AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

DRIFT: Difference-Aware Reinforcement Through Iterative Fine-Tuning for Language Model

Wenjie Liao, Xiaohui Song, Haonan Lu

AAAI 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning generalization and delta compression · 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
add DRIFT iteration 1 after SFT epoch 1Qwen2.5-3B-Instruct · Summarization 5-task averageROUGE-1: 44.00 -> 47.29Improved
add DRIFT iteration 1 after SFT epoch 1Llama-3.1-8B-Instruct · Summarization 5-task averageROUGE-1: 45.15 -> 51.61Improved
add DRIFT iteration 1 after SFT epoch 1Llama-3.1-8B-Instruct · SAMSumROUGE-1: 49.93 -> 65.94Improved
add DRIFT iteration 1 after SFT epoch 1Qwen2.5-3B-Instruct · SAMSumROUGE-1: 47.47 -> 54.64Improved
train a second SFT epoch instead of DRIFTQwen2.5-3B-Instruct · Summarization 5-task averageROUGE-1: 44.00 -> 43.71Hurt
train a second SFT epoch instead of DRIFTLlama-3.1-8B-Instruct · Summarization 5-task averageROUGE-2: 27.03 -> 25.50Hurt
run more DRIFT iterations (iteration 4 vs 1)Llama-3.1-8B-Instruct · Summarization 5-task averageROUGE-1: 51.61 -> 55.04Improved
use DRIFT instead of SPIN at iteration 1Qwen2.5-3B-Instruct · Summarization 5-task averageROUGE-1: 46.03 -> 47.29Improved
add Difference-Aware Masking (DAM) (ablation)DRIFT · Summarization 5-task averageROUGE-1: 46.49 -> 47.29Improved
add Occurrence-Aware Loss (OAL) (ablation)DRIFT · Summarization 5-task averageROUGE-1: 46.99 -> 47.29Improved
update opponent each iteration instead of training more epochs on fixed data (ablation)DRIFT · Summarization 5-task averageROUGE-1: 47.45 -> 48.35Improved
add DRIFT iteration 1 after SFT epoch 1Qwen2.5-3B-Instruct · Open LLM Leaderboard averageaccuracy: 61.47 -> 63.07Improved

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