AI Paper Atlas

Research map / Continual learning / Continual fine-tuning of LLMs

Self-Distillation Enables Continual Learning

Idan Shenfeld, Mehul Damani, Jonas Hübotter, Pulkit Agrawal

ICML 2026 · Continual learning › Continual fine-tuning of LLMs · approach: Low-rank and model-merging adaptation · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use on-policy self-distillation (SDFT) instead of SFTQwen2.5-7B-Instruct · Science Q&Anew-task accuracy (%): 66.2 -> 70.2Improved
use SDFT instead of SFTQwen2.5-7B-Instruct · Science Q&Aprior-task average (%): 53.4 -> 64.5Improved
use SDFT instead of SFTQwen2.5-7B-Instruct · Tool Useprior-task average (%): 56.0 -> 65.4Improved
use SDFT instead of SFTQwen2.5-7B-Instruct · Medicalnew-task accuracy (%): 35.5 -> 40.2Improved
add re-invocation after SFTSFT · Science Q&Aprior-task average (%): 53.4 -> 60.2Improved
use SDFT instead of SFTQwen2.5-7B-Instruct · Knowledge acquisitionOOD accuracy (%): 80 -> 98Improved
use SDFT instead of SFTQwen2.5-7B-Instruct · Knowledge acquisitionstrict accuracy (%): 80 -> 89Improved
use SDFT instead of SFT on answer-only dataOlmo-3-7B-Think · Medicalaccuracy (%): 23.5 -> 43.7Improved
condition the teacher on article text plus answer rather than text only (ablation)SDFT · Knowledge acquisitionstrict accuracy (%): 75 -> 89Improved
train on-policy instead of offline distillation from the same teacher (ablation)SDFT · Tool UseaccuracyImproved
increase model size from 3B to 14B (ablation)SDFT vs SFT · Science Q&Agap over SFT: four-point (7B) to seven points (14B)Improved
use EMA teacher instead of frozen base or current student teacher (ablation)SDFT · training stabilitytraining stabilityImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy