AI Paper Atlas

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

Retaining by Doing: The Role of On-Policy Data in Mitigating Forgetting

Howard Chen, Noam Razin, Karthik R Narasimhan, Danqi Chen

ICML 2026 · Continual learning › Continual fine-tuning of LLMs · approach: Continual pretraining of language models · 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 RL (GRPO) instead of SFTpost-training of instruct LMs · IFEval/MMLU/Countdown, Llama and Qwennon-target tasks drop: Llama 8B MMLU drop 38.5 -> -0.2Improved
remove KL regularization from GRPO (beta 0.05 -> 0.0) (ablation)GRPO · IFEval/MMLU/Countdown, four modelsgain-drop tradeoffNo clear effect
remove KL regularization from GRPO (ablation)GRPO · Llama models on IFEvalnon-target tasks dropHurt
replace GRPO advantage estimate with plain REINFORCE (ablation)GRPO · Llama 3.1 8B Instruct, IFEvalnon-target tasks drop: 3.4 -> 7.7No clear effect
replace GRPO with REINFORCE (ablation)GRPO · Llama 3.1 8B Instruct, Countdowntarget gain: 60.4 -> 7.5Hurt
replace GRPO with REINFORCE (ablation)GRPO · Qwen 2.5 7B Instruct, IFEvaltarget gain: 17.0 -> 5.7Hurt
regenerate SFT data at the start of each epoch (Iterative-SFT) instead of only from initial policy (ablation)Self-SFT · Qwen 2.5 1.5B/7B, IFEval and MMLUnon-target drop relative to GRPOImproved
use data generated only from the initial policy (Self-SFT) instead of expert data (ablation)SFT · IFEval/MMLU/Countdownnon-target tasks dropMixed
lower SFT learning rate to 1e-5 (ablation)Self-SFT · IFEval/MMLU/Countdowngain and dropMixed
SFT on data generated during an RL run (ablation)SFT · Appendix A.4.1non-target dropImproved
start with a multi-modal rather than uni-modal training policy (ablation)reverse KL vs forward KL simulation · Gaussian mixture simulationnon-target drop (overlap): 0.7 -> 0.03Improved

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