AI Paper Atlas

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

Spectrum Tuning: Post-Training for Distributional Coverage and In-Context Steerability

Taylor Sorensen, Benjamin Newman, Jared Moore, Chan Young Park, Jillian Fisher, Niloofar Mireshghallah, Liwei Jiang, Yejin Choi

ICLR 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning for reasoning data · 11 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
apply instruction tuning to the pretrained modelPT models · Spectrum Suite categorical tasksaccuracy: significantly decreases accuracy in 35 of 76 comparisons, increases in 7Hurt
apply instruction tuning to the pretrained modelPT models · general capability ICL (8 datasets)accuracy: increases in 8 of 24 cases, same in 13, decreases in 2No clear effect
apply instruction tuning to the pretrained modelPT models · free-text Spectrum Suiteloss: IT loss worse than PT in 117 of 144 comparisons, better in 2Hurt
use Spectrum Tuning instead of the pretrained modelPT models · Spectrum Suite-Test multiple-choiceloss: lower loss in 14 of 15 comparisonsImproved
use Spectrum Tuning instead of the pretrained modelPT models · distributional alignment (7 datasets)JS-divergence: improves in 10 of 21 cases, matches in 10, degrades in 1Improved
use Spectrum Tuning instead of the instruction-tuned modelIT models · zero-shot verifiable tasks, gemma-3-12byield: 6.2 -> 40.5Improved
use Spectrum Tuning instead of the instruction-tuned modelIT models · zero-shot verifiable tasks, Llama-3.1-8Byield: 24.0 -> 37.4Improved
compute loss on first output only instead of all outputs (ablation)Spectrum Tuning · ablation averageyield, 3-shot: 33.0 vs 42.1 with defaultImproved
compute loss on all outputs rather than last output only (ablation)Spectrum Tuning · ablation averagedistributional alignment JS-divergence: .103 -> .069Improved
mask loss on description/input tokens (outputs only) (ablation)Spectrum Tuning · ablation averageyield, description only: 33.0 -> 36.7Improved
train on Spectrum Suite instead of capability data (3.9k sequences) (ablation)Spectrum Tuning · ablation averagedistributional alignment JS-divergence: .111 -> .086Improved
initialize special tokens from IT embeddings instead of random (ablation)Spectrum Tuning · ablation averageyield, description only: 21.0 -> 36.7Improved

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