlink evidence structures. Qwen3-14B models pretrained on 1B tokens with DRP outperform matched no-DRP baselines on deep research benchmarks, even with less supervised fine-tuning data.
Proposes Graphs of Research (GoR), a supervised fine-tuning method that uses citation evolution graphs as supervision for LLM-based research idea generation, achieving state-of-the-art results against gpt-4o-driven baselines.