AI Paper Atlas

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

Learning from the Undesirable: Robust Adaptation of Language Models Without Forgetting

Yunhun Nam, Jaehyung Kim, Jongheon Jeong

AAAI 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning generalization and delta compression · 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 LfU consistency regularization to SFTLlama-3.1-8B · Math avg, GSM8k fine-tunedaccuracy: 51.5 -> 54.2Improved
add LfU to SFT on Alpagasus Dolly 3kLlama-3.1-8B · Math avgaccuracy: 37.0 -> 43.2Improved
add LfU to SFTLlama-3.1-8B · HEx-PHI ASR after 5-step attackASR: 34.8 -> 5.8Improved
add LfU to SFTLlama-3.1-8B-Instruct · Math avg, GSM8k fine-tunedaccuracy: 65.7 -> 66.7Improved
use representation steering instead of LoRA auxiliary (ablation)LfU · Math avg, GSM8kaccuracy: 54.2 -> 53.2; time 4142.4 -> 2332.1 msMixed
raise consistency weight lambda from 5.0 to 10.0 (ablation)LfU · Math avgaccuracy: 54.2 -> 53.4Mixed
large step size alpha 0.5 (ablation)LfU · Math/Knowledge avgaccuracy: 54.2 -> 44.5Hurt
apply consistency loss to all layers rather than a subset (ablation)LfU · Math avgaccuracy: 53.3-53.6 -> 54.2Improved
use representation-level consistency instead of logit-level (ablation)LfU · Math avgaccuracy: 53.9 -> 54.2Improved
use undesirable-update consistency instead of feature L1/L2 penalties (ablation)LfU · Knowledge avgaccuracy: 72.2/73.0 -> 73.5Improved
use undesirable-update consistency vs direct SAM-style update (ablation)LfU · Math avgaccuracy: 53.3 -> 54.2Improved
add LfU regularizationLlama-3.1-8B · GSM8k prompt variationsstd of accuracy: 92.1% lower stdImproved

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