Beyond Conciseness: Adaptive Self-Distillation Guided by Problem Structure and Uncertainty

TL;DR: Instead of always pushing for shorter reasoning, let's teach models to be concise only when it's safe—using signals from model confidence and problem structure. Try training with dynamic instructions (e.g., "be concise if confident, elaborate if unsure") and test if adaptive conciseness outperforms one-size-fits-all OPSDC, especially for ambiguous or multi-step problems.

Research Question: Can conditioning self-distillation on model uncertainty and problem complexity, rather than always conciseness, yield better trade-offs between compression and accuracy in reasoning models?

Hypothesis: Adapting the distillation signal to encourage brevity only when the model is confident (as measured by entropy or calibration proxies) will further improve accuracy, robustness, and avoid harmful under-explanation on difficult or ambiguous problems compared to always enforcing conciseness.

Experiment Plan: Modify OPSDC to incorporate confidence-based dynamic instructions, e.g., "be concise if the model is confident" or "provide more steps if uncertain". Use uncertainty estimates (entropy, margin, etc.) and task metadata (number of reasoning steps required, from datasets like FDA/THR [4]) to scaffold the instruction. Train models on GSM8K, MATH-500, and synthetic datasets with controllable reasoning complexity and label noise. Measure accuracy, token reduction, and the incidence of under-explained errors compared to vanilla OPSDC and baselines like ReCUT [ReCUT, 3]. Analyze where adaptive conciseness helps or hinders, especially on long-horizon or noisy tasks.

References:

• 1. Sang, H., Xu, Y., Zhou, Z., He, R., Wang, Z., & Sun, J. (2026). On-Policy Self-Distillation for Reasoning Compression.
• 2. Jin, Z., Li, X., Ji, Y., Peng, C., Liu, Z., Shi, Q., Yan, Y., Wang, S., & Peng, F., Yu, G. (2025). ReCUT: Balancing Reasoning Length and Accuracy in LLMs via Stepwise Trails and Preference Optimization. Conference on Empirical Methods in Natural Language Processing.
• 3. Du, Y., Mondorf, P., Casola, S., Yao, Y., Litschko, R., & Plank, B. (2025). Reason to Rote: Rethinking Memorization in Reasoning. Conference on Empirical Methods in Natural Language Processing.