Wei Dai

Flock (United Kingdom)

Unclaimed · AI-structured profileOpenAlex

Machine learning researcher at Flock (UK): efficient small decision models and alignment against LLM sycophancy

papers here

reviewed by Evolve

first-author

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—

No verified bibliometric output record is linked to this profile yet — no publication count, venue list, or authorship-role breakdown. This axis would measure cumulative paper volume across venues and years.

Impact—

Citation counts, h-index, and field-weighted influence metrics are unavailable for this profile. This axis would measure how widely the two 2026 preprints and any other work are cited and taken up by the community.

Focus0/100

Focus scores 0 because the field distribution is an exact 50/50 split between two distinct directions — efficient NLP systems and model deployment versus NLP/LLM alignment and evaluation — across only two 2026 arXiv papers. With one paper per subfield, this reflects a deliberately dual-track agenda rather than measurable concentration; the axis needs more papers to resolve a coherent primary field.

Collaboration22/100

Collaboration scores 22, grounded in raw counts: only 2 distinct coauthors — Zehua Cheng and Jiahao Sun — appear across both catalog papers, and Wei Dai sits in a middle authorship position on each without corresponding authorship. This points to a small, tight-knit team producing both 2026 preprints together, with limited demonstrated external network breadth so far; the score should be read against a two-paper sample.

Activity—

Publication-velocity data is insufficient: the catalog holds only two 2026 arXiv entries with no linked multi-year record. This axis would measure output cadence — papers per year and sustained publishing activity over time.

Reproducibility78/100

Reproducibility scores 78, the mean across 2 scored papers. The typed decision model paper rates high readiness, while the Euston sycophancy paper rates medium. Evolve's verdicts repeatedly flag candor about limitations — the decision-head paper discloses its small 68-item cohort, and Euston is notably candid — but also flag weaknesses: in-distribution training and reliance on recalled ablation numbers. Strong self-reporting discipline, with verification still pending.

Research summary

Wei Dai is a machine learning researcher at Flock (United Kingdom) whose Evolve catalog profile currently rests on two 2026 arXiv preprints, both produced with a small team — Zehua Cheng and Jiahao Sun are the only 2 distinct coauthors on record. The work splits evenly, 50/50 by field weight, between two directions: efficient NLP systems and model deployment, and NLP/LLM alignment and evaluation. The first strand produced this-that-model-1.0, a 2B typed decision head claiming 30.9 ms answers at a reported 0.941 on a 68-item cohort, at a per-query cost of a millionth of a cent; Evolve rates its reproducibility readiness high while cautioning that small samples and in-distribution training weaken the claim. The second produced Euston, which trains an 8B model to reject false mathematical claims — a reported +27.5pp gap — while claiming AIME performance is retained, using GRPO under reinforcement learning from verifiable rewards; Evolve calls it notably candid but notes attribution rests on recalled ablation numbers. Bibliometrically this is an early-stage profile: reproducibility is the standout axis at 78 (mean of 2 scored papers), collaboration is low at 22 given the two-person coauthor set and middle, non-corresponding authorship on both papers, and the output, impact, and activity axes cannot be scored because no verified publication record, citation data, or multi-year history is linked. Wei Dai is a common name, so this profile is name-resolved from public bibliographic data only.

Collaboration pattern

Both catalog papers come from the same small team: Wei Dai appears as a middle author on each, with only 2 distinct coauthors on record — Zehua Cheng and Jiahao Sun, who recur across both 2026 preprints. No corresponding authorship appears in the catalog, and no external collaborations are visible yet. Treat the collaboration footprint as provisional given the two-paper sample.

Research keywords

typed decision modelstructured outputsproper scoring rulesBrier score calibrationlinear attentionsmall language modelsmathematical sycophancyGRPOreinforcement learning from verifiable rewardsclaim verificationsynthetic data generationcapability retention

Highlights

this-that-model-1.0: a 2B typed decision head claiming 30.9 ms answers at 0.941
Reviewed as an efficient small decision model claiming answers in 30.9 ms at a reported 0.941 on a 68-item cohort, at a per-query cost of a millionth of a cent. Evolve rates its reproducibility readiness high, but cautions that small samples and in-distribution training weaken the claim.
Euston: cutting mathematical sycophancy with a +27.5pp gap while claiming AIME retained
Trains an 8B model to reject false mathematical claims with a reported +27.5pp gap while claiming AIME performance is retained, using GRPO under reinforcement learning from verifiable rewards. Evolve calls the work notably candid, but notes its attribution rests on recalled ablation numbers; reproducibility readiness is medium.

Timeline

2026Two arXiv preprints enter the Evolve catalog, both listing Wei Dai in a middle authorship position alongside the only coauthors on record, Zehua Cheng and Jiahao Sun.
2026this-that-model-1.0 appears on arXiv: a 2B typed decision model claiming 30.9 ms answers at a reported 0.941 on a 68-item cohort; Evolve rates its reproducibility readiness high.
2026Euston appears on arXiv: an 8B model trained to reject false mathematical claims with a reported +27.5pp gap while claiming AIME is retained; Evolve rates its reproducibility readiness medium.

Papers in the Evolve catalog

Frequent co-authors

Zehua Cheng×2 Jiahao Sun×2

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.

The name Wei Dai is common; this profile is resolved from public bibliographic data (arXiv records), with no ORCID linked to confirm identity.
Bibliometric scores are derived from OpenAlex-linked data only; the output, impact, and activity axes lack sufficient underlying data to score.
The catalog contains only two 2026 arXiv preprints, neither peer-reviewed, making this an early-stage profile.
No citation counts, h-index, or multi-year publication history are available, so influence and output trends cannot be characterized.