AI Paper Atlas

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

CPQS-Tuning: A Model Self-Perception-Based Data Filtering Algorithm for Efficient Instruction Fine-Tuning

Yi Ren, Yanhui Li, Tianyi Zhang, Diandong Liu

ICLR 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning generalization and delta compression

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
filter to 5k CPQS-selected samples instead of full 52kLlama2-7B on Alpaca GPT4 · Open LLM averageaverage accuracy: 54.37 -> 54.40Improved
filter Reasoning-DeepSeek to 10k CPQS samples instead of full 146kQwen2.5-7B-Instruct · Reasoning-DeepSeek averageaverage accuracy: 60.92 -> 65.45Improved
increase CPQS subset from 10k to 50kQwen2.5-7B-Instruct · Reasoning-DeepSeek averageaverage accuracy: 65.45 -> 65.02No clear effect
use CPQS selection instead of Superfiltering at 1kLlama2-7B · AlpacaEvalwin rate: 55.87 -> 55.98Improved
use CPQS selection on GSM8K 500 samples vs original modelQwen2.5-7B-Instruct · GSM8Kaccuracy: 76.27 -> 81.12Improved
use CPQS selection on Magicoder 1000 samplesLlama 2-7B-Chat · HumanEval avgpass@1: 12.5 -> 14.95Improved
use CPQS selection on Magicoder 1000 samplesQwen2.5-7B-Instruct · HumanEval avgpass@1: 79.25 -> 77.2Hurt
use hidden states from all layers instead of only early 9 layers (ablation)CPQS CNN selector · GSM8Kaccuracy: 84.23 -> 84.91Improved
use hidden states from all layers instead of final layer only (ablation)CPQS CNN selector · HumanEvalpass@1: 78.7 -> 80.0Improved
use CNN selector instead of MLP (ablation)CPQS · Reasoning-DeepSeek 10k averageaverage accuracy: 62.38 -> 65.45Improved
use CNN selector instead of Transformer (ablation)CPQS · Reasoning-DeepSeek 10k averageaverage accuracy: 63.81 -> 65.45Improved
use own-model selected data instead of swapped dataset (ablation)Qwen2.5-7B-Instruct · GSM8Kaccuracy: 83.24 -> 84.91Improved

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