Reinforcement learning raised Qwen3.5-9B’s score on a test of whether language models can build research ideas from earlier papers, the authors report. On 924 papers accepted to ICLR 2026, the model met 54.2% of the benchmark’s expert-refined criteria, versus 31.1% for its base version. The score measures agreement with a rubric, not whether a proposed study works.
IdeaAnchor teaches a model to connect the roles earlier papers play in a later study. Its guides mark a paper as a direct predecessor, an inspiration, a source of a research gap, or a method ingredient; they also spell out how the works relate and what a plausible new idea should address. When generating an idea, the model receives the papers without the guide.
The authors reconstructed these guides from 14,183 machine-learning and natural-science papers published from 2023 to 2025. They trained models in three ways: copying teacher-written examples, learning from their own outputs produced with a guide, or reinforcement learning against the guide’s criteria. For Qwen3-8B, reinforcement learning raised the score to 24.6%, from 10.9% for the base model. Qwen3.5-9B rose from 31.1% to 54.2%; supervised fine-tuning reached 48.7%.
The evaluation used 924 ICLR 2026 papers. Two expert annotators checked each paper’s proposed prior-work set, assigned roles and refined its criteria. A GPT-5.4 judge then scored whether generated ideas met those criteria while remaining concrete, literature-grounded, sound and feasible.
The study does not test whether generated proposals succeed in experiments. Its main benchmark also covers machine-learning papers from a single conference year, so the results do not establish performance across other research fields or settings.
Sources
- Ziyu Chen et al., “IdeaAnchor: Teaching LLMs to Turn Literature into Research Ideas” (arXiv:2610.08781v1)
- Full text of the study
Figure: IdeaAnchor authors, Figure 1, used unchanged under CC BY 4.0.