Research map / Large language models / Instruction and fine-tuning
Reducing the Scope of Language Models
David Yunis, Siyu Huo, Chulaka Gunasekara, Danish Contractor
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 |
|---|---|---|---|
| add system prompt only (no training) for scoping | Mistral-7B-Instruct-v0.2 · Classification and Generation accept set, OOD Reject | OOD rejection rate: 0.42 vs 0.91 (Probe), 0.79 (CB) | Hurt |
| use CB instead of SFT for out-of-distribution rejection | Mistral-7B-Instruct-v0.2 · Math and Programming accept set, OOD Reject | OOD rejection rate: 0.52 -> 0.96 | Improved |
| use SFT instead of CB for accept-set generation quality | Mistral-7B-Instruct-v0.2 · Classification and Generation accept set, Accept Score | Accept Score: 0.25 -> 0.46 | Improved |
| layer CB on top of SFT | SFT · Classification and Generation accept set, OOD Reject | OOD rejection rate: 0.28 -> 0.54 | Improved |
| layer CB on top of SFT | SFT · Math and Programming accept set, Accept rejection rate | rejection rate on Accept set: 0.0 -> 0.26 | Hurt |
| layer CB on top of SFT | SFT · Math and Programming accept set, OOD Reject | OOD rejection rate: 0.52 -> 0.64 | Improved |
| increase diversity of the rejection set (ablation) | SFT vs CB · Sentiment Analysis accept, multiple models | rejection rates | Improved |
| scale model size from 1B to 8B (Llama-3.x) (ablation) | all scoping methods · Sentiment Analysis accept | scoping quality | Improved |
| use probe on small models (1B) (ablation) | Probe · Sentiment Analysis accept, Llama-1B/3B | scoping quality | Hurt |
| use CB instead of SFT/DPO under adversarial prompts | Mistral-7B-Instruct-v0.2 · Adversarial evaluation (Appendix A.4) | rejection robustness | Improved |
| increase training data quantity for CB (ablation) | CB · Sentiment Analysis accept | scoping quality | Hurt |
| increase training data quantity for DPO (ablation) | DPO · Sentiment Analysis accept | scoping quality | Improved |
Closest papers by meaning
- Supervised Knowledge Makes Large Language Models Better In-context Learners (ICLR 2024 · similarity 0.89)
- Listening to the Wise Few: Query–Key Alignment Unlocks Latent Correct Answers in Large Language Models (NeurIPS 2026 · similarity 0.88)
- HonestLLM: Toward an Honest and Helpful Large Language Model (NeurIPS 2024 · similarity 0.87)
- TACT: Advancing Complex Aggregative Reasoning with Information Extraction Tools (NeurIPS 2024 · similarity 0.87)
- TAIA: Large Language Models are Out-of-Distribution Data Learners (NeurIPS 2024 · similarity 0.87)
- Large Language Model Cascades with Mixture of Thought Representations for Cost-Efficient Reasoning (ICLR 2024 · similarity 0.87)
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.
