Advanced LLMs show inverse scaling in forecasting, study finds

A new research paper released on arXiv suggests that more capable large language models may perform worse on forecasting tasks involving financial or epidemiological data. The study found that these advanced models tend to produce less accurate distributional forecasts when the underlying data exhibits superlinear growth or a risk of regime change. This inverse scaling effect was observed across simulated and real-world datasets, with the failure concentrated in the upper tail of predictions, where models aggressively extrapolate growth. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Suggests current LLM evaluation metrics may not capture real-world forecasting risks, potentially impacting deployment in critical domains.

RANK_REASON The cluster contains an academic paper detailing new research findings on LLM capabilities. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• COVID-19
• ForecastBench-Sim
• Large Language Models
• Llama-3.1

• paper
• safety