Research map / Probabilistic and scientific modeling / Bayesian optimization
Scaling Multi-Task Bayesian Optimization with Large Language Models
Yimeng Zeng, Natalie Maus, Haydn Thomas Jones, Jeffrey Tao, Fangping Wan, Marcelo Der Torossian Torres, Cesar de la Fuente-Nunez, Ryan Marcus, Osbert Bastani, Jacob R. Gardner
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 |
|---|---|---|---|
| fine-tune LLM on more BO trajectories (1426 vs 1138 tasks) (ablation) | BOLT · Query plan optimization, Best@50 | summed runtime: 78.16 -> 63.68 | Improved |
| fine-tune on 1426 tasks versus 1138 tasks, one-shot (ablation) | BOLT · Query plan optimization, Best@1 | summed runtime: 202.04 -> 160.22 | Improved |
| replace old solutions with newer better ones (same training size) versus keeping old data (ablation) | BOLT-1138 · Query plan optimization, Best@50 | summed runtime: 78.16 -> 64.03 | Improved |
| add more data on top of replaced better data (BOLT-1426 vs BOLT-1138*) (ablation) | BOLT · Query plan optimization, Best@50 | summed runtime: 64.03 -> 63.68 | Improved |
| add self-augmentation (fine-tune on own best generations) without extra BO runs (ablation) | BOLT-1138 and BOLT-1426 · Query plan optimization | summed runtime on 100 validation queries: about 62 seconds for both self-augmented models | Improved |
| use greedy decoding (temperature 0) instead of 0.7 (ablation) | BOLT · Query plan optimization, one-shot | one-shot performance | Improved |
| fine-tune open-source Qwen-2.5-7B instead of GPT-4o-mini on BOLT-1426 data (ablation) | BOLT · Query plan optimization, Best@50 | summed runtime: 61.54 -> 62.04 | No clear effect |
| fine-tune open-source Llama-3.1-8B instead of GPT-4o-mini (ablation) | BOLT · Query plan optimization, Best@50 | summed runtime: 61.54 -> 155.55 | Hurt |
| initialize BO with best solutions of previously solved tasks instead of BOLT samples (ablation) | BOLT-1426 · Query plan optimization, first 10 validation tasks | summed runtime: 7.21 -> 9.18 | Hurt |
| run full BO after BOLT-1426 initialization (ablation) | BOLT-1426 · Query plan optimization, first 10 validation tasks | summed runtime: 7.21 -> 6.43 | Improved |
| scale BOLT training tasks beyond about 20 versus multi-task GP (DKT/FSBO) | BOLT vs DKT/FSBO · Query plan optimization BO curves | BO performance | Improved |
| more experts in POGPE/SGPE ensembles | POGPE/SGPE · Query plan and peptide design | BO performance | Mixed |
Closest papers by meaning
- Large Language Models to Enhance Bayesian Optimization (ICLR 2024 · similarity 0.90)
- OptScale: Probabilistic Optimality for Inference-time Scaling (AAAI 2026 · similarity 0.88)
- Adaptive Kernel Design for Bayesian Optimization Is a Piece of CAKE with LLMs (NeurIPS 2025 · similarity 0.87)
- BOFormer: Learning to Solve Multi-Objective Bayesian Optimization via Non-Markovian RL (ICLR 2025 · similarity 0.87)
- REASONING COMPILER: LLM-Guided Optimizations for Efficient Model Serving (NeurIPS 2025 · similarity 0.87)
- AgentTTS: Large Language Model Agent for Test-time Compute-optimal Scaling Strategy in Complex Tasks (NeurIPS 2025 · 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.
