HapBO: Harnessed $\alpha$-$\pi$ Bayesian Optimization via Expert Natural Language Priors
HapBO is designed to make expert knowledge useful to Bayesian optimization even when that knowledge is written as ordinary language. Experts may have useful intuitions or documentation, but traditional methods often need prior knowledge in an explicit probability form. HapBO offers a way to bridge that gap. A human expert, an LLM, or both can provide beliefs once. HapBO deterministically compiles a structured specification into a reusable belief function, written as π: X → [0,1]. The function can use a rich representation that is not limited to a Gaussian family. During optimization, HapBO applies it as a bounded positive multiplier on a nonnegative acquisition function. The positive multiplier preserves the acquisition function’s positive support, so the belief does not eliminate candidates that otherwise have positive acquisition. HapBO makes no LLM calls during optimization. The authors report results across 40 HPOBench tasks, synthetic problems, and two real-world applications, with zero in-loop token cost.
What is HapBO designed to do with expert knowledge expressed in natural language?
HapBO is designed to make expert knowledge useful to Bayesian optimization even when that knowledge is written as ordinary language. Experts may have useful intuitions or documentation, but traditional methods often need prior knowledge in an explicit probability form. HapBO offers a way to bridge that gap.
A human expert, an LLM, or both can provide beliefs once. HapBO deterministically compiles a structured specification into a reusable belief function, written as π: X → [0,1]. The function can use a rich representation that is not limited to a Gaussian family. During optimization, HapBO applies it as a bounded positive multiplier on a nonnegative acquisition function.
The positive multiplier preserves the acquisition function’s positive support, so the belief does not eliminate candidates that otherwise have positive acquisition. HapBO makes no LLM calls during optimization. The authors report results across 40 HPOBench tasks, synthetic problems, and two real-world applications, with zero in-loop token cost.
Why is it difficult for traditional prior-guided Bayesian optimization to use expert advice written in ordinary language?
Expert advice often arrives as intuitions, explanations, or documentation in natural language. Traditional prior-guided Bayesian optimization can be difficult to use with that material because the methods may require an explicit probability distribution describing the problem’s landscape. A sentence of advice is not already that distribution.
This means useful knowledge can be difficult to encode before the search begins. The challenge is not just translating words into numbers; the prior must express beliefs about the search space in a form the optimization method can use. The article identifies both the need for an expressive language-to-prior mapping and the need to handle advice that may be misleading.
HapBO addresses this gap by eliciting beliefs once from a human expert, an LLM, or both, then compiling a structured specification into a reusable belief function. That function is not restricted to a Gaussian family. It guides optimization through a bounded positive multiplier, and the method requires no LLM calls during the search itself.
How does HapBO turn an expert's description into a reusable belief function, and when does it need an LLM?
HapBO first elicits beliefs from a human expert, a large language model, or both. The aim is to capture knowledge that may otherwise exist only as expert intuition or documentation, rather than as a probability distribution ready for a prior-guided optimizer. The elicitation happens once, before optimization proceeds.
Next, HapBO deterministically compiles a structured specification into a reusable belief function, π: X → [0,1]. Its representation is rich and is not restricted to a Gaussian family. During optimization, the belief function modifies a nonnegative acquisition function through a bounded positive multiplier. This lets the belief guide candidate selection while preserving positive acquisition support.
An LLM is optional: a human expert can provide the beliefs, or the two can contribute together. If an LLM is used, the paper describes its role in the initial elicitation, not in repeated optimization steps. HapBO therefore makes no LLM calls during optimization, which the authors report as zero in-loop token cost.
How large and varied were HapBO's tests, from benchmark tasks and synthetic dimensions to real-world applications and query budgets?
The evaluation covered several kinds of optimization problems rather than a single benchmark. HapBO was tested on 40 HPOBench tasks and on a synthetic suite with dimensions ranging from 6 to 15. These tests let the researchers examine its behavior across benchmark and constructed problems.
The evaluation also included two real-world applications: cortical neurostimulation and constrained solar design. The article reports that HapBO improved the optimization outcome in these applications even when budgets were only 12–20 queries. That is a small number of opportunities to evaluate candidate settings, so performance in this range is a central practical result reported in the abstract.
Taken together, the tests span benchmark tasks, synthetic spaces of varying dimensionality, and applied problems. The abstract gives the number of HPOBench tasks, the synthetic dimension range, and the two application areas, but it does not provide further details here about the individual tasks or the exact query allocation for each one.
What happened to optimization performance and costs when HapBO was tested against other methods?
The reported results combine optimization quality with computational and language-model costs. On 40 HPOBench tasks, HapBO achieved the lowest aggregate regret among the evaluated approaches. The abstract also describes its performance on real-world tasks as strong, without giving a numerical comparison for those applications.
For cortical neurostimulation and constrained solar design, HapBO improved the optimization outcome at budgets of 12–20 queries. The article reports that it did so at a small fraction of the wall-clock and token costs of other LLM-based methods. It also states that HapBO incurs zero in-loop token cost because no LLM calls are made during optimization.
These findings suggest a practical combination: language-derived beliefs can help guide a search, while the method avoids repeated LLM use as candidates are selected. The abstract does not provide exact regret values, cost ratios, or a full account of every comparison, so the reported advantage should be understood in the terms and evaluation settings it gives.
How can HapBO use a belief that may be misleading without completely ruling out any candidate that the acquisition function would consider?
A belief based on expert language may be helpful, but it can also be misleading. If a method treated a doubtful belief as an absolute rule, it could exclude a candidate that the optimization process would otherwise consider. The abstract highlights robustness to misleading information as an important design goal.
HapBO applies its belief function through a bounded positive multiplier on any nonnegative acquisition function. Because the multiplier is positive and bounded, it can adjust the acquisition values without turning a positive value into zero. In the paper’s terminology, this preserves positive acquisition support: candidates with positive acquisition remain supported after applying the belief.
The authors prove that, for Expected Improvement, HapBO preserves the standard simple-regret rate up to a bounded factor, even under adversarial beliefs. This is a theoretical guarantee about the method’s behavior, not a claim that every prior belief will be accurate. It provides a safeguard against a belief completely shutting off promising options.
What is Bayesian optimization, and how do acquisition functions such as Expected Improvement help choose the next experiment?
In general, Bayesian optimization is used to optimize a function when evaluating a candidate can be costly. It builds a model from evaluations so far, then uses that model to decide where to evaluate next. This explanation is general background; the supplied abstract focuses on HapBO rather than giving a full definition of Bayesian optimization.
An acquisition function turns the model’s information into a score for possible next candidates. Expected Improvement, or EI, estimates how much a candidate is expected to improve on the best result observed so far. The optimizer can use those scores to choose a next experiment, balancing the goal of finding better results with uncertainty in the model.
The article specifically uses EI in its theoretical result: it proves that HapBO preserves the standard simple-regret rate up to a bounded factor, even under adversarial beliefs. HapBO modifies a nonnegative acquisition through a bounded positive multiplier, so its belief can guide the choice without removing positive acquisition support.
This brief was written by AI from the original reporting and checked by other models. Names, figures and quotes come from the source; read it for full context.
Read more in the JupiteX app
Pulse is free. New stories every 4 hours, each one broken into the questions that explain it.
Or read more news on the web