First-author NLP researcher at Oxford working on fast typed decision models and training away LLM sycophancy
2
papers here
2
reviewed by Evolve
2
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-09) 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 is not scored: the indexed record for this profile contains too few verified publications to measure productivity reliably. This axis would count total papers, venue distribution, and first/corresponding-author share across the researcher's full career.
Impact—
Impact is not scored: no citation counts, h-index, or downstream-use signals are available in the indexed data. This axis would measure citations to the researcher's work and its uptake by other studies and systems.
Focus0/100
The focus score of 0 reflects an even split in the indexed field distribution: 0.5 for machine learning on efficient NLP systems and model deployment, and 0.5 for machine learning on NLP/LLM alignment and evaluation. Both reviewed 2026 papers sit on opposite sides of this divide — a fast inference decision model versus alignment training — so the profile currently reads as two parallel threads within NLP-focused machine learning rather than one dominant specialty.
Collaboration22/100
The collaboration score of 22 is computed from only 2 distinct coauthors in the indexed record, with Wei Dai and Jiahao Sun as the recurring partners. Both reviewed 2026 papers are first-authored (and neither is corresponding-author), consistent with a compact, small-team working style rather than large consortium science; the co-authorship footprint is still too thin to reveal stable network patterns.
Activity—
Activity is not scored: the indexed record spans too short a window to characterize publishing cadence. This axis would track output per year, recent trend, and consistency of arXiv or peer-reviewed submissions over time.
Reproducibility78/100
The reproducibility score of 78 is the mean readiness rating across 2 reviewed papers. The efficient decision-model paper rates high, while the sycophancy-training paper rates medium because key ablation attributions rely on recalled numbers; Evolve's verdicts also note the author is notably candid about limitations — a trait that supports independent replication efforts.
Research summary
Zehua Cheng is a machine learning researcher at the University of Oxford whose indexed record centers on two first-authored 2026 arXiv preprints in natural language processing. The work splits evenly, at 0.5 each in the indexed field distribution, between efficient NLP systems and model deployment on one side and NLP/LLM alignment and evaluation on the other. The first line, this-that-model-1.0, is a 2B typed decision model reporting 30.9 ms answers at 0.941 on a 68-item cohort, positioned as deciding for a millionth of a cent; Evolve rates its reproducibility readiness high while cautioning that the small cohort and in-distribution training limit generalization. The second line, Euston, trains an 8B model with GRPO and reinforcement learning from verifiable rewards to reject false mathematical claims, reporting a +27.5pp gap while claiming AIME performance is retained; its readiness is medium because attribution relies on recalled ablation numbers. Both verdicts describe the author as notably candid about limitations. The collaboration footprint is compact: 2 distinct coauthors, with Wei Dai and Jiahao Sun recurring, and neither paper lists Cheng as corresponding author. Output, impact, and activity dimensions lack sufficient data to score. The profile is name-resolved from public bibliographic data tied to ORCID 0000-0002-9325-2252.
Collaboration pattern
Collaboration in the indexed record is minimal: the radar reports 2 distinct coauthors, and the frequent-coauthor list names Wei Dai and Jiahao Sun. Both 2026 papers are first-authored and neither carries the corresponding-author flag, suggesting Cheng leads the writing and modeling within a small fixed team. With a collaboration score of 22, no large or cross-institution networks are visible yet; co-authorship breadth is a clear growth area.
Research keywords
Highlights
Typed decision model that answers in 30 ms, for a millionth of a cent
Training away mathematical sycophancy without losing the mathematics
Timeline
- 2026First-authored the arXiv preprint this-that-model-1.0, a typed decision model reporting 30.9 ms decisions at a millionth of a cent.
- 2026First-authored the arXiv preprint Euston, training an 8B model to reject false mathematical claims with a reported +27.5pp gap.
- 2026Both indexed works appeared in 2026 as first-authored arXiv preprints, spanning efficient model deployment and LLM alignment.
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 are computed from OpenAlex-indexed public metadata; only 2 reviewed papers back this profile, so all figures carry small-sample uncertainty.
- Output, impact, and activity axes lack sufficient underlying data and are left unscored rather than estimated.
- Both catalog entries are 2026 arXiv preprints; peer-review status, citation counts, and venue acceptance are not captured in the input data.
- Identity is name-resolved from public bibliographic data (ORCID 0000-0002-9325-2252); affiliation is listed as University of Oxford with no country recorded.