LLAMBO (sampler slot)

Cite: Liu et al., ICLR 2024¹.

Paper Code

Does not replace the GP. Use it to propose candidates from natural-language context and recent trials, then score them against lenz.

LLM configuration

Sampling uses Sara's LLM by default. Pass --sampler-llm-* or llambo set-llm only to use a different model:

lenz llambo set-llm --state ./state.json --provider openai --model gpt-4.1
lenz llambo warmstart --state ./state.json --n 5 --context ./context.md
lenz llambo sample --state ./state.json --n 8

Making it the default sampler

To make every suggest go through LLAMBO instead of BoTorch acquisition optimization:

lenz set-sampler --state ./state.json --sampler llambo

Strongest early in a run, when observations are scarce. After the GP is trustworthy, switch back to botorch.

---

1. Liu, Astorga, Seedat, van der Schaar. Large Language Models to Enhance Bayesian Optimization. ICLR 2024. ↩