AI learns to forecast events by searching and gathering evidence

Researchers have developed a novel approach to training AI models for event forecasting by integrating a learning-to-search mechanism. This method allows the AI agent to actively gather evidence through web searches and data analysis before making a prediction, with the skill of evidence gathering being shaped by the reward signal. The trained model, based on Qwen3.5-35B-A3B, demonstrated improved calibration and reduced search attempts, outperforming frontier models like Claude Opus 4.5 in forecasting accuracy on complex questions at a fraction of the inference cost. The researchers are releasing the environment, dataset, and tools to facilitate further research in temporal forecasting agents. AI

IMPACT This research demonstrates a novel method for improving AI forecasting capabilities by teaching models to actively seek and evaluate information, potentially enhancing their reliability in complex decision-making scenarios.

RANK_REASON The cluster contains a research paper detailing a new method for training AI models, including a dataset and environment release. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• CatalyzeX
• Claude Opus 4.5
• DagsHub
• Gotit.pub
• Hugging Face
• IArxiv
• Influence Flower
• Polymarket
• Qwen3.5-35B-A3B
• ScienceCast

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →