AI Paper Atlas

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

Token-level Data Selection for Safe LLM Fine-tuning

Yanping Li, Zhening Liu, Zijian Li, Zehong Lin, Jun Zhang

ICLR 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning generalization and delta compression · 4 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 progressive refinement (TOSS-Pro) on top of TOSS (ablation)TOSS, Llama-3-8B-Instruct · HEx-PHIwin rate: 87.54 -> 93.79Improved
use metric-driven instead of random selection of samples for refining the safety-degraded model (ablation)TOSS-Pro · HEx-PHIwin rate: 87.37 -> 93.79Improved
use metric-driven instead of random selection (ablation)TOSS-Pro · SlimOrca testwin rate: 63.12 -> 68.85Improved
add a second progressive iteration (ablation)TOSS-Pro · HEx-PHIwin rate: 91.13 -> 93.79Improved
add a second progressive iteration (ablation)TOSS-Pro · ANTHROPIC HH testwin rate: 88.54 -> 88.85No clear effect
token-level TOSS instead of sample-level SEALLlama-3-8B-Instruct · HEx-PHIwin rate: 68.83 -> 87.54Improved
token-level TOSS instead of sample-level SEALLlama-3-8B-Instruct · SlimOrca testwin rate: 57.41 -> 68.37Improved
token-level TOSS instead of SEAL at 30% harmful datafine-tuning mixture · HEx-PHIwin rate: 63.37 -> 69.71Improved
token-level TOSS instead of SEAL at 70% harmful datafine-tuning mixture · SlimOrca testwin rate: 51.30 -> 64.48Improved
token-level TOSS instead of SEAL on benign datafine-tuning on benign data · HEx-PHIwin rate: 55.57 -> 55.75No clear effect
token-level TOSS instead of SEAL on GSM8K fine-tuningLlama-3-8B-Instruct · HEx-PHIwin rate: 53.04 -> 93.18Improved
token-level TOSS instead of SEAL on GSM8K fine-tuningLlama-3-8B-Instruct · GSM8Kwin rate: 60.62 -> 62.97Improved

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