Julia Hockenmaier

Research profile

Research summary

Julia Hockenmaier is a researcher at the University of Illinois Urbana-Champaign, profiled here from public bibliographic data resolved against ORCID 0000-0003-2796-1752. The Evolve-visible slice of her record is a recent, preprint-centered snapshot built on two 2026 arXiv papers. The first, "A Survey on the Linear Representation Hypothesis," places her in the last (senior) author position; Evolve's review credits it with arguing that the linear representation hypothesis is ill-defined unless pinned to a fixed model-layer-feature-dataset tuple, proposing a falsifiable Definition 4, and delivering a rigorous taxonomy spanning linear probing, sparse autoencoders, activation steering and causal abstraction — while noting it adds no new experiments, which matches its low reproducibility-readiness rating. The second, "MineCEraft: Evaluating Language Models as Construction Engineers in the World of Minecraft," has her in a middle author position and has not yet been reviewed, so it carries no field tag, keywords, verdict or slug. Radar scores reflect this thin but concentrated slice: focus is maximal at 100 (the field distribution lists a single field, machine learning interpretability / mechanistic interpretability), collaboration is 36 with 5 distinct co-authors recorded, and reproducibility is 20, computed as a mean of 20 over one assessed paper. Output, impact and activity axes remain unscored for lack of data. Frequent co-authors in the snapshot are Sewoong Lee, Marc E. Canby, Ikhyun Cho, Risham Sidhu and Yoonhwa Jung.

Research keywords

Highlights

• A falsifiable definition for the linear representation hypothesis

  As last author on this 2026 arXiv survey, Hockenmaier argues the linear representation hypothesis is ill-defined unless pinned to a fixed model-layer-feature-dataset tuple, and the paper proposes falsifiable Definition 4. Evolve's review praises the rigorous taxonomy across linear probing, sparse autoencoders, activation steering and causal abstraction, while noting it adds no new experiments — hence its low reproducibility-readiness rating.

• Benchmarking language models as Minecraft construction engineers

  In this 2026 arXiv paper, MineCEraft evaluates language models as construction engineers inside the Minecraft world, with Hockenmaier in a middle author position. It has not yet been through Evolve review, so no field tag, keywords, verdict or slug exist for it; the profile links it to her only through authorship on the record.

• A senior-position interpretability record inside a single-year snapshot

  Both Evolve-tracked papers fall in 2026, and the interpretability survey places her last among authors without a corresponding-author flag — consistent with a senior supervisory role. With only 5 distinct co-authors recorded and a collaboration score of 36, the visible network is compact, so this highlight reads as early visibility rather than an established multi-year record.

Timeline

1. 2026arXiv preprint "A Survey on the Linear Representation Hypothesis" appears with Hockenmaier in the last author position; Evolve review rates its reproducibility readiness low and records the machine learning interpretability field tag.
2. 2026Co-authors "MineCEraft: Evaluating Language Models as Construction Engineers in the World of Minecraft" in a middle author position; the record is not yet reviewed by Evolve.
3. 2026Both catalog records land in the same year, so the visible profile is a single-year, preprint-stage snapshot; output, impact and activity axes stay unscored for lack of longitudinal data.

Papers in the Evolve catalog

• 2026-09-19

  A Survey on the Linear Representation Hypothesis

Frequent co-authors