LLM mechanistic interpretability researcher; first author of a four-stage decomposition of math word-problem reasoning
1
papers here
1
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-09-20) 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—
Output cannot be scored: this snapshot holds a single reviewed arXiv paper with no aggregated publication counts. The axis would measure yearly output volume, venue mix, and growth across the career record once fuller bibliometric data accumulates.
Impact—
Impact cannot be scored: no citation counts or influence signals accompany the 2026 preprint, which is too recent to have accrued citations. The axis would measure citation uptake, downstream reuse, and scholarly reach over time.
Focus100/100
Focus scores a perfect 100 because every catalogued output sits in one field: the field distribution lists exactly one entry, machine learning (LLM mechanistic interpretability), with count 1 of 1 papers. The 2026 study is squarely on-theme, applying activation patching, linear probing, and direct logit attribution to GSM8K word problems — a deliberately specialized, early-stage research identity.
Collaboration14/100
The collaboration score of 14 reflects a very narrow footprint: the snapshot records exactly 1 distinct co-author — Carla P. Gomes — on the 2026 first-author paper. Concentrated two-person authorship is plausible for early mechanistic-interpretability work, but it limits independent corroboration; a broader co-author network should emerge as more of Qu's papers enter the catalog.
Activity—
Activity cannot be scored from a single paper year (2026); the snapshot lacks a multi-year publication stream. The axis would measure publication cadence and research continuity across recent years once more records accumulate.
Reproducibility100/100
Reproducibility scores 100 (reproMean 100 across 1 reproducibility-assessed paper), matching Evolve's 'high' reproducibility readiness for the 2026 study. The review cites strong, well-controlled multi-method evidence spanning activation patching, linear probing, and direct logit attribution, while flagging interpretive leaps and model-specific caveats — so the perfect score rests on a single-paper sample and should be read as early but encouraging.
Research summary
Zhongdi Qu is an early-career researcher working in machine learning, specifically large language model (LLM) mechanistic interpretability. This profile is name-resolved from public bibliographic data (ORCID 0009-0002-7789-712X) and rests on a single reviewed paper currently in the Evolve catalog, so it should be read as an early-stage profile rather than a mature career record. The reviewed work is the 2026 arXiv preprint "A Four-Stage Decomposition of Word-Problem Solving and Mechanistic Fragility in LLM Math Reasoning", on which Qu is first author (non-corresponding), co-authored with Carla P. Gomes. Evolve's review characterizes the paper as claiming a four-stage internal pipeline for how LLMs solve word problems and as localizing distractor fragility to "Operation Planning" heads, supported by strong, well-controlled multi-method evidence — activation patching, linear probing, and direct logit attribution — while noting interpretive leaps and model-specific caveats. Bibliometrically, the snapshot shows a highly focused output profile (one field in the distribution), maximal reproducibility readiness (reproMean 100 across 1 paper), and a narrow collaboration footprint (1 distinct co-author). Output volume, citation impact, and multi-year activity cannot yet be scored for lack of data.
Collaboration pattern
The snapshot records exactly 1 distinct co-author — Carla P. Gomes — on Zhongdi Qu's 2026 first-author paper, which the low collaboration score (14) reflects. The current pattern is a concentrated two-person collaboration rather than a broad network. As Qu's catalog grows, additional co-authors and cross-group ties will be needed to independently corroborate the mechanistic-interpretability findings and to reveal how collaborative this research practice truly is.
Research keywords
Highlights
Four-stage internal pipeline for LLM word-problem solving
Distractor fragility localized to Operation Planning heads
Multi-method evidence earns high reproducibility readiness
Timeline
- 2026Debuts in the Evolve catalog as first author of an arXiv study decomposing LLM word-problem solving into four internal stages and tracing mechanistic fragility.
- 2026Evolve's review finds strong, well-controlled multi-method evidence and grades reproducibility readiness as high (reproMean 100 across 1 paper).
- 2026Catalog debut records Carla P. Gomes as the only distinct co-author (distinctCoauthors = 1), marking a concentrated two-person collaboration.
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 come from a thin bibliometric snapshot: output, impact, and activity are unscored, and the scored axes rest on a single paper.
- Institution and country are absent from the input data and are omitted rather than inferred.
- Only the 2026 arXiv paper in the Evolve catalog is reviewed here; other work by this author may exist outside the catalog.
- No citation, h-index, or venue-quality data is available; the 2026 preprint is too recent for meaningful citation accumulation.