Research map / Large language models / Language, safety and interpretability
Don't Throw Away Your Pretrained Model
Shangbin Feng, Wenhao Yu, Yike Wang, Hongming Zhang, Yulia Tsvetkov, Dong Yu
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 trained switcher to alternate pretrained, finetuned, aligned models by segment | aligned Tulu-3-8B · TruthfulQA | task score: 29.01 -> 39.22 | Improved |
| add trained switcher (switch-task) | aligned Tulu-3-8B · GSM8k | task score: 56.80 -> 59.60 | Improved |
| add trained switcher (switch-task) | aligned Tulu-3-8B · BBH | task score: 35.20 -> 58.30 | Improved |
| add trained switcher (switch-task) | aligned Tulu-3-8B · Alpaca | task score: 57.46 -> 56.89 | Hurt |
| add trained switcher (switch-task) | aligned Tulu-3-8B · Poem | task score: 77.70 -> 70.25 | Hurt |
| use trained switcher instead of greedy soup | greedy soup · BBH | task score: 36.50 -> 58.30 | Improved |
| switcher outperforms all baselines and individual models | switch generation · 18 datasets | task score: best on 13 of 18 datasets, 12.9% average relative improvement | Mixed |
| fine-tune the switcher instead of random switching (ablation) | switch-task · TruthfulQA | task score: 27.07 -> 39.22 | Improved |
| fine-tune the switcher instead of untuned switching (ablation) | switch-task · BBH | task score: 41.80 -> 58.30 | Improved |
| random switching instead of trained switcher (ablation) | switch-task · BBH | task score: 58.30 vs 53.10 | Mixed |
| use patch size 50 instead of 100 (ablation) | switch-task · GSM8k | task score: 44.70 -> 59.60 | Improved |
| use smaller patch sizes (20) instead of 50 (ablation) | switch-task · Pluralism | task score: 53.20 -> 56.80 | Improved |
Closest papers by meaning
- Aligning Language Models with Demonstrated Feedback (ICLR 2025 · similarity 0.88)
- Your Weak LLM is Secretly a Strong Teacher for Alignment (ICLR 2025 · similarity 0.88)
- Aligner: Efficient Alignment by Learning to Correct (NeurIPS 2024 · similarity 0.88)
- Training on the Test Task Confounds Evaluation and Emergence (ICLR 2025 · similarity 0.88)
- Theoretical Limits of Language Model Alignment (NeurIPS 2026 · similarity 0.87)
- Cross-model Control: Improving Multiple Large Language Models in One-time Training (NeurIPS 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.
