Is Capability a Liability? More Capable Language Models Make Worse Forecasts When It Matters Most

1. 1Analyze the performance of LLMs on forecasting tasks with superlinear growth and tail risk of regime change
2. 2Evaluate the distributional forecasts produced by more capable models on these tasks
3. 3Use the ForecastBench-Sim benchmark to test the models' performance on synthetic SIR epidemics
4. 4Compare the results with a matched linear control to identify the inverse scaling pattern
5. 5Investigate the causes of this pattern and its implications for real-world forecasting problems

Data scientists and AI engineers working on forecasting problems in finance and epidemiology can benefit from understanding this phenomenon to improve their models' performance, and researchers can use this knowledge to develop more accurate forecasting models

💡 Inverse scaling in LLMs can lead to worse distributional forecasts in tasks with superlinear growth and tail risk of regime change

💡 More capable LLMs can make worse forecasts in critical situations, highlighting the need for careful evaluation and improvement of forecasting models #LLMs #forecasting