A preprint posted on November 9, 2026 tests TabPFN and TabICL as surrogate models for Bayesian optimization in materials discovery. Klemens Flöge and co-authors report competitive results on two benchmarks, while warning that retrospective performance does not establish superiority in live laboratory campaigns. The arXiv record lists a revision on November 11.
What the materials optimization benchmarks measured
The authors evaluated 15 recorded campaigns from the Liang and Olympus benchmarks, comparing TabPFN and TabICL with Gaussian processes, random forests, CatBoost and random search. Their study reports that a TabPFN configuration ranked first on each benchmark. Rankings changed with the acquisition rule, which selects the next candidate, and the metric used to score progress. Neither foundation model received task-specific optimization training.
These were replays of finite candidate pools whose outcomes had already been measured. The authors caution that the pools reflect earlier experimental selection policies and that results depend on evaluation protocol. The finding therefore supports further investigation, without demonstrating faster discovery in a prospective experiment.
Why the model version matters for reproduction
The authors’ benchmark repository explicitly selects TabPFN-3 and pins library versions in its lockfile. By contrast, the current TabPFN documentation identifies TabPFN-3.5 as the default. For a reproduction attempt, installing the package and accepting its current default would change the model choice. The benchmark’s explicit checkpoint selection is therefore a material detail.
The TabICL documentation provides a regression interface and distinguishes default in-context learning from optional fine-tuning. Researchers assessing this result should preserve that distinction: adapting weights to their own dataset would constitute a different evaluation. Recording the checkpoint and adaptation settings belongs in a reproducible AI experiment.
What researchers and research offices can assess next
The repository contains reference traces recording candidate selections across campaigns and seeds, together with scripts that regenerate figures from those traces without fitting models. This gives university research software teams a concrete starting point: examine the recorded optimization paths before budgeting a full rerun. The repository also documents CPU/GPU differences that can change which candidate is selected.
Our interpretation is that a local pilot should compare model-and-acquisition pairs against the laboratory’s existing approach at its intended experimental budget. Research offices evaluating a proposal should separate reproducing archived benchmark results from validating a prospective laboratory workflow. Teams should also record the cited preprint version, software environment and model checkpoint, so that later revisions do not blur the evidence behind an adoption decision.
Sources
- Flöge et al., arXiv submission record. Checked 2026-11-15.
- Out-of-the-Box Tabular Foundation Models for Bayesian Optimization in Materials Discovery, version 2. Checked 2026-11-15.
- Prior Labs, matsci-bo-bench code and reference results. Checked 2026-11-15.
- Prior Labs, TabPFN documentation. Checked 2026-11-15.
- SODA, TabICL documentation. Checked 2026-11-15.







