Probing LLM Forecaster Calibration and CoT Faithfulness
July 8, 2026
Research on the Eternis-Forecaster 8B model shows that representation-pooling probes on intermediate activations provide better calibration than direct model outputs. Probes also act as lie detectors, identifying when chain-of-thought reasoning diverges from the actual evidence used for a forecast.
HOW THIS AFFECTS YOU
●
researcherYou can use activation probing to assess the calibration of forecasting models.
●
policyYou can implement more robust auditing for the faithfulness of AI-generated reasoning traces.