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

This paper from Goodfire and Eternis investigates whether an LLM’s internal activations reveal more about confidence and reasoning than its verbal explanations. Using Eternis-Forecaster 8B, or EF-8B, on OpenForesight, the authors train lightweight mean, attention, and covariance-pooling probes over frozen intermediate representations; a layer-21 covariance probe reduces expected calibration error to 0.044, compared with 0.093 for the model’s verbalized probability. The approach also transfers without updating model weights to Zhipu AI’s GLM-4.7-Flash and GLM-4.5-Air, and supports out-of-distribution correctness monitoring for Qwen3-8B on AIME and AMC mathematics. Faithfulness tests remove real news evidence or inject fabricated articles generated with Claude Sonnet. In evidence ablation, 23% of high-impact perturbations changed the forecast while leaving the chain-of-thought unchanged, and written reasoning correlated with behavioral change at only Spearman rho 0.215. A layer-20 attention probe tracked those shifts at rho 0.565 and predicted their direction correctly in 83.6% of cases, including many silent changes. Finally, forced answering with the larger EF-32B shows that confidence and much of the committed answer are fixed before chain-of-thought begins. Routing questions by the entropy of this pre-reasoning answer distribution saves 30% to 47% of generated tokens with no measurable accuracy loss. The central conclusion is that internal representations provide a practical channel for calibration, auditing, and inference-time triage beyond what the model can faithfully say.