Cambridge ML researcher bridging feature-superposition interpretability and label-free LLM confidence calibration
2
papers here
2
reviewed by Evolve
1
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—
Insufficient data: Evolve has no verified publication-count or productivity metrics for this researcher yet. This axis would measure total and per-year output across venues, including preprints and peer-reviewed papers, once a fuller bibliographic record is linked.
Impact—
No citation counts, h-index, or influence signals are available in the current snapshot. This axis would measure citations, downstream uptake, and reuse of the two reviewed papers, so no impact score can be reported or estimated here.
Focus0/100
The focus score is 0, reflecting an even 50/50 split in the reviewed record between machine learning (mechanistic interpretability) and machine learning (LLM confidence calibration). In substance the two threads are adjacent — both concern how neural language models can be understood and made more reliable — but the bibliometric field distribution shows no single dominant field yet, so concentration cannot be scored favorably on the reviewed sample.
Collaboration41/100
The collaboration score is 41 on the 0–100 radar, based on 7 distinct coauthors in the reviewed record. The frequent-coauthor list includes José Miguel Hernández-Lobato, Xiaoyu Li, Andi Han, Linhan Luo, Lequan Lin, Feng Chen, and Junbin Gao. Authorship positions differ by project: first author on the feature-superposition survey and middle author on SupportCal, suggesting a mix of lead-conceptual writing and team-based calibration experiments with a stable, moderately sized circle.
Activity—
Insufficient data: no multi-year publication cadence or venue timeline is available beyond two 2026 preprints. This axis would measure publishing frequency and recency over time once a longer bibliographic history is linked.
Reproducibility38/100
Reproducibility scores 38 (reproMean over 2 papers). Evolve rates SupportCal as medium readiness: the method is concrete — label-free temperature calibration weighted by PLM support with consistent ECE gains over agreement-only baselines — but base-model baselines were untested. The first-authored superposition survey is rated low readiness: a rigorous L1–L6 claim taxonomy with honest hedging, yet no new experiments to replicate.
Research summary
Dai Shi is a machine learning researcher at the University of Cambridge whose Evolve-reviewed work spans two adjacent threads: mechanistic interpretability of neural networks and confidence calibration of large language models. The catalog contains two 2026 arXiv papers. The first-authored 'Feature Superposition in Neural Networks: From Theory to Practice' is a careful survey bridging superposition theory and interpretability practice, organized around a rigorous L1–L6 claim taxonomy with honest hedging, though it adds no new experiments. The middle-authored 'SupportCal' introduces label-free temperature calibration of post-trained LLMs, weighting disagreement examples by reference-model (PLM) support; Evolve reports consistent ECE gains over agreement-only baselines while noting that base-model baselines were untested. Bibliometric signals mirror this split: the reviewed record divides evenly (0.5/0.5) across the two fields, with 7 distinct coauthors including frequent collaborators José Miguel Hernández-Lobato, Xiaoyu Li, Andi Han, and Junbin Gao. Radar scores are moderate — collaboration 41 and reproducibility 38, the latter averaged over 2 papers — while output, impact, and activity axes lack sufficient data to score. Overall this is an early-stage profile built from a narrow, very recent slice of work, best read alongside its caveats.
Collaboration pattern
With 7 distinct coauthors and a collaboration score of 41, Dai Shi's reviewed work shows a compact, stable circle rather than a broad network. Recurring collaborators include José Miguel Hernández-Lobato, Xiaoyu Li, Andi Han, Linhan Luo, Lequan Lin, Feng Chen, and Junbin Gao. The pattern pairs first-author conceptual synthesis (the superposition survey) with middle-author team experiments (SupportCal), consistent with small-group collaboration at Cambridge spanning interpretability and calibration.
Research keywords
Highlights
First-authored survey organizes feature superposition, theory to practice
SupportCal: label-free calibration of post-trained LLMs
Timeline
- 2026First-authored 'Feature Superposition in Neural Networks: From Theory to Practice' released on arXiv, organizing superposition research with an L1–L6 claim taxonomy.
- 2026Middle-authored 'SupportCal' released on arXiv, proposing label-free temperature calibration of post-trained LLMs via reference support and corroboration.
- 2026Both cataloged papers land in 2026, splitting the reviewed record evenly between mechanistic interpretability and LLM confidence calibration.
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.
- Radar scores are computed from a limited OpenAlex-based bibliometric snapshot; the output, impact, and activity axes have insufficient data and are not scored.
- The profile rests on only two Evolve-reviewed 2026 arXiv preprints, not a complete or peer-reviewed publication record.
- 'Dai Shi' is a common name; this profile is name-resolved from public bibliographic data, and no ORCID is linked to confirm identity.
- Country and affiliation details beyond 'University of Cambridge' are unverified.