Reinforcement Learning for LLM-based Event Forecasting

A new study on arXiv demonstrates that reinforcement learning can sharpen the event-forecasting ability of large language models, enabling a relatively small model to outperform a much larger commercial system on the same prediction tasks [1]. The paper, submitted June 14, 2026, applies Group Relative Policy Optimization (GRPO) to fine-tune pretrained large language models (LLMs) ranging from 1.5 billion to 14 billion parameters [1]. GRPO is described as a sample- and memory-efficient reinforcement learning method [2]. The models were equipped with tools to access current information beyond their original training data, including a Wikipedia revisions tool and news summaries, to forecast real-world events [2]. The researchers used the results to examine how LLM forecasting scales and to place judgmental forecasting within a taxonomy of verifiable and unverifiable domains, accounting for inherent aleatoric uncertainty in future events [2]. A key finding was that GRPO training elevated a 1.5-billion-parameter transformer, Qwen 2.5 1.5B, to a forecasting performance superior to that of Claude Sonnet 3.5 on the same dataset, as measured by cross entropy against market-agreed probabilities [2]. The paper also documents several unsuccessful approaches explored before arriving at this result [2]. The work sits within a broader machine-learning field that has seen rapid advances since the introduction of the transformer architecture in 2017 and the subsequent AI boom of the 2020s [5]. Forecasting real-world events presents a challenge distinct from many benchmark tasks because it involves irreducible uncertainty, such as the outcome of a die roll [2]. The study's approach of connecting models to live information sources also touches on the persistent problem of AI hallucination, where models generate false or misleading information presented as fact [4]. The research appears on arXiv, a preprint repository that has integrated with platforms like Hugging Face Spaces to allow authors and the community to attach interactive demos directly to paper pages, making results more accessible without requiring users to write code [8][9]. The Qwen model family used in the study originates from the same ecosystem that produced models like DeepSeek, a Chinese AI company that gained attention for training competitive open-weight models at a fraction of the cost of larger rivals [11].