Jian Kang

Ant Group (China)

Unclaimed · AI-structured profileOpenAlex

Ant Group LLM researcher: on-policy distillation, gradient estimation, and mechanistic analysis of reasoning models.

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-17) 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—

This axis is not scored yet: publication-volume data such as paper counts over time, venues, and authorship roles is insufficient in the current snapshot. It would measure research productivity and output consistency across the researcher's career, and should be assessed only once more dated records are available.

Impact—

No citation-based data is available, so impact is left unscored. The axis would measure scholarly influence through citation counts, citing works, and downstream adoption of the distillation and reasoning-model methods described in the catalog entries; none of this can be estimated from the inputs.

Focus100/100

The focus dimension scores at the maximum of the scale (100), and the underlying field distribution explains why: every catalogued field assignment falls into a single area, machine learning (large language models). Both 2026 entries — on-policy distillation and the hierarchical reasoning model study — sit inside this one field, indicating an unusually concentrated research identity rather than a spread across disciplines.

Collaboration41/100

The collaboration dimension scores 41, computed over a compact network of 7 distinct co-authors observed in the bibliometric record. Recurring names include Huanxin Sheng, Zhiling Ye, Haonan Wang, Jian Wang, Jinjie Gu, and Leo Raphael Rodrigues. Both catalogued papers place Jian Kang in the last author position, consistent with a senior, supervisory role within a small, stable team rather than large multi-institution consortia.

Activity—

Recent-activity data is not yet available, so this axis is left unscored. It would measure publishing cadence and sustained engagement over time, which requires more dated records than the two 2026 arXiv entries currently on file; no trend or productivity pace is inferred here.

Reproducibility55/100

The reproducibility dimension averages 55 across the single paper assessed so far (reproMean 55, reproPapers 1). Evolve's review of the on-policy distillation paper rates reproducibility readiness as medium and cautions that many reported gains fall within the reported noise bounds — a useful signal that replication requires careful variance reporting, though conclusions remain provisional with a one-paper sample.

Research summary

Jian Kang is a researcher at Ant Group (China) whose Evolve profile is tightly concentrated in machine learning for large language models. The bibliometric snapshot records two arXiv papers, both from 2026, with Jian Kang in the last author position on each. The reviewed paper, '1% of Tokens Can Be Enough: On Gradient Estimation in On-Policy Distillation,' proposes IER, a method for selecting reliable-gradient tokens in sparse on-policy distillation while arguing that about 1% of tokens can suffice; Evolve's verdict finds the theory and experiments broad but flags that many reported gains fall within the reported noise, and reproducibility readiness is rated medium. The second entry, 'Dissecting Hierarchical Reasoning Models: A Mechanistic Study,' applies mechanistic analysis to hierarchical reasoning architectures and has not yet received a full Evolve review. The radar reflects this early, narrow footprint: focus scores 100 because all field assignments fall in a single field, collaboration scores 41 over 7 distinct co-authors, and reproducibility averages 55 from one assessed paper, while output, impact, and activity remain unscored for lack of data. Recurring co-authors include Huanxin Sheng, Zhiling Ye, Haonan Wang, Jian Wang, Jinjie Gu, and Leo Raphael Rodrigues. The profile is name-resolved from public bibliographic data and should be read as early-stage.

Collaboration pattern

The observed collaboration pattern is compact: 7 distinct co-authors appear in the record, and the collaboration dimension scores 41. Recurring names include Huanxin Sheng, Zhiling Ye, Haonan Wang, Jian Wang, Jinjie Gu, and Leo Raphael Rodrigues. On both catalogued 2026 papers Jian Kang sits in the last author position without corresponding-author flags, consistent with a senior supervisory role within a small, stable team rather than wide multi-institution networks.

Research keywords

On-policy distillationGradient estimationToken selectionLarge language modelsKnowledge distillationMechanistic interpretabilityHierarchical reasoning modelsReliable-gradient tokensTraining efficiencyLLM reasoning

Highlights

IER: selecting reliable-gradient tokens in sparse on-policy distillation
The reviewed 2026 arXiv paper proposes IER to pick the tokens whose gradients can be trusted in sparse on-policy distillation, arguing that about 1% of tokens can be enough. Evolve notes the theory and experiments are broad, but flags that many reported gains sit within the reported noise.
Reproducibility caution on the 1% claim
Evolve rates the paper's reproducibility readiness as medium, and that single assessment supplies the 55 mean of the reproducibility dimension. Readers replicating the result should attend to variance: the verdict explicitly states that many reported gains fall within the reported noise.
Mechanistic dissection of Hierarchical Reasoning Models
A second 2026 arXiv entry, 'Dissecting Hierarchical Reasoning Models: A Mechanistic Study,' again with Jian Kang in the last author position, applies mechanistic analysis to hierarchical reasoning architectures. It has not yet received a full Evolve review, so no verdict or reproducibility rating exists.

Timeline

2026Published the arXiv study '1% of Tokens Can Be Enough: On Gradient Estimation in On-Policy Distillation' in the last author position; Evolve reviewed it, noting broad theory and experiments but gains within reported noise, and rated reproducibility readiness medium.
2026Appeared as last author on 'Dissecting Hierarchical Reasoning Models: A Mechanistic Study,' extending the 2026 output into mechanistic analysis of hierarchical reasoning architectures.
2026Both catalogued papers landed in the same field — machine learning (large language models) — producing the maximum focus score of 100 in the Evolve radar.

Papers in the Evolve catalog

2026-09-21
1% of Tokens Can Be Enough: On Gradient Estimation in On-Policy Distillation

Frequent co-authors

Huanxin Sheng×1 Zhiling Ye×1 Haonan Wang×1 Jian Wang×1 Jinjie Gu×1 Haonan Wang×1 Leo Raphael Rodrigues×1

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.

Bibliometrics are name-resolved from public bibliographic data; no ORCID is on file and Jian Kang is a common Chinese name, so identity confidence is limited.
Only two arXiv entries from 2026 are in the Evolve catalog; the output, impact, and activity dimensions are unscored because their underlying data is missing.
The reproducibility score of 55 rests on a single assessed paper (reproPapers = 1) and should be treated as provisional.
No citation counts, h-index, venue acceptance statistics, or award information was provided; none is quoted or inferred in this profile.