IdeaAnchor: Teaching LLMs to Turn Literature into Research Ideas
IdeaAnchor trains models with structured signals about how prior papers combine into a new research idea.
The paper frames literature-grounded ideation as a supervision problem, not just a prompting problem. Its instances encode each input paper’s role, relationships, and synthesis criteria, mined from published papers. The authors train with demonstration, self-distillation, reinforcement learning, and retrieval at inference time. They report consistent gains, with anchor training improving creative synthesis and retrieval adding detail. ArXiv · AI/CL/LG's note
The paper frames literature-grounded ideation as a supervision problem, not just a prompting problem. Its instances encode each input paper’s role, relationships, and synthesis criteria, mined from published papers. The authors train with demonstration, self-distillation, reinforcement learning, and retrieval at inference time. They report consistent gains, with anchor training improving creative synthesis and retrieval adding detail. ArXiv · AI/CL/LG's note
score 5