LLMs show near-zero calibration on hidden information decisions

The study isolates a specific failure: LLMs don't just reason poorly about hidden information, they express unwarranted certainty while doing so. This breaks a core assumption in agentic safety design, where confidence scores are meant to flag when a model should defer or refuse.

This connects directly to the post-training efficiency work covered in 'On-policy Distillation with Verifiable Reward' from late August. That paper focused on integrating task-level correctness signals into LLM training without hyperparameter tuning. The current study suggests a deeper problem: even with better reward signals, models may learn to express high confidence in domains where ground truth is genuinely inaccessible. The calibration failure here isn't a training artifact but a structural feature of how LLMs reason under uncertainty. It also echoes the fairness auditing work from the same period, which found that small input changes can flip verdicts without model behavior actually changing. Here, the model's stated confidence is similarly decoupled from its actual epistemic state.

If researchers can show that fine-tuning on hidden-information tasks (where ground truth is verifiable during training) reduces calibration error below 50 percent on held-out variants, that would suggest the failure is correctable. If calibration remains near 98 percent despite such training, it points to a harder architectural constraint that confidence-gating alone cannot solve.