Tsinghua researcher on LLM post-training and continual learning: forgetting mitigation 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—
No output score is available: publication-volume data — total and per-year paper counts — is missing from the record. This axis would measure how much and how consistently Guan publishes; only two 2026 preprints are known, too few to score.
Impact—
No impact score exists: citation and influence data — citation counts, h-index, downstream usage — are absent from the input. This axis would measure how widely the two preprints are read and built upon; it cannot be estimated yet.
Focus0/100
The recorded focus score is 0, which sits oddly beside the raw field distribution: the two cataloged papers split exactly 50/50 between machine learning (LLM post-training) and machine learning (LLM fine-tuning and continual learning), two adjacent subfields. Read together the work is thematically narrow — LLM training dynamics — but the two-way split and the zero score mean this axis should be read cautiously until more papers accrue.
Collaboration58/100
The collaboration score is 58, derived from 17 distinct coauthors in the cataloged record. The frequent-coauthor list — Letian Tao, Jingliang Duan, Wei Xiong, Kehua Sheng, Bo Zhang, Shengbo Eben Li, Shiqi Liu, Zeyu He, Guojian Zhan, Jiaxin Gao, and Feihong Zhang — points to a stable, moderately sized working group rather than solo work. Both papers are middle-authored and non-corresponding, consistent with a contributor role inside a Tsinghua-based team.
Activity—
No activity score is available: recent publication cadence and momentum signals are not in the data. The axis would measure publishing rate over recent years; both known papers date to 2026, which alone cannot support a score.
Reproducibility55/100
Reproducibility scores 55, the mean over the 2 papers assessed; both carry medium repro-readiness. Evolve's reviews name concrete gaps: the JANUS paper tests a narrow scope and its replay set overlaps the evaluation set, while γOPD reports consistent single-run gains without seeds or confidence intervals. The methods are described clearly enough to attempt, but current evidence does not yet support confident independent replication.
Research summary
Yang Guan is a machine-learning researcher at Tsinghua University whose cataloged work centers on how large language models are trained after pretraining, splitting evenly — 50/50 — between LLM post-training and LLM fine-tuning with continual learning. The Evolve catalog holds two 2026 arXiv preprints, on both of which Guan is a middle author and not the corresponding author, consistent with a contributing role inside a larger team. The first, 'Past, Future, All at Once,' proposes post-hoc JANUS rectification via Jacobian null-space projection to ease the stability-plasticity dilemma; Evolve's review finds the first-order theory reasonable but flags a narrow test scope and overlap between the replay and evaluation sets. The second, 'Beyond Token-Local Imitation,' introduces γOPD, adding discounted temporal credit assignment and bounded reward mixing to on-policy distillation; the review deems the theory sound but elementary, with consistent single-run gains and no seeds or confidence intervals. Bibliometric signals record 17 distinct coauthors and a collaboration score of 58, with recurring partners including Letian Tao, Jingliang Duan, Wei Xiong, Kehua Sheng, Bo Zhang, and Shengbo Eben Li. Both papers carry medium repro-readiness and a mean reproducibility score of 55. Output, impact, and activity are unscored for lack of data, so this profile should be read as early-stage.
Collaboration pattern
Guan's cataloged work is team-based: 17 distinct coauthors and a collaboration score of 58, with middle, non-corresponding authorship on both papers. Recurring partners include Letian Tao, Jingliang Duan, Wei Xiong, Kehua Sheng, Bo Zhang, and Shengbo Eben Li, suggesting a stable core group spanning both threads — continual-learning-oriented fine-tuning and RL-flavored post-training.
Research keywords
Highlights
Post-hoc Jacobian null-space rectification against catastrophic forgetting
γOPD: reward-compatible temporal credit assignment for on-policy distillation
Timeline
- 2026First cataloged preprint appears: 'Past, Future, All at Once' proposes post-hoc JANUS rectification for the stability-plasticity dilemma (arXiv, middle author).
- 2026Second cataloged preprint appears: 'Beyond Token-Local Imitation' introduces γOPD for on-policy distillation (arXiv, middle author).
- 2026Both 2026 preprints sit in the Evolve catalog with medium repro-readiness and a mean reproducibility score of 55 across the 2 assessed papers.
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 public bibliographic data (OpenAlex-sourced) only; no curated citation database is involved.
- 'Yang Guan' is a common name; this profile is name-resolved from public bibliographic records, and no ORCID is provided to confirm identity.
- Catalog coverage is limited to two 2026 arXiv preprints; no peer-reviewed venue records, citation counts, h-index, or career facts are in scope.
- Output, impact, and activity axes are unscored because their underlying data is missing; the focus score of 0 should be interpreted cautiously.