Rebranding BO away from "black-box" and towards "model-based"

Source

In a recent blog post (link) I described my "model-centric" view of Bayesian optimization (BO), essentially arguing that the model is the most important component of BO and BO users (and researchers) should do more to get it right. Under the assumption that the reader broadly agrees with the content of that post (or at least thinks this view of BO is one of several valid views), here I want to argue that the BO community should re-brand itself more towards model-based optimization and away from black-box optimization.

What is a "rebrand"?

BO is not a company or specific product, and therefore does not officially have any kind of brand. However, most BO papers and textbooks mention similar points in their introductions and descriptions of BO: black-box functions, use of uncertainty, exploration-exploitation trade-off, Bayesian models (especially Gaussian processes). Since papers and textbooks are some of the primary ways that people learn about BO, collectively I think they do form a kind of "brand". If I were to describe this brand as an elevator pitch of a company, I'd say:

You give us an expensive black-box function, we will efficiently optimize it using a Gaussian process surrogate model that precisely balances exploration and exploitation.

I bet most BO researchers wouldn't fully agree with this statement. However, from talking with a lot of non-BO experts at conferences, this is how I think the typical person in ML would describe the value proposition of BO.

Therefore, in my opinion a "re-brand" is changing this brand: doing something to make the typical BO user or ML researcher view BO differently. Most likely this "something" will be describing BO differently in papers, but I'll come back to "how" to do the rebrand later. Let's first focus on what.

What brand should BO have?

My elevator pitch for BO's value proposition would be:

You give us a function and your model of the function, we turn this model into an interpretable sequential decision algorithm tailored to the constraints and trade-offs of your particular problem.

Compared to the current brand, this emphasizes:

  • Model as an input.
  • The value add being producing a decision algorithm from the model.
  • Customization of the algorithm for particular constraints and trade-offs.

It deemphasizes:

  • Black-box nature of the function (still true, but BO is not exclusively black box).
  • Uncertainty (this comes in naturally as part of the model).
  • Exploration/exploitation (I think this is one of many trade-offs you can customize).

Why should it have this brand?

It's 2026 and everybody's attention is focused on the amazing advances of LLMs and agents. Implicitly everybody is asking the questions

  1. Should I use an LLM agent to do this task?
  2. Should I replace my existing solution with an LLM agent?

I don't believe BO is obsolete, but there are parts of BO which don't work well and I believe can be augmented or replaced by LLMs. Essentially my suggested rebrand is to emphasize the use cases where I think BO will be most relevant (ie not replaced by LLMs) in the next 5-10 years.

Let's start by focusing on the positives of LLMs and how they might disrupt BO. First, LLMs are amazing at multi-modality, meaning that any black box optimization problem whose inputs/outputs can be represented as tokens can easily be fed to an LLM. BO used to stand out by making fewer assumptions than other optimization algorithms (eg not requiring gradients). Unfortunately, I think BO can no longer compete on "breadth": LLMs are the clear winner here. LLMs are also clearly easier to use than BO (at least on small datasets): you can put all data into the LLM's context window, describe the problem, and ask for a decision. BO is difficult to set up and I don't see that changing. Clearly BO's value needs to come from other areas (eg "quality" or "reliability").

Intuitively I would expect BO to perform better than LLMs, but there are many situations where LLMs will probably perform better. Especially on problems with tiny amounts of data (eg < 10 data points), I bet LLMs might actually be better-calibrated surrogate models than GPs. A short textual description of what the output variable actually is means the LLMs can leverage pre-training knowledge to suggest values within a reasonable range, and some basic knowledge about what influences a property. For example, LLMs probably won't suggest a house price of $1, a binding affinity of -100 kcal/mol, or a learning rate of 1e-10. The same isn't true for GPs: on small datasets the learned mean and outputscale can be severely misspecified.

LLMs can also capture user preferences much more easily than standard BO. BO algorithms are controlled by a bunch of semi-interpretable hyperparameters, tuning of which is often more of an art than a science. LLMs can be given specific behavioural instructions in text form. You can tell the LLM to "explore" or "exploit" more directly, whereas in BO this is indirectly controlled by a bunch of hyperparameters. Ultimately I expect BO to make better long horizon decisions than LLMs (since its explore/exploit trade-off is governed by theory and LLMs have limited context lengths), but over short horizons LLMs might actually be better.

That being said, a huge disadvantage for LLMs is their lack of interpretability (which is even worse for closed-weight models). Although the outputs usually seem reasonable, they can be completely "hallucinated" and appear for the wrong reasons. Reasoning models and RAG models are in some ways attempts to work around this and ground LLM responses in something more concrete, but as far as I am aware there is still evidence that they can be "gamed": models can output an answer that does not match the reasoning trace, and they can also ignore context retrieved as part of a RAG process. This means LLM-based systems might appear to be following user instructions closely, while actually not following them at all.

In contrast, basically all versions of BO guarantee that the decision will be supported by the outputs of the surrogate model. This is how I believe BO will continue to provide value into the future. For applications where one has a trusted predictive model and wants to use it to make decisions in a precise way, LLMs introduce uncertainty and unpredictability, while BO can be set up to behave exactly as implied by the predictive model.

Therefore, it feels appropriate for me for BO's "brand" to align more closely to this core value proposition (unlike the current black-box brand which obscures this advantage).

Why think about this now?

The strength of the BO research community depends on people outside the community seeing value in BO. Most BO research is supported by government research grants or tech companies supporting an in-house BO team. A strong value proposition helps convince these people to keep funding and supporting research in this area.

BO has not yet been "eaten" by LLMs. But it is possible that LLMs might be able to do as well as BO on standard black-box benchmarks within a couple of years. If this happens, funders who previously funded BO research because they expected it to yield efficient general-purpose black-box optimizers may start to lose faith that this is the correct approach. That might inevitably force the BO community to adopt a different "brand", at least in grants and internal presentations. My argument is simply that it would be better to do this proactively. Rebranding takes time, and emphasizing a different value proposition after the research has been perceived as "not the right approach" might come across as opportunistic and dishonest.

Even if the scenario above (LLMs beating BO at black-box optimization benchmarks) does not occur, I still don't think much is lost from doing this kind of rebrand. The key points are arguably still true, and the new branding does not in any way deny BO's effectiveness as a black-box optimizer (just as the current branding does not deny BO's effectiveness at turning models into decision algorithms).

How would the "rebrand" actually be done?

I don't have a complete answer to this, but the BO community is small, so I bet a lot of changes could be achieved if BO researchers:

  1. Described BO in a different way in their papers and open source software.
  2. Wrote BO software to be more accommodating of custom models.
  3. Clearly articulated this benefit of BO (perhaps via blogs like this one).

Summary

In this post I explained why I think BO has the brand "black-box optimizer" as a result of its description across many papers, why I'd prefer a brand more like "turn models into decision algorithms", why I think this is more aligned with BO's long-term value proposition over LLMs, and why we should do this proactively by describing BO differently now.

If you disagree with my argument, I encourage you to reflect on what you think BO's brand should be. Arguably the current brand was not an intentional choice, and may not line up perfectly with how you see BO's value. At the very least, I think we should try to describe methods in a way that emphasizes their strengths and what problems they are suitable for, rather than just describe methods in the same way as all previous papers!