Hangoo Kang

Stanford University · United States

Unclaimed · AI-structured profileOpenAlex

Stanford researcher working across LLM alignment and distributed training systems for diffusion language models

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-15) 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—

Not yet scored: the current OpenAlex record offers no usable publication totals beyond 7 recent works, and citation-linked fields read zero, so this axis — total output, growth trajectory, and venue mix — cannot be measured from the data at hand.

Impact—

Not yet scored: citation counts, h-index trajectory, and downstream uptake are absent or zero in the current record (h-index 0, i10-index 0), so this axis — citation influence and field-normalized impact — has insufficient data to report.

Focus0/100

Scores 0/100 because the field distribution is split exactly 50/50 between LLM alignment and post-training and distributed LLM training systems. The two cataloged 2026 papers mirror that even split, and OpenAlex topics — natural language processing (5), topic modeling (4), reinforcement learning in robotics (4), parallel computing (2), AI applications (2) — reinforce a deliberately two-sided, cross-stack profile rather than narrow specialization.

Collaboration51/100

Scores 51/100. The recent window records 12 distinct coauthors, and the frequent-coauthor list spans NLP (Yejin Choi, Natasha Jaques, Liwei Jiang), robotics-RL and systems (Vikram Iyer, Azalia Mirhoseini, Hermann Kumbong), plus a cluster including Jiayi Yuan, James Jihao Liu, Tarun Suresh, Pranshu Chaturvedi, Parth Shroff, and Ishan S. Khare. Both cataloged 2026 papers are middle-author team efforts with no corresponding role — a mid-range, group-based collaboration footprint.

Activity61/100

Scores 61/100. OpenAlex logs 7 works in the trailing three-year window, with all 7 dated 2026 (countsByYear shows only 2026: 7), indicating a concentrated recent burst rather than long-run accumulation. Two of these are Evolve-reviewed arXiv preprints spanning alignment and distributed-training systems, consistent with an active early-stage publishing cadence.

Reproducibility55/100

Scores 55/100, the mean across the 2 reproducibility-scored papers, both rated medium readiness. The systems paper reports proven-exact block parallelism with 1.2–7.6x gains, though its evaluation deltas lack significance tests; the alignment paper's claimed +265% SBERT diversity gain lacks variance estimates. Stronger statistical reporting on both would lift this axis.

Research summary

Hangoo Kang is a Stanford University researcher whose recorded work splits evenly between natural language processing — specifically LLM alignment and post-training — and machine learning systems for distributed LLM training. The OpenAlex record names 12 distinct coauthors in the recent window and lists 7 works over the past three years, all dated 2026, suggesting a concentrated early burst of output rather than an accumulated back catalog. Two of those works are reviewed in the Evolve catalog: a systems paper introducing proven-exact block parallelism (CSBP) for long-context block diffusion language model training with reported 1.2–7.6x efficiency gains, and Forty Shades of Blue, which uses mode-conditioned reinforcement learning to raise LLM output diversity — a claimed 265% gain on SBERT-based measures — without quality loss. Evolve rates both papers medium on reproducibility readiness, averaging 55 across the two, and flags missing significance tests and evaluator-overlap concerns respectively. Radar dimensions place activity at 61, collaboration at 51, and reproducibility at 55, while output and impact remain unscored because citation data is not yet available. Topic frequencies — natural language processing techniques (5), topic modeling (4), reinforcement learning in robotics (4), parallel computing (2), and AI applications (2) — round out a two-sided, systems-plus-alignment profile.

Collaboration pattern

Collaboration scores 51/100. The recent window records 12 distinct coauthors, and the frequent-coauthor list mixes senior NLP researchers (Yejin Choi, Natasha Jaques, Liwei Jiang), robotics-RL and systems names (Vikram Iyer, Azalia Mirhoseini, Hermann Kumbong), and a peer cluster including Jiayi Yuan, James Jihao Liu, Tarun Suresh, Pranshu Chaturvedi, Parth Shroff, and Ishan S. Khare. Both cataloged papers are middle-author team efforts with no corresponding role, consistent with contributing inside larger groups.

Research keywords

LLM alignmentpost-trainingquality-diversity trade-offmode collapseGRPOmulti-agent reinforcement learningdiffusion language modelscontext parallelismdistributed traininglong-context trainingspeculative decodingtopic modeling

Highlights

Proven-exact block parallelism for long-context diffusion LM training
Evolve's review credits the CSBP approach with proven-exact block parallelism for long-context block diffusion language model training and 1.2–7.6x efficiency gains, calling it strong systems work while noting the evaluation deltas lack significance tests.
Mode-conditioned RL for quality-diversity alignment
Forty Shades of Blue claims mode-conditioned reinforcement learning lifts LLM output diversity by 265% on SBERT-based evaluation without quality loss; Evolve flags evaluator overlap and absent variance estimates as reasons for caution.
Medium reproducibility readiness across both reviewed papers
Both Evolve-reviewed papers are rated medium on reproducibility readiness, averaging 55 across the 2 reproducibility-scored papers; adding significance tests and variance estimates is the clearest path to a stronger score.

Timeline

2026First Evolve-cataloged systems work: block parallelism (CSBP) for efficient distributed long-context block diffusion language model training, reporting 1.2–7.6x gains.
2026First Evolve-cataloged alignment work: Forty Shades of Blue applies mode-conditioned reinforcement learning to widen LLM output diversity without quality loss.
2026OpenAlex logs 7 works in the trailing three-year window, all dated 2026, with 12 distinct coauthors — the profile's entire recorded activity sits in this year.

Papers in the Evolve catalog

Frequent co-authors

Jiayi Yuan×1 James Jihao Liu×1 Yejin Choi×1 Vikram Iyer×1 Liwei Jiang×1 Natasha Jaques×1 Tarun Suresh×1 Pranshu Chaturvedi×1 Parth Shroff×1 Ishan S. Khare×1 Hermann Kumbong×1 Azalia Mirhoseini×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 scores, topics, and coauthor counts come from an OpenAlex record with no ORCID linkage; identity is name-resolved from public bibliographic data and may merge or split namesakes.
OpenAlex lists h-index and i10-index as 0, and the output and impact dimensions are unscored, so citation-based impact cannot yet be assessed.
Only two Evolve-reviewed papers are in scope, both 2026 arXiv preprints with middle authorship and no corresponding-author role; no peer-reviewed venue data is included.
Counts by year show activity only in 2026 (7 works), so career-length output trends are unknown from this record.