Cambridge machine learning researcher working on LLM confidence calibration and mechanistic interpretability
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.
Research profile
Output—
Publication counts and output volume are not available in this snapshot, so this axis is unscored. It would measure the number and distribution of the researcher's publications over time and across venues.
Impact—
Citation-based impact indicators such as citation counts or h-index are missing from the current data, so no score is given. This axis would measure the influence of the researcher's work on the wider literature.
Focus0/100
The focus score is 0, grounded in an evenly split field distribution: 0.5 mechanistic interpretability and 0.5 LLM confidence calibration. Rather than concentrating on a single topic, the reviewed profile divides attention between two adjacent machine-learning subfields — calibration of post-trained LLMs and theory-to-practice interpretability of neural networks — as reflected in the two 2026 papers.
Collaboration41/100
The collaboration score is 41, based on 7 distinct coauthors recorded in the snapshot. The recurring collaborator group — Dai Shi, Xiaoyu Li, Andi Han, Linhan Luo, Lequan Lin, Feng Chen, and Junbin Gao — suggests a small, stable working cluster. Across the two reviewed papers the researcher appears as a middle author on SupportCal and as last author on the superposition survey, in neither case as corresponding author.
Activity—
Recent-year publication cadence cannot be computed from the available snapshot, so this axis is unscored. It would measure how consistently the researcher has produced new work in recent years.
Reproducibility38/100
The reproducibility score is 38, a mean (reproMean 38) computed over 2 papers (reproPapers 2). The two contributions sit at different readiness levels: SupportCal is a concrete method with reported ECE gains, rated medium for reproduction, though base-model baselines are untested; the superposition survey is rated low because it adds no new experiments. Together they yield a modest reproducibility profile.
Research summary
José Miguel Hernández-Lobato is a machine learning researcher at the University of Cambridge whose Evolve-reviewed work spans two adjacent subfields: confidence calibration of post-trained large language models and mechanistic interpretability. The bibliometric snapshot splits his field distribution evenly (0.5/0.5) between these two areas, and both reviewed papers appeared in 2026 on arXiv. On the calibration side, SupportCal (2026) introduces a label-free temperature-scaling method that weights disagreement examples by pretrained-language-model support and reports consistent expected calibration error (ECE) gains over agreement-only baselines, though base-model baselines remain untested. On the interpretability side, he last-authored "Feature Superposition in Neural Networks: From Theory to Practice" (2026), a careful survey bridging superposition theory and interpretability practice with a rigorous L1–L6 claim taxonomy and honest hedging, but no new experiments. The collaboration axis scores 41 with 7 distinct coauthors, anchored by a recurring group including Dai Shi, Xiaoyu Li, Andi Han, Linhan Luo, Lequan Lin, Feng Chen, and Junbin Gao. Reproducibility scores 38 over 2 papers, consistent with one medium-readiness method paper and one low-readiness survey. Output, impact, and activity scores are not yet computed; this profile should be read as a focused, early-coverage view of the researcher rather than a comprehensive record.
Collaboration pattern
The snapshot records 7 distinct coauthors (collaboration score 41), with a recurring cluster of Dai Shi, Xiaoyu Li, Andi Han, Linhan Luo, Lequan Lin, Feng Chen, and Junbin Gao. Across the two reviewed papers he appears as a middle author on SupportCal and as last author on the superposition survey; neither lists him as corresponding author.
Research keywords
Highlights
SupportCal: label-free temperature calibration via PLM support
Open question: base-model baselines left untested
A rigorous theory-to-practice survey of feature superposition
Timeline
- 2026SupportCal posted on arXiv: label-free temperature calibration for post-trained LLMs, weighting disagreement examples by PLM support.
- 2026Last-authored "Feature Superposition in Neural Networks: From Theory to Practice", a survey with a rigorous L1–L6 claim taxonomy.
- 2026Reviewed output splits evenly between LLM confidence calibration and mechanistic interpretability (field distribution 0.5/0.5).
Papers in the Evolve catalog
Frequent co-authors
About this profile
This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.
- Bibliometric scores (collaboration, focus, reproducibility) are computed from a small public bibliographic snapshot; output, impact and activity data are missing entirely.
- Only two Evolve-reviewed papers (both 2026, arXiv) underpin the qualitative assessment; this is an early-coverage profile.
- ORCID and country are not provided; the profile is name-resolved from public bibliographic data and may merge or split identities.
- The reproducibility score of 38 is a mean over just 2 papers and should not be read as a stable estimate.