# LLMs Can Teach Themselves to Better Predict the Future

**Authors:** Benjamin Turtel, Danny Franklin, Philipp Schoenegger  
**arXiv:** [2502.05253](https://arxiv.org/abs/2502.05253)  
**Date:** 2025-02-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2502.05253v1/

## Key points

- Authors introduce an outcome-driven fine-tuning framework for enhancing forecasting capabilities of large language models (LLMs)
- Approach leverages model self-play to generate diverse reasoning trajectories and probabilistic forecasts for a wide range of questions
- Ranking reasoning traces based on proximity to actual outcomes before fine-tuning using Direct Preference Optimization (DPO) is key innovation
- Method significantly improves prediction accuracy of Phi-4 14B and DeepSeek-R1 14B by 7--10% compared to base models
- Enhancement brings models on par with larger frontier models like GPT-4o in forecasting capabilities
- Research showcases novel approach allowing LLMs to autonomously improve predictive abilities without extensive human intervention

## Summary

In their paper titled "LLMs Can Teach Themselves to Better Predict the Future," authors Benjamin Turtel, Danny Franklin, and Philipp Schoenegger introduce an innovative outcome-driven fine-tuning framework aimed at enhancing the forecasting capabilities of large language models (LLMs). This approach differs from traditional methods that rely on human-curated reasoning samples and instead leverages model self-play to generate diverse reasoning trajectories and probabilistic forecasts for a wide range of questions. These questions are resolved after the models' knowledge cutoff date. The key innovation lies in ranking these reasoning traces based on their proximity to actual outcomes before fine-tuning the model using Direct Preference Optimization (DPO). The results are impressive, with their method significantly improving the prediction accuracy of Phi-4 14B and DeepSeek-R1 14B by 7--10% compared to base models. This enhancement brings these models on par with much larger frontier models like GPT-4o in terms of forecasting capabilities. Overall, this research showcases a novel approach that allows LLMs to autonomously improve their predictive abilities without extensive human intervention. By harnessing self-play and sophisticated optimization techniques, the proposed framework represents a significant step forward in advancing the capabilities of language models for future prediction tasks.