Actionable research limitations: forward-citation supervision for automated ideation
In brief
- LLMs from 0.6B to 32B parameters better recognize limitations addressed in later work but fail to generate new ones.
- ActLim-Bench provides 18,279 human-verified supervision triples based on forward citations.
- The three-stage pipeline boosts recall and achieves 64.9% preference in human comparisons for ideation.
Research shows LLMs struggle to generate new, actionable limitations despite identifying previously addressed ones. The authors introduce ActLim-Bench, a dataset of 18,279 forward-citation-based triples, and a three-stage post-training pipeline to improve this critical step in automated research ideation.
This summary gives you the essentials. The full article, its figures and quotes are on cs.AI updates on arXiv.org.
Read the original articleAI-generated summary of the source, checked by rules. Report an error