AI Paper Atlas

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

TokenSeek: Memory Efficient Fine Tuning via Instance-Aware Token Ditching

Runjia Zeng, Qifan Wang, Qiang Guan, Ruixiang Tang, Lifu Huang, Zhenting Wang, XUELING ZHANG, Cheng Han, Dongfang Liu

ICLR 2026 · Large language models › Instruction and fine-tuning · approach: Efficient LLM fine-tuning systems · 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 TokenSeek instead of random token selection (TokenTune) with QLoRAQwen2.5-0.5B QLoRA · Few-shot QA averageaverage score: 46.34 -> 48.45Improved
add TokenSeek instead of random token selection with QLoRALlama3.2-1B QLoRA · Few-shot QA averageaverage score: 49.22 -> 52.61Improved
add TokenSeek instead of random token selection with LoHaLlama3.2-3B LoHa · Few-shot QA averageaverage score: 59.13 -> 61.02Improved
add TokenSeek (10% tokens) to full-parameter tuningQwen2.5-0.5B full tuning · Few-shot QA averageaverage score: 48.43 -> 41.94Hurt
add TokenSeek (10% tokens) to full-parameter tuningLlama3.2-1B full tuning · Few-shot QA averageaverage score: 40.82 -> 41.13No clear effect
add TokenSeek to QLoRA, reducing memoryLlama2-7B QLoRA · Few-shot QA averageaverage score: 62.55 -> 62.14Hurt
add gradient information to context-only scoring (alpha=5, beta=5) (ablation)TokenSeek on Qwen2.5-0.5B QLoRA · Few-shot QA averageaverage score: 48.45 -> 48.65Improved
add context information to gradient-only scoring (alpha=5, beta=5) (ablation)TokenSeek on Qwen2.5-0.5B QLoRA · Few-shot QA averageaverage score: 46.39 -> 48.65Improved
raise tunable token ratio from 10% to 30% (ablation)TokenSeek on Llama3.2-1B QLoRA · Few-shot QA averageaverage score: 52.61 -> 52.75No clear effect
lower tunable token ratio from 10% to 3% (ablation)TokenSeek on Qwen2.5-0.5B QLoRA · Few-shot QA averageaverage score: 48.45 -> 43.79Hurt
raise tunable token ratio from 10% to 60% (ablation)TokenSeek on Qwen2.5-0.5B QLoRA · Few-shot QA averageaverage score: 48.45 -> 48.43No clear effect
add 100K MiniPile examples to Open-Platypus training data (ablation)TokenSeek QLoRA on Llama2-7B · Few-shot QA averageaverage score: 62.14 -> 63.26Improved

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