Reproduce “FAME: Forecasting Academic Impact via Continuous-Time Manifold Evolution” on OpenResearch

About this paper

Jianrong Ding, Jianyuan Zhong, Zhengyan Shi, Qiang Xu

FAME introduces a continuous-time manifold evolution framework to forecast the future impact of scientific papers by modeling topic dynamics as trajectories in a latent space. The model achieves average Spearman rank correlations over 0.50, outperforming current LLMs and machine learning baselines, and boosts augmented LLM evaluators by more than +0.150 in correlation.

Key insights

• Static text-based evaluation by LLMs is inadequate for prospective scientific impact forecasting due to its inability to capture dynamic spatiotemporal field evolution.
• Modeling local impact relative to a spatiotemporal neighborhood (Vanguard Loss) is crucial for geometrically differentiating vanguard papers from ordinary contributions within their evolving topic.
• FAME's dynamic latent space resolves the "geometric bottleneck" of static text embeddings, structurally disentangling impactful research by aligning it with the continuous evolutionary trajectories of scientific topics.

Summary by alphaXiv.