AI Paper Atlas

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

Reducing the Scope of Language Models

David Yunis, Siyu Huo, Chulaka Gunasekara, Danish Contractor

AAAI 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning generalization and delta compression · 3 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 system prompt only (no training) for scopingMistral-7B-Instruct-v0.2 · Classification and Generation accept set, OOD RejectOOD rejection rate: 0.42 vs 0.91 (Probe), 0.79 (CB)Hurt
use CB instead of SFT for out-of-distribution rejectionMistral-7B-Instruct-v0.2 · Math and Programming accept set, OOD RejectOOD rejection rate: 0.52 -> 0.96Improved
use SFT instead of CB for accept-set generation qualityMistral-7B-Instruct-v0.2 · Classification and Generation accept set, Accept ScoreAccept Score: 0.25 -> 0.46Improved
layer CB on top of SFTSFT · Classification and Generation accept set, OOD RejectOOD rejection rate: 0.28 -> 0.54Improved
layer CB on top of SFTSFT · Math and Programming accept set, Accept rejection raterejection rate on Accept set: 0.0 -> 0.26Hurt
layer CB on top of SFTSFT · Math and Programming accept set, OOD RejectOOD rejection rate: 0.52 -> 0.64Improved
increase diversity of the rejection set (ablation)SFT vs CB · Sentiment Analysis accept, multiple modelsrejection ratesImproved
scale model size from 1B to 8B (Llama-3.x) (ablation)all scoping methods · Sentiment Analysis acceptscoping qualityImproved
use probe on small models (1B) (ablation)Probe · Sentiment Analysis accept, Llama-1B/3Bscoping qualityHurt
use CB instead of SFT/DPO under adversarial promptsMistral-7B-Instruct-v0.2 · Adversarial evaluation (Appendix A.4)rejection robustnessImproved
increase training data quantity for CB (ablation)CB · Sentiment Analysis acceptscoping qualityHurt
increase training data quantity for DPO (ablation)DPO · Sentiment Analysis acceptscoping qualityImproved

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