[FEATURE REQUEST]: New High-Dimensional BO Algorithm from ICLR 2026

[rosenyu304]

Motivation

Hi, my paper "GIT-BO: High-Dimensional Bayesian Optimization with Tabular Foundation Models" has been accepted to ICLR 2026 (https://openreview.net/forum?id=9iTdKS4SRQ), and I am wondering if I can contribute it to BoTorch's library.

Describe the solution you'd like to see implemented in BoTorch.

There are two parts in this paper that can be implemented in BoTorch:

1. The dimension reduction techniques (GI-subspace) using the gradient from the surrogate. The surrogate can either be a GP or a TabPFN (Appendix C).
2. The TabPFN wrapper for the latest TabPFN-2.5. I noticed that Issue [FEATURE REQUEST]: TabPFN2 Wrapper #2677 brought this topic up, but I couldn't find this feature in BoTorch. I have the code ready to contribute both the GI-subspace with GP and TabPFN. If we implement more on TabPFN, I also have BAxUS's dimensional reduction implemented using TabPFN as a surrogate, and its performance is very competitive (Appendix B.4).
   I am now envisioning these features implemented under the "Bite-Sized Tutorials," but would like to know what the team thinks.

Describe any alternatives you've considered to the above solution.

No response

Is this related to an existing issue in BoTorch or another repository? If so please include links to those Issues here.

No response

Pull Request

Yes

Code of Conduct

• I agree to follow BoTorch's Code of Conduct

[saitcakmak]

Hi @rosenyu304. We welcome contributions of recent methods from literature under botorch_community (see https://github.com/meta-pytorch/botorch/blob/main/CONTRIBUTING.md#community-contributions-beta). The notebooks are similarly contributed under notebooks_community and are displayed on our website here.

I know we have some PFN support implemented in botorch_community (link) but I am not familiar with the details and whether the support extends to the TabPFN models. Based on the open issue, it seems that it is not currently supported. Let's continue the discussion about this there.

[Balandat]

+1 we'd be excited to include this in botorch_community (and potentially graduate it into the main library if it works well and there is demand from users for it)

[rosenyu304]

Thanks for the reply! I guess I can start working on a pull request that goes under the botorch_community for the GP version's code.

For the PFN support, I see that here BoTorch loads the PFN4BO version of PFN, not the latest PFN model. The only solution I see for using the latest model is to download the latest tabpfn, place it locally, and write a wrapper around it (at least that is how I did it for my experiments). Would this part be feasible for the same pull request to botorch_community?

[Balandat]

@rosenyu304 I think that may be the easiest setup, yes. It would also simplify some of the licensing questions in the sense that we just provide the adapter and what kinds of TabPFN weights are being used (non-commercial, Prior Labs License, enterprise) is up to the user.