AI Paper Atlas

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

Explainable Token-level Noise Filtering for LLM Fine-tuning Datasets

Yuchen Yang, Wenze Lin, Enhao Huang, Zhixuan Chu, Hongbin zhou, Lan Tao, Yiming Li, Zhan Qin, Kui Ren

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
add XTF token-level gradient masking to fine-tuningLLM fine-tuning · GSM8K (average of 10 cases)accuracy: 37.1 -> 45.7Improved
add XTF token-level gradient masking to fine-tuningLLM fine-tuning · PubMedQA (average of 10 cases)accuracy: 30.4 -> 37.1Improved
use XTF instead of sample-level data filteringDeepseek-distilled-qwen-1.5B · GSM8Kaccuracy: 47.0 -> 56.2Improved
use XTF instead of selective language model trainingLlama-3.1-8B full fine-tuning · GSM8Kaccuracy: 60.3 -> 58.7Hurt
train for twice as many epochsNormal fine-tuning · PubMedQA (average)accuracy: 30.4 -> 29.9No clear effect
add data augmentationNormal fine-tuning · GSM8K (average)accuracy: 37.1 -> 36.5No clear effect
add all three attribute filters (RI, KN, TR) instead of none (ablation)XTF · GSM8K / PubMedQA (average of 6)accuracy: 30.7 -> 40.1Improved
add KN filtering to RI+TR (ablation)XTF · GSM8K / PubMedQA (average of 6)accuracy: 36.9 -> 40.1Improved
add TR filtering to RI+KN (ablation)XTF · GSM8K / PubMedQA (average of 6)accuracy: 36.1 -> 40.1Improved
add RI filtering to KN+TR (ablation)XTF · GSM8K / PubMedQA (average of 6)accuracy: 36.3 -> 40.1Improved
filter with KN alone versus RI alone (ablation)XTF single attribute · GSM8K / PubMedQA (average of 6)accuracy: 32.0 -> 34.5Improved
apply XTF to larger base models rather than smallerXTF on code · HumanEvalpass@kImproved

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