AI Paper Atlas

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

Low-cost Full Fine-tuning: Learning What to Update for LLMs

Jin Li, Yaming Guo, Shenghao Gao, Xinlong Chen, Zuhao Xu, Ying Sun, Chao Wang, Hui Xiong

ICML 2026 · Large language models › Instruction and fine-tuning · approach: Instruction data selection

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
raise per-step budget from M=32 to M=128 (ablation)Think-Touch bandit selector (Qwen2.5-1.5B) · Commonsense reasoning avgaccuracy (%): 77.10 -> 81.14Improved
raise per-step budget from M=32 to M=128 (ablation)Think-Touch bandit selector (Llama-3.2-3B) · Commonsense reasoning avgaccuracy (%): 81.70 -> 83.83Improved
use online bandit selector (M=128) instead of LoRA r=64Qwen2.5-1.5B · Commonsense reasoning avgaccuracy (%): 80.16 -> 81.14Improved
use online bandit selector (M=128) instead of LoRA r=64Llama-3.2-3B · Commonsense reasoning avgaccuracy (%): 81.83 -> 83.83Improved
use bandit selector M=32 instead of LoRA r=32Qwen2.5-1.5B · Commonsense reasoning avgaccuracy (%): 79.70 -> 77.10Hurt
use bandit selector (M=128) instead of RandLoRA r=64Qwen2.5-1.5B · Commonsense reasoning avgaccuracy (%): 80.83 -> 81.14No clear effect
use bandit selector (M=128) instead of RandLoRA r=64Llama-3.2-3B · Commonsense reasoning avgaccuracy (%): 83.35 -> 83.83No clear effect
use bandit selector (M=128) instead of DoRA r=64Llama-3.2-3B · Commonsense reasoning avgaccuracy (%): 82.74 -> 83.83No clear effect
use bandit selector instead of uniform group selection (ablation)Think-Touch · Commonsense reasoning training losstraining lossImproved
update only selected rows (M=128) instead of full fine-tuningQwen2.5-1.5B · Commonsense reasoning avgaccuracy (%): 82.49 -> 81.14Hurt
update only selected rows (M=128) instead of full fine-tuningLlama-3.2-3B · Commonsense reasoning avgaccuracy (%): 84.50 -> 83.83Hurt
update selected rows per step instead of full fine-tuningQwen2.5-1.5B · wall-clock timeseconds per step: 1.2803s -> 0.1632sImproved

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