What LLM Forecasters Know but Don't Say: Probing Internal Representations for Calibration and Faithfulness
The paper says model activations reveal forecast confidence and hidden evidence use better than the model’s own reasoning.
Researchers probed intermediate representations in forecasting models and found better calibration than the stated outputs. In evidence-ablation tests, forecasts often changed while the chain-of-thought did not show that the removed source mattered. The probes tracked those hidden shifts and predicted the direction of change in 84% of cases. A pre-reasoning pass also recovered much of the committed answer and confidence, cutting generated tokens by 30-47% without accuracy loss. HF Daily Papers' note
Researchers probed intermediate representations in forecasting models and found better calibration than the stated outputs. In evidence-ablation tests, forecasts often changed while the chain-of-thought did not show that the removed source mattered. The probes tracked those hidden shifts and predicted the direction of change in 84% of cases. A pre-reasoning pass also recovered much of the committed answer and confidence, cutting generated tokens by 30-47% without accuracy loss. HF Daily Papers' note
score 5