A new training method teaches large language models to turn a pile of related papers into an actual research proposal, not just a summary.
The paper introduces IdeaAnchor, a system for training LLMs on what the authors call literature-grounded ideation. Rather than relying on prompting tricks or ad hoc feedback, the researchers mined real published papers to extract structured specifications showing how each source paper functioned when a human researcher turned it into a new idea, including its role, its relationship to other sources, and how it fed into the final synthesis. Models are then trained on those mined examples through demonstration, self-distillation, and reinforcement learning, with a retrieval step added at inference time to pull in supporting details. The authors report consistent improvements in ideation quality over existing approaches.
Most AI-for-research tools today amount to summarizing a stack of abstracts and gesturing at a gap. This work tries to teach the mechanics of synthesis itself, treating idea generation as a function of specific inputs rather than a vibe. The authors' own breakdown is the most useful finding: anchor-based training handles the creative leap between papers, retrieval fills in supporting detail, and the two together outperform either alone.
It's a lab paper, not a product, but it points at where AI research assistants are headed: from summarizing what's already written to proposing what should be written next.