Ask or search anything...
Outcome-based Reinforcement Learning to Predict the Future
Stanford UniversityResearchers successfully adapted Reinforcement Learning with Verifiable Rewards (RLVR) to predict real-world future events, training a 14-billion-parameter model to achieve superior probabilistic calibration and predictive accuracy compared to a frontier LLM (OpenAI o1). The fine-tuned models demonstrated a 10% hypothetical return on investment in simulated prediction market trading, particularly excelling where market consensus was uncertain.
View blog
LLMs Can Teach Themselves to Better Predict the Future
07 Feb 2025
We present an outcome-driven fine-tuning framework that enhances the
forecasting capabilities of large language models (LLMs) without relying on
human-curated reasoning samples. Our method leverages model self-play to
generate pairs of diverse reasoning trajectories and probabilistic forecasts
for a set of diverse questions that resolve after the models' knowledge cutoff
date. We then rank pairs of these reasoning traces by their distance to the
actual outcomes before fine-tuning the model via Direct Preference Optimization
(DPO). On a separate test set, our approach increases prediction accuracy of
Phi-4 14B and DeepSeek-R1 14B by between 7--10\% over a base model and a DPO
fine-tuned control model with randomized labels, bringing them on par with
forecasting capabilities of much larger frontier models like GPT-4o.
There are no more papers matching your filters at the moment.