Jiahao Sun

Flock (United Kingdom)

Unclaimed · AI-structured profileOpenAlex

ML researcher at Flock (UK): efficient typed decision models and alignment work curbing LLM mathematical 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—

Output metrics such as total publication count, venue record, and yearly volume are not yet available for this profile — the catalog holds only two reviewed 2026 arXiv papers, too few to score productivity.

Impact—

Impact data — citations, h-index, and downstream usage — is not available in the current snapshot, so this axis is left unscored rather than estimated from the two reviewed arXiv papers.

Focus0/100

This axis scores 0, and the raw field distribution explains why: research effort splits exactly 50/50 between machine learning for efficient NLP systems and model deployment, and machine learning for NLP/LLM alignment and evaluation. With two 2026 arXiv papers each anchoring one subfield, no dominant concentration has emerged yet, so focus reads as minimal rather than missing.

Collaboration22/100

Collaboration scores 22, consistent with a compact network: only 2 distinct co-authors appear in this profile, with Zehua Cheng and Wei Dai listed as the recurrent partners. Both catalogued papers carry Sun in the last author position without a corresponding-author flag, a pattern often read as a supervisory role within a small, stable team. No large consortium or cross-institution collaboration is visible in the current data.

Activity—

Activity metrics — publication cadence, submission frequency, and recent output volume — are not yet computable: the snapshot covers only two 2026 arXiv papers, so this axis is left unscored until a longer publication window is observed.

Reproducibility78/100

Reproducibility scores 78, the mean across the 2 papers Evolve has assessed. this-that-model-1.0 earns a high readiness rating, with structured outputs underpinned by proper scoring rules and Brier score calibration, while Euston is rated medium because its attribution depends on recalled ablation numbers. Evolve's verdicts repeatedly note candid treatment of limitations, though small evaluation cohorts temper confidence in both.

Research summary

Jiahao Sun is a machine learning researcher at Flock (United Kingdom) whose profile is name-resolved from public bibliographic data and currently rests on two 2026 arXiv papers reviewed by Evolve. Both papers place him in the last author position, a pattern consistent with a senior or supervising role, although neither lists him as corresponding author. His work splits evenly — 50/50 in Evolve's field distribution — between efficient NLP systems and model deployment, and NLP/LLM alignment and evaluation. The first line of work is this-that-model-1.0, a 2B-parameter typed decision head that reportedly answers in 30.9 ms at 0.941 on a 68-item cohort, framed around structured outputs, proper scoring rules, and Brier score calibration. The second, Euston, uses GRPO-based reinforcement learning from verifiable rewards to train an 8B model to reject false mathematical claims, reporting a +27.5pp gap while claiming AIME capability is retained. Evolve rates reproducibility high for the first paper and medium for the second, yielding a reproducibility mean of 78 across 2 papers. Recurring co-authors are Zehua Cheng and Wei Dai; the distinct-coauthor count of 2 indicates a compact collaboration footprint. Evolve's verdicts praise candour about limitations but flag small evaluation cohorts and reliance on recalled ablation numbers.

Collaboration pattern

Jiahao Sun's recorded collaboration footprint is small: 2 distinct co-authors, with Zehua Cheng and Wei Dai as the recurring partners. On both catalogued papers he sits in the last author position, and neither flags him as corresponding author — a signature consistent with supervising a compact, stable two-person team rather than joining large consortia. No cross-institution or industry-academia network is visible in the current snapshot; all catalogued work appears on arXiv from his Flock (UK) affiliation.

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: 30.9 ms typed decisions at 0.941 on a 68-item cohort
Evolve's review describes a 2B-parameter typed decision head that answers in 30.9 ms — pitched in the title as deciding for a millionth of a cent — scoring 0.941 on a 68-item cohort, built on structured outputs, proper scoring rules, and Brier score calibration. The verdict credits candour about limits while flagging small samples and in-distribution training; reproducibility readiness is rated high.
Euston: +27.5pp gap in rejecting false math, with claimed AIME retention
Evolve's review covers GRPO-based reinforcement learning from verifiable rewards that trains an 8B model to reject false mathematical claims, widening a reported +27.5pp gap while claiming the mathematics itself is retained on AIME, supported by claim verification and synthetic data generation. The verdict calls the paper notably candid but cautions that attribution rests on recalled ablation numbers; reproducibility readiness is medium.
Candid-by-design reproducibility profile: 78/100 across two reviewed papers
Across Evolve's two reviews, both papers are flagged as candid about limitations, producing a reproducibility mean of 78 with high and medium readiness ratings. The recurring caveats — small evaluation cohorts, in-distribution training, and recalled ablation numbers — define exactly what larger-scale, held-out evaluation would need to fix next.

Timeline

2026Publishes this-that-model-1.0 on arXiv — a 2B typed decision model reporting 30.9 ms latency at 0.941 on a 68-item cohort; Evolve rates reproducibility readiness high and notes candid treatment of limits.
2026Publishes Euston on arXiv — GRPO-based reinforcement learning from verifiable rewards training an 8B model to reject false math with a +27.5pp gap and claimed AIME retention; Evolve rates reproducibility readiness medium.
2026Both papers appear with Sun in the last author position and a two-person recurring co-author set (Zehua Cheng, Wei Dai), establishing the current two-track profile across efficient decision models and LLM alignment.

Papers in the Evolve catalog

Frequent co-authors

Zehua Cheng×2 Wei Dai×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.

Radar scores and bibliometrics come from Evolve's public bibliometric snapshot (OpenAlex-based) only; no citation counts or h-index were available for this profile.
Output, impact, and activity axes are unscored (dimMissing): the profile rests on just two reviewed arXiv papers, with no citation or venue-acceptance data behind them.
No ORCID or verified author disambiguation is attached; Jiahao Sun is a common name, so this profile is name-resolved from public bibliographic data and records may merge or split.
Both catalogued papers are 2026 arXiv preprints reviewed by Evolve; peer-reviewed venue acceptance is not reflected in the current data.