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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| apply instruction tuning to the pretrained model | PT models · Spectrum Suite categorical tasks | accuracy: significantly decreases accuracy in 35 of 76 comparisons, increases in 7 | Hurt |
| apply instruction tuning to the pretrained model | PT models · general capability ICL (8 datasets) | accuracy: increases in 8 of 24 cases, same in 13, decreases in 2 | No clear effect |
| apply instruction tuning to the pretrained model | PT models · free-text Spectrum Suite | loss: IT loss worse than PT in 117 of 144 comparisons, better in 2 | Hurt |
| use Spectrum Tuning instead of the pretrained model | PT models · Spectrum Suite-Test multiple-choice | loss: lower loss in 14 of 15 comparisons | Improved |
| use Spectrum Tuning instead of the pretrained model | PT models · distributional alignment (7 datasets) | JS-divergence: improves in 10 of 21 cases, matches in 10, degrades in 1 | Improved |
| use Spectrum Tuning instead of the instruction-tuned model | IT models · zero-shot verifiable tasks, gemma-3-12b | yield: 6.2 -> 40.5 | Improved |
| use Spectrum Tuning instead of the instruction-tuned model | IT models · zero-shot verifiable tasks, Llama-3.1-8B | yield: 24.0 -> 37.4 | Improved |
| compute loss on first output only instead of all outputs (ablation) | Spectrum Tuning · ablation average | yield, 3-shot: 33.0 vs 42.1 with default | Improved |
| compute loss on all outputs rather than last output only (ablation) | Spectrum Tuning · ablation average | distributional alignment JS-divergence: .103 -> .069 | Improved |
| mask loss on description/input tokens (outputs only) (ablation) | Spectrum Tuning · ablation average | yield, description only: 33.0 -> 36.7 | Improved |
| train on Spectrum Suite instead of capability data (3.9k sequences) (ablation) | Spectrum Tuning · ablation average | distributional alignment JS-divergence: .111 -> .086 | Improved |
| initialize special tokens from IT embeddings instead of random (ablation) | Spectrum Tuning · ablation average | yield, description only: 21.0 -> 36.7 | Improved |
Closest papers by meaning
- Focus On This, Not That! Steering LLMs with Adaptive Feature Specification (ICML 2025 · similarity 0.87)
- DELIFT: Data Efficient Language model Instruction Fine-Tuning (ICLR 2025 · similarity 0.87)
- Steering Information Utility in Key-Value Memory for Language Model Post-Training (NeurIPS 2025 · similarity 0.87)
- Fixing It in Post: A Comparative Study of LLM Post-Training Data Quality and Model Performance (NeurIPS 2025 · similarity 0.87)
- Steering When Necessary: Flexible Steering Large Language Models with Backtracking (NeurIPS 2025 · similarity 0.86)
- Consolidating Rewarded Perturbations for LLM Post-Training (NeurIPS 2026 · similarity 0.86)
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.
