GOLLuM: Gaussian Process Optimized LLMs – Reframing LLMs as Principled Bayesian Optimizers 🧙‍♂️📈

GOLLuM – Gaussian Process Optimized LLMs are here!
One representation to rule them all!

📄 Paper:

---

🔍 Overview

🎯 GOLLuM addresses the challenge of harnessing LLMs for optimization under uncertainty by introducing:

• LLM-based deep kernels, jointly optimized with GPs to preserve the benefits of both
• LLMs to provide a rich and flexible input space for Bayesian optimization
• GPs to model this space with predictive uncertainty for more efficient sampling

🌌 The framework enables a bidirectional feedback loop:

1. The GP guides updates to LLM weights to produce more effective embeddings
2. These embeddings enhance the GP's probabilistic modeling

---

🧠 Key Features

• Unified Representation Learning: Uses textual templates to represent heterogeneous parameter types (categorical, numerical, structural)
• GP-Guided LLM Finetuning: Optimizes LLM embeddings through GP marginal likelihood
• Implicit Contrastive Learning: Automatically organizes the latent space into distinct performance regions
• Chemical reasoning in the latent space: Uncovering chemical patterns under extreme low-data regimes
• Architecture Agnostic: Works with various LLM architectures (encoder, decoder, encoder-decoder)
• Domain Agnostic: No requirement for domain-specialized models or pretraining

---

🚀 Quickstart

📦 Project Dependencies

You can install the environment with uv:

uv sync
uv pip install rxnfp --no-deps

This will create a .venv and install the required packages For manual setup or more details, see DEPENDENCIES.md.

---

🛠 Install GOLLuM in editable mode

uv pip install -e .

---

⚙️ Running Experiments

All configuration files for reproducing experiments are included in the configs/ directory. You can launch an experiment with:

python train.py --config=configs/pllm_phi.yaml

Replace pllm_phi.yaml with other config files for variants such as llm_phi.yaml, pllm.yaml, etc.

---

📚 Citation

@inproceedings{
rankovic2025gollum,
title={{GOLL}uM: Gaussian Process Optimized {LLM}s {\textemdash} Reframing {LLM} Finetuning through Bayesian Optimization},
author={Bojana Rankovi{\'c} and Philippe Schwaller},
booktitle={ICLR 2025 Workshop on World Models: Understanding, Modelling and Scaling},
year={2025},
url={https://openreview.net/forum?id=2ORViHAUbf}
}

---

⚖️ License

This project is licensed under the Apache 2.0 License. See the LICENSE file for details.

---

🤝 Acknowledgements

This work was supported by NCCR Catalysis (grant number 225147), a National Centre of Competence in Research funded by the Swiss National Science Foundation.