Curated papers & projects (2023→) that combine Large Language Models (LLMs) with Bayesian Optimization (BO).
Goal: help researchers quickly survey strategies, applications, and design patterns for LLM-based BO.
LLMs are increasingly used to warm-start, guide acquisition, embed complex spaces, and even generate BO algorithms.
Each entry lists year, scope, strategy, highlights, and a source link.
We include papers, preprints, and project pages that either:
- use BO with LLMs (e.g., LLM-guided acquisition, embedding, warm-start), or
- use LLMs to guide/accelerate a BO procedure.
Coverage: 2023–2025 (and growing). Please contribute via PRs/issues (see How to Contribute).
Longer notes follow the table to keep things tidy.
| Year | Paper | Application / Domain | LLM + BO Strategy | Highlights | Source |
|---|---|---|---|---|---|
| 2025 | LLaMEA-BO: LLM-aided Evolutionary Algorithm for BO | Meta-optimization | LLM + evolution to auto-generate BO algorithm code (init/surrogate/acquisition) | Outperforms several BO baselines on BBOB; no extra finetuning | https://arxiv.org/abs/2505.21034 |
| 2025 | Reasoning BO | General & chemistry | LLM reasoning + multi-agent + knowledge graph coupled to BO | Real-time sampling recs; improved yields in a direct arylation example | https://arxiv.org/abs/2505.12833 |
| 2025 | Distilling and Exploiting Quantitative Insights from LLMs for Chemical Reaction BO | Chemical reaction optimization | Prompt LLM to elicit priors; train a utility guiding BO | Learned utility correlates with yields; speeds optimization | https://arxiv.org/abs/2504.08874 |
| 2025 | GOLLuM: Gaussian Process Optimized LLMs | Reaction optimization & finetuning | Treat finetuning via GP ML objective; BO tunes hyperparams with LLM kernel | Higher discovery rate; +14% vs domain-specific reps | https://arxiv.org/abs/2504.06265 |
| 2024 | LLAMBO: Large Language Models to Enhance BO | General black-box | Frame BO in natural language; LLM proposes candidates + warm-starts | Strong early-stage performance; modular components | https://openreview.net/forum?id=OOxotBmGol |
| 2024 | Language Model Embeddings for BO (Embed-then-Regress) | General black-box | Use LLM to embed string inputs; regress in embedding space; apply BO | Enables BO over arbitrary strings; GP-competitive | https://arxiv.org/abs/2410.10190 |
| 2024 | PEBOL: BO with LLM-Based Acquisition Functions for NL Preference Elicitation | Conversational recommenders | Use NLI + LLM-guided TS/UCB queries in BO | Better MRR after few rounds vs monolithic LLM baselines | https://arxiv.org/abs/2405.00981 |
| 2024 | LLANA: LLM-Enhanced BO for Analog Layout Constraint Generation | Analog circuit synthesis | LLM generates design-dependent constraints that guide BO | Faster exploration; SOTA-level performance | https://arxiv.org/abs/2406.05250 |
| 2023 | Bayesian Approach for Prompt Optimization in Pre-trained LMs | Prompt tuning | Relax discrete prompts to embeddings; run BO (BoTorch) | Finds hard prompts w/o LM changes; analyzes accuracy/time trade-offs | https://arxiv.org/abs/2312.00471 |
| 2023 | BoChemian: LLM Embeddings for BO of Chemical Reactions | Chemistry | Map text procedures via LLM embeddings; optimize with BO | Open-source LLMs yield effective reaction features for BO | https://neurips.cc/virtual/2023/78776 |
Feel free to open a PR to add: Bayesian Optimization for Instruction Generation (BOInG), Optimal RoPE Extension via BO, BOPRO, HOLLM, ADO-LLM, LEDRO, HbBoPs, Bilevel-BO-SWA, Multi-task BO with LLM inits, model fusion via BO, and more.
-
Reasoning BO / LLM research assistants for BO.
Works like Reasoning BO integrate long-context reasoning, multi-agent setups, and knowledge graphs to steer sampling; similar ideas appear in LLM assistants that warm-start BO and supply interpretable commentary. See sources for concrete gains and case studies. -
Analog design (LLANA et al.).
Analog layout/sizing benefits from LLMs generating constraints or candidate points that focus BO, cutting exploration cost while hitting competitive figures of merit. -
Prompt & instruction optimization.
Papers relax discrete prompts to embeddings or use stochastic mini-batch surrogates, allowing BO to search combinatorial prompt spaces without backprop access to the LM. -
Chemical reaction optimization.
LLM embeddings and elicited priors provide strong features/priors for BO, improving yields and sample efficiency in reaction optimization campaigns. -
Meta-optimization & algorithm generation.
LLaMEA-BO shows LLMs can design BO algorithms themselves; other works learn to initialize or fuse models, creating a virtuous cycle where LLMs and BO co-improve.
- Add a row to the table (keep it concise).
- Include a short 1–2 line summary and a public source link (arXiv/OpenReview/project page).
- PRs welcome for new categories (e.g., robotics, materials, compiler/hardware).
Template row:
| YEAR | **Paper title** | Domain | LLM + BO strategy | Key highlight | https://link |