This work concentrates on the initial step of the research ideation pipeline—identifying actionable limitations from the literature. A pilot study shows that as large language models (LLMs) scale from 0.6B to 32B parameters, their ability to recognize which limitation prior work addressed improves dramatically, while their capability to generate new limitations does not, indicating that the bottleneck lies in the supervision signal rather than model capacity. Consequently, the authors build ActLim-Bench by extracting forward citations, yielding 18,279 evidence‑linked supervision triples verified by humans. On this signal they propose a three‑stage post‑training pipeline: (1) supervised fine‑tuning; (2) verifier‑filtered rejection‑sampling fine‑tuning; (3) reinforcement learning with the verifier as an online dense reward. Experiments demonstrate that the method attains the highest recall among frontier models and is preferred in 64.9% of non‑tied human comparisons, also improving downstream ideation systems. The dataset and code will be released upon publication.
Review