Machine learning researcher on LLM post-training: continual-learning forgetting and temporal credit assignment.
2
papers here
2
reviewed by Evolve
0
first-author
0
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: this snapshot lacks sufficient data on publication volume. This axis would count total papers, preprints versus peer-reviewed outputs, and yearly productivity to gauge research production.
Impact—
Not yet scored: citation and influence data are unavailable in this snapshot. This axis would aggregate citations, citing works, and uptake of the two 2026 preprints to gauge the community's response.
Focus0/100
The focus axis is recorded as 0, and the underlying field distribution is split exactly 50/50 between 'machine learning (LLM post-training)' and 'machine learning (LLM fine-tuning and continual learning)'. With only two reviewed papers, both 2026 arXiv preprints on adjacent LLM-training topics, the visible footprint is narrow; the zero score should be read against this very thin record rather than as demonstrated breadth of interests.
Collaboration58/100
Collaboration scores 58, with 17 distinct coauthors visible across the record. The recurring pool includes Letian Tao, Jingliang Duan, Wei Xiong, Bo Zhang, Yang Guan, Shengbo Eben Li, Shiqi Liu, Zeyu He, Guojian Zhan, Jiaxin Gao, and Feihong Zhang. On both cataloged papers Sheng holds a middle author position without corresponding-author status, consistent with work carried out inside larger teams rather than solo-led efforts.
Activity—
Not yet scored: recent-activity signals are insufficient in this snapshot. This axis would track submission cadence and recency of outputs; only the two 2026 arXiv entries currently anchor the timeline.
Reproducibility55/100
Reproducibility scores 55, a mean over the 2 papers Evolve has assessed, each rated medium repro-readiness. The reviews qualify the claims: the JANUS paper tests a narrow range of settings and its replay set overlaps the evaluation set, while γOPD reports consistent but single-run gains without seeds or confidence intervals. The methods are theoretically sound on paper, yet independent verification would need extra experimental controls.
Research summary
Kehua Sheng (Vision pour la Robotique; ORCID 0009-0008-3370-7711) is a machine learning researcher whose visible record sits squarely in post-training of large language models. Evolve's catalog holds two 2026 arXiv preprints, and the field distribution splits evenly: 50% machine learning (LLM post-training), 50% machine learning (LLM fine-tuning and continual learning). The first, 'Past, Future, All at Once,' attacks the stability-plasticity dilemma via post-hoc Jacobian null-space rectification (JANUS) to curb catastrophic forgetting in fine-tuned LLMs; Evolve finds the first-order theory reasonable but flags narrow test coverage and an overlap between the replay and evaluation sets. The second, 'Beyond Token-Local Imitation,' proposes γOPD, adding discounted temporal credit assignment and bounded reward mixing to on-policy distillation; the review calls the theory sound but elementary, with consistent single-run gains, no seeds, and no confidence intervals. Sheng sits in a middle author position on both papers without corresponding-author status, and 17 distinct coauthors appear in the record, including recurring collaborators such as Letian Tao, Jingliang Duan, Wei Xiong, and Shengbo Eben Li. Radar scores collaboration at 58 and reproducibility at 55, a mean over the two assessed papers, both rated medium repro-readiness; focus registers 0 against this two-paper footprint, while output, impact, and activity axes are unscored for insufficient data. The profile is name-resolved from public bibliographic data.
Collaboration pattern
The record shows 17 distinct coauthors around a two-paper footprint, scoring 58 on the collaboration axis. Recurring collaborators include Letian Tao, Jingliang Duan, Wei Xiong, Bo Zhang, Yang Guan, Shengbo Eben Li, Shiqi Liu, Zeyu He, Guojian Zhan, Jiaxin Gao, and Feihong Zhang. Both cataloged papers place Sheng in a middle author position without corresponding-author status, suggesting contributions within larger teams rather than solo-led projects.
Research keywords
Highlights
JANUS: post-hoc Jacobian null-space rectification against forgetting in LLM fine-tuning
γOPD: reward-compatible temporal credit assignment for on-policy distillation
Timeline
- 2026Posts 'Past, Future, All at Once' (JANUS) to arXiv, targeting the stability-plasticity dilemma in LLM fine-tuning and continual learning.
- 2026Posts 'Beyond Token-Local Imitation' (γOPD) to arXiv, adding temporal credit assignment and bounded reward mixing to on-policy distillation.
- 2026Both preprints enter the Evolve review catalog, each rated medium repro-readiness, with a reproducibility mean of 55 across the two assessed papers.
- 2026The visible collaboration footprint reaches 17 distinct coauthors across these outputs, with the field distribution split evenly between LLM post-training and LLM fine-tuning/continual learning.
Papers in the Evolve catalog
Frequent co-authors
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 are computed from OpenAlex-derived public bibliometric data only; no publisher or author-curated records are included.
- Output, impact, and activity axes are unscored in this snapshot, so citation counts, h-index-style aggregates, and publication cadence are not quoted anywhere in this profile.
- Only the two Evolve-reviewed 2026 arXiv preprints are covered; both are middle-authorship, non-corresponding works, so the visible footprint may under-represent the full publication record.
- 'Kehua Sheng' is name-resolved from public bibliographic data (ORCID 0009-0008-3370-7711); records from namesakes cannot be fully excluded.