Tung Sum Thomas Kwok

University of California, Los Angeles · United States

Unclaimed · AI-structured profileOpenAlex

UCLA researcher spanning discrete diffusion generative models and LLM-agent systems for financial QA.

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-13) 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 are unavailable for this profile: publication counts by venue and over time could not be assembled from the indexed record, so this axis, which would measure publication volume and productivity, is left unscored.

Impact—

Impact cannot be assessed: the record shows an h-index of 0 and i10-index of 0 with incomplete citation coverage, so this axis, which would measure citations, uptake, and influence, is left unscored rather than estimated.

Focus0/100

The focus score of 0 reflects an evenly split portfolio: the OpenAlex field distribution assigns 50% to NLP/LLM agents (financial QA) and 50% to machine learning (generative models/discrete diffusion), and five topics share the record — Topic Modeling (11), NLP Techniques (5), Domain Adaptation and Few-Shot Learning (3), Multimodal Applications (3), and Information Retrieval (2). The two 2026 catalog papers sit on opposite sides of this divide.

Collaboration72/100

With a collaboration score of 70, this is a strongly team-based profile: 32 distinct coauthors appear in the recent OpenAlex window, and 12 recur often enough to be listed, including Yanzhang Ma, Zhenghan Tai, Hanwei Wu, and Sizhe Guan. Both reviewed 2026 papers are multi-author arXiv efforts — LEDFlow first-authored, FINSKILLOPS in a middle position — consistent with contribution to larger group projects rather than solo work.

Activity79/100

Activity scores 79 on the strength of a concentrated recent record: OpenAlex logs 14 works in the last three years, with countsByYear placing all 14 in 2026. The burst is recent — both catalog papers appeared as 2026 arXiv preprints — so this is a fast-starting profile whose rhythm is visible only over one year; whether the pace is sustained beyond 2026 cannot be judged from the data provided.

Reproducibility58/100

Reproducibility averages 38 (reproMean) across the 2 Evolve-reviewed papers (reproPapers: 2). LEDFlow is rated medium readiness — its entropy-ordering claims include a local theory and large Sudoku gains, though broader gains are partly within noise. FINSKILLOPS is rated low: its six-benchmark wins rest on author-built evaluations, judge-based scoring, and artifacts that could not be verified. Artifact availability is the main limiting factor across both.

Research summary

Tung Sum Thomas Kwok is a UCLA-affiliated researcher whose recorded work splits evenly between two active fronts: machine learning with generative models and discrete diffusion, and NLP / LLM-agent systems aimed at financial question answering. The catalog's anchor is LEDFlow (2026, first author, arXiv), which introduces entropy-guided generation order into uniform discrete flow: entropy-ordered absorption fixes reliable tokens early, supported by local theory and large gains on Sudoku, though Evolve notes broader gains are modest and partly within noise. The second reviewed paper, FINSKILLOPS (2026), is a self-evolving multi-agent system for SEC filing QA built around gated skill maintenance and regression gating; Evolve judges its six-benchmark wins credible in direction but tempered by author-built evaluations, judge bias, and unverifiable artifacts. OpenAlex records 14 works in the recent three-year window, all attributed to 2026, with 32 distinct coauthors and 12 recurring collaborators, so the profile is team-based and fast-starting. Topic coverage is broad: Topic Modeling (11), NLP Techniques (5), Domain Adaptation and Few-Shot Learning (3), Multimodal Applications (3), and Information Retrieval (2). Citation-based output and impact cannot yet be assessed (h-index and i10-index recorded as 0 under incomplete coverage), and reproducibility evidence averages 38 across the two reviewed papers. This is an early-stage, rapidly producing profile whose long-term trajectory is not yet established.

Collaboration pattern

Collaboration is team-centered: the recent OpenAlex window lists 32 distinct coauthors, 12 of whom recur — including Yanzhang Ma, Zhenghan Tai, Hanwei Wu, Sizhe Guan, Jianliang Lei, and Hailin He. LEDFlow is first-authored while FINSKILLOPS places Kwok mid-list, consistent with rotating roles inside larger arXiv-group projects rather than a fixed two-person partnership; no senior mentor can be identified from the data provided.

Research keywords

discrete flow matchingdiffusion language modelsentropy-guided decodingsampling algorithmstext-to-image generationLLM agentsfinancial question answeringSEC filing QAmulti-agent systemsretrieval-augmented generationself-evolving agentstopic modeling

Highlights

Entropy-ordered generation for uniform discrete flow (LEDFlow, 2026)
First-authored arXiv paper introducing entropy-guided generation order into uniform discrete flow matching: reliable tokens are absorbed first, backed by a local theory and large Sudoku gains; Evolve notes broader-task improvements are modest and partly within noise. Reproducibility readiness: medium.
Self-evolving multi-agent system for SEC filing QA (FINSKILLOPS, 2026)
Co-authored arXiv system combining gated skill maintenance, retrieval-augmented generation, and regression gating for SEC filing question answering; Evolve finds the six-benchmark wins directionally real but tempered by author-built evaluations, judge bias, and unverifiable artifacts. Reproducibility readiness: low.
Fast-starting, team-based 2026 output burst
OpenAlex logs 14 works in the recent three-year window, all attributed to 2026, alongside 32 distinct coauthors and 12 recurring collaborators — a rapid, group-oriented start spanning two research lines.

Timeline

2026First-authored the arXiv preprint LEDFlow, introducing entropy-guided generation order into uniform discrete flow matching.
2026Co-authored FINSKILLOPS, a self-evolving multi-agent system for SEC filing QA, in a middle author position.
2026OpenAlex recorded 14 works in the recent three-year window, all attributed to 2026, with 32 distinct coauthors.

Papers in the Evolve catalog

Frequent co-authors

Zhenghan Tai×2 Xinyu Wang×1 Tung Sum Thomas Kwok×1 Zhenghan Tai×1 Guang Cheng×1 Yanzhang Ma×1 Hanwei Wu×1 Sizhe Guan×1 Jianliang Lei×1 Hailin He×1 Chaolong Jiang×1 Jijun Chi×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.

Bibliometric indicators come from an OpenAlex name-resolved profile (no ORCID linked), so identity matching for this name may be imperfect.
Output and impact axes are unscored: citation coverage is incomplete, with h-index and i10-index recorded as 0.
Qualitative assessments rest on only two 2026 arXiv preprints reviewed by Evolve; peer-reviewed venues and earlier work are absent from the evidence base.
The reproducibility score averages just two reviewed papers (reproMean 38, reproPapers 2) and is indicative rather than stable.