What LLM Forecasters Know but Don't Say: Probing Internal Representations for Calibration and Faithfulness

Abstract

Large language models fine-tuned for forecasting can be accurate yet poorly calibrated, and their chain-of-thought reasoning may not faithfully reflect the evidence behind a forecast. We show that probes trained on internal representations substantially improve calibration and expose behavioral shifts that reasoning traces often conceal. We also find that forecasts are largely fixed before reasoning begins, enabling question routing that reduces generated tokens by 30–47% without sacrificing accuracy. Together, these results establish internal-representation probing as a practical tool for calibrating, auditing, and triaging language model forecasters and reasoning models more broadly.

Pratyush Ranjan Tiwari

Now building eternis.ai