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

This paper studies whether hidden representations reveal more about LLM forecasting than chain-of-thought. Using Eternis-Forecaster 8B on OpenForesight, plus GLM-4.7-Flash and GLM-4.5-Air, the authors train probes over intermediate activations and report improved calibration. Evidence ablation and diversionary prompt injections often change forecasts without changing the visible reasoning trace. Probe activations track these behavioral shifts better than CoT and predict the direction of change in 84% of cases. Forced-answer experiments suggest forecasts and confidence are largely committed before reasoning, while routing based on the pre-reasoning answer distribution saves 30-47% of generated tokens without reported accuracy loss.