LLM Future Prediction via Time-Truncated Data Synthesis

A new paper on arXiv introduces a time-truncation harness method to improve future event prediction using large language models (LLMs). The approach addresses temporal leakage in historical queries by enforcing a temporal cut-off at each step, enabling Tool-Integrated Reasoning (TIR) style sampling from historical events without reliance on rejection sampling or unsolved queries. This increases data synthesis efficiency for forecasting tasks.