Causal Analysis of Epistemic Suppression: Can Explicit Uncertainty Training Reverse Reasoning Degradation?

TL;DR: What if we explicitly train models to identify and verbalize uncertainty? Try fine-tuning LLMs to generate and explain their uncertainty, then see if this counteracts the negative effects of standard self-distillation on complex reasoning tasks.

Research Question: If we explicitly train LLMs to generate, explain, and reflect on their own uncertainty, can this intervention reverse the reasoning degradation observed after standard self-distillation, especially for mathematical or logic-heavy tasks?

Hypothesis: Directly supervising models to generate epistemic markers and self-explanations will mitigate the performance drop seen after self-distillation, especially for problems requiring extrapolation.

Experiment Plan: - Setup: Fine-tune models on a mix of tasks, with supervision to generate both answers and explicit uncertainty explanations (e.g., "I am unsure because...").

• Data: Use datasets with annotated uncertainty rationales, or generate such annotations using strong LLMs.
• Measurements: Compare reasoning accuracy, response diversity, and OOD performance with and without explicit uncertainty training, following self-distillation.
• Expected Outcomes: Models trained with explicit uncertainty rationales will resist self-distillation-induced degradation seen by Kim et al. (2026), especially on OOD math problems.

References:

• Kim, J., Luo, X., Kim, M., Lee, S., Kim, D., Jeon, J., Li, D., & Yang, Y. (2026). Why Does Self-Distillation (Sometimes) Degrade the Reasoning Capability of LLMs?
• Zong, Q., Liu, J., Zheng, T., Li, C., Xu, B., Shi, H., Wang, W., Wang, Z., Chan, C., & Song, Y. (2025). CritiCal: Can Critique Help LLM Uncertainty or Confidence Calibration? arXiv.org.