Frank Hutter

Research profile

Research summary

Frank Hutter (Microsoft, United States; ORCID 0000-0002-2037-3694) is profiled here from an OpenAlex-resolved bibliographic record. The snapshot indicates a long, high-volume publishing history: h-index 72, i10 index 166, works indexed from 2002 onward, and a steep recent climb — 27 works in each of 2018 and 2019, 34 in 2021, a record 39 in 2024, 37 in 2025, and 98 works across the last three years. Topic counts concentrate on Machine Learning and Data Classification (202 works) with substantial footprints in metaheuristic optimization (69), neural network applications (60), bandit algorithms (59), and few-shot learning (57). The two papers Evolve has reviewed, both 2026 arXiv releases, define a two-track current agenda that matches the record's 50/50 field distribution: Complex KDA extends Kimi Delta Attention's gate to [-1,1] and beta to [0,2] with strong expressivity theory for linear RNNs, while Attention Quantization reports FP8 attention kernels for tabular foundation models (TabPFN, TabICL) delivering up to 1.7x speedup at negligible accuracy loss. Collaboration is anchored by 14 distinct coauthors and a recurring core including Kübler, Hollmann, Siems, Zela, Grazzi, Jitsev, and Cevher; on both reviewed papers this researcher holds a middle or last rather than corresponding position. Output and impact radar axes remain unscored pending complete citation-level records.

Research keywords

Highlights

• Complex KDA: extending Kimi Delta Attention's gate to [-1,1] with single-transform rotation

  Evolve's review finds this 2026 arXiv paper extends KDA's gating to [-1,1] and beta to [0,2], enabling rotation in a single transform, with strong theoretical support for improved linear-RNN expressivity and state tracking; the caveat is that language-modeling gains over the KDA baseline remain limited.

• FP8 attention kernels push tabular foundation models to up to 1.7x speedup

  According to Evolve's verdict, this 2026 arXiv paper quantizes attention via Triton FP8 kernels for tabular foundation models such as TabPFN and TabICL, reporting up to 1.7x speedup with negligible accuracy loss; the review calls the study rigorous but limited to Blackwell GPUs, 3 seeds, and no stated code release.

• A two-track 2026 agenda mirroring a 50/50 field split

  The catalog's only two reviewed papers, both from 2026, split evenly between sequence modeling and expressivity theory and efficient inference for tabular deep learning, matching the record's 0.5/0.5 field distribution and showing simultaneous engagement with attention theory and attention systems engineering.

Timeline

1. 2002Earliest works appear in the OpenAlex record (2 indexed in 2002, rising to 8 by 2007).
2. 2014Output accelerates: 13 works indexed, more than double 2013's 6.
3. 2018High-output phase begins: 27 works in 2018, repeated in 2019.
4. 202134 works indexed, then a dip to 19 in 2022 and 17 in 2023 before rebounding.
5. 2024Record annual output: 39 works indexed, followed by 37 in 2025.
6. 2026Both Evolve-reviewed papers appear on arXiv — Complex KDA and Attention Quantization for Tabular Foundation Models; 22 works already indexed this year.

Papers in the Evolve catalog

• 2026-09-21

  Complex KDA: Understanding and Enhancing the Expressivity of Kimi Delta Attention

• 2026-09-11

  Attention Quantization for Tabular Foundation Models

Frequent co-authors