Skip to content

Why Does Self-Distillation (Sometimes) Degrade the Reasoning Capability of LLMs?

Jeonghye Kim Xufang Luo Minbeom Kim Sangmook Lee Dohyung Kim Jiwon Jeon Dongsheng Li Yuqing Yang
Aug 2026
Machine Learning Natural Language Processing

Abstract

Self-distillation has emerged as an effective post-training paradigm for LLMs, often improving performance while shortening reasoning traces. However, in mathematical reasoning, we find that it can reduce response length while degrading performance. We trace this degradation to the suppression of epistemic verbalization - the model's expression of uncertainty during reasoning. Through controlled experiments varying conditioning context richness and task coverage, we show that conditioning the teacher on rich information suppresses uncertainty expression, enabling rapid in-domain optimization with limited task coverage but harming OOD performance, where unseen problems benefit from expressing uncertainty and adjusting accordingly. Across Qwen3-1.7B/8B, DeepSeek-Distill-Qwen-7B, and Olmo3-7B-Instruct, we observe performance drops of up to 40%. Our findings highlight that exposing appropriate levels of uncertainty is crucial for robust reasoning and underscore the importance of optimizing reasoning behavior beyond merely reinforcing correct answer traces.

View source

Similar papers

#machine learning Open access Oct 2020

Physics-Informed and Hybrid Machine Learning in Additive Manufacturing: Application to Fused Filament Fabrication

This article investigates several physics-informed and hybrid machine learning strategies that incorporate physics knowledge in experimental data-driven deep-learning models for predicting the bond quality and porosity of fused filament fabrication (FFF) parts. Three types of strategies are explored to incorporate physics constraints and multi-physics FFF simulation results into a deep neural network (DNN), thus ensuring consistency with physical laws: (1) incorporate physics constraints within the loss function of the DNN, (2) use physics model outputs as additional inputs to the DNN model, and (3) pre-train a DNN model with physics model input-output and then update it with experimental data. These strategies help to enforce a physically consistent relationship between bond quality and tensile strength, thus making porosity predictions physically meaningful. Eight different combinations of the above strategies are investigated. The results show how the combination of multiple strategies produces accurate machine learning models even with limited experimental data.

B. Kapusuzoglu, S. Mahadevan · 79 citations · ⚡2

Spikformer V2: Join the High Accuracy Club on ImageNet with an SNN Ticket

This work introduces a pioneering exploration of Self-Supervised Learning (SSL) within the SNN, and proposes a novel Spiking Self-Attention (SSA) and Spiking Transformer (Spikformer) that achieves 80+% accuracy on ImageNet.

Zhaokun Zhou, Kaiwei Che, Wei Fang et al. · 69 citations · ⚡10
#machine learning Open access May 2021

Information fusion and machine learning for sensitivity analysis using physics knowledge and experimental data

This paper considers global sensitivity analysis (GSA) for situations where both a physics-based model and experimental observations are available, and investigates physics-informed machine learning strategies to effectively combine the two sources of information in order to maximize the accuracy of the sensitivity estimate.

B. Kapusuzoglu, S. Mahadevan · 45 citations
#machine learning Conference Open access Feb 2023

EquiPocket: an E(3)-Equivariant Geometric Graph Neural Network for Ligand Binding Site Prediction

EquiPocket is proposed, an E(3)-equivariant Graph Neural Network for binding site prediction, which comprises three modules: the first one to extract local geometric information for each surface atom, the second one to model both the chemical and spatial structure of protein and the last one to capture the geometry of the surface via equivariant message passing over the surface atoms.

Yang Zhang, Wenbing Huang, Zhewei Wei et al. · 43 citations · ⚡4
#artificial intelligence Open access Oct 2022

Adaptive surrogate modeling for high-dimensional spatio-temporal output

An adaptive surrogate modeling method for problems with very high-dimensional spatio-temporal outputs is developed that combines exploration and exploitation to improve the surrogate model accuracy with the fewest possible runs of the expensive physics-based model.

B. Kapusuzoglu, S. Mahadevan, Shunsaku Matsumoto et al. · 17 citations
#machine learning Open access Nov 2022

Doubly robust nearest neighbors in factor models

An improved variant of nearest neighbors (NN) for estimation with missing data in latent factor models that provides a (near-)quadratic improvement in the non-asymptotic error and admits a significantly narrower asymptotic confidence interval when compared to both unit-unit or time-time NN.

Raaz Dwivedi, Katherine Tian, Sabina Tomkins et al. · 11 citations

Related blog posts

MIT News · Artificial Intelligence Aug 27, 2026

Looking beyond natural sequences

A new machine-learning framework aims to improve the success rate of computational protein design while moving away from results that reproduce sequences found in nature.