Tsinghua machine learning researcher; last-position author of 2026 work on LLM post-training and continual learning
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 publication counts, venue records or output totals are available in this snapshot, so output is not scored. This axis would measure the volume, cadence and venue spread of the researcher's published work.
Impact—
No citation counts, h-index or influence metrics are provided, so impact is not scored. This axis would measure how widely the work is cited and its downstream influence across the field.
Focus0/100
The radar focus score is 0, the lowest recorded value in this snapshot, while the underlying field distribution splits evenly — 50% machine learning (LLM post-training) and 50% machine learning (LLM fine-tuning and continual learning). Read together, recorded output spans two adjacent LLM subfields rather than a single narrow niche.
Collaboration58/100
The collaboration score of 58 rests on 17 distinct coauthors recorded in the snapshot, with a recurring core including Letian Tao, Jingliang Duan, Wei Xiong, Kehua Sheng and Bo Zhang. Both reviewed 2026 papers place this researcher in the last author position, consistent with a senior, group-leading role rather than solitary work.
Activity—
Year-by-year productivity data is insufficient in this snapshot, so activity is not scored. This axis would measure recent publication cadence and research momentum over time.
Reproducibility55/100
Reproducibility scores 55, matching a reproMean of 55 computed over 2 reviewed papers, each rated medium reproducibility readiness. The reviews flag concrete gaps: the JANUS work shows narrow test coverage with replay/evaluation overlap, and the γOPD work reports consistent but single-run gains without seeds or confidence intervals — sound method design, limited empirical auditing.
Research summary
Shengbo Eben Li is a machine learning researcher at Tsinghua University whose recent recorded work concentrates on post-training of large language models. The Evolve catalog holds two 2026 arXiv papers, both with this researcher in the last author position: a continual-learning study proposing post-hoc Jacobian null-space projection (JANUS) to mitigate the stability-plasticity dilemma in LLM fine-tuning, and a post-training study introducing γOPD, which adds reward-compatible temporal credit assignment to on-policy distillation under reinforcement learning with verifiable rewards. In both cases Evolve's review finds the first-order theory sound but the empirics limited — narrow test coverage and replay/evaluation overlap for JANUS, consistent single-run gains without seeds or confidence intervals for γOPD — yielding a reproducibility mean of 55 across the 2 reviewed papers, both rated medium readiness. The bibliometric snapshot records 17 distinct coauthors, including recurring collaborators such as Letian Tao, Jingliang Duan, Wei Xiong, Kehua Sheng and Bo Zhang, with a collaboration score of 58. The field distribution splits evenly between machine learning (LLM post-training) and machine learning (LLM fine-tuning and continual learning), while the focus score is 0 and the output, impact and activity axes lack sufficient data to score. The result is an early-stage, review-based picture of a senior author working at the methodological frontier of LLM post-training.
Collaboration pattern
The snapshot records 17 distinct coauthors and a collaboration score of 58. Recurring names include Letian Tao, Jingliang Duan, Wei Xiong, Kehua Sheng and Bo Zhang, alongside Yang Guan, Shiqi Liu, Zeyu He, Guojian Zhan, Jiaxin Gao and Feihong Zhang. Both reviewed papers place this researcher in the last author slot, suggesting a senior supervisory role on the cataloged work.
Research keywords
Highlights
JANUS: post-hoc Jacobian null-space rectification against catastrophic forgetting
γOPD: reward-compatible temporal credit assignment for on-policy distillation
Timeline
- 2026Last-position author on "Past, Future, All at Once" (JANUS), proposing post-hoc Jacobian null-space rectification for continual LLM fine-tuning (arXiv).
- 2026Last-position author on "Beyond Token-Local Imitation" (γOPD), adding reward-compatible temporal credit assignment to on-policy distillation (arXiv).
- 2026Both catalog papers fall in this single year, spanning LLM post-training and fine-tuning/continual learning; no earlier years are present in the provided data.
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.
- Radar dimension scores come from the platform's bibliometric snapshot; output, impact and activity axes lack sufficient data and are left unscored.
- Only 2 papers have been reviewed in the Evolve catalog; the narrative reflects those reviews plus the snapshot, not the researcher's complete publication record.
- No citation counts, h-index, venue list, country or ORCID were provided, and none are inferred here.
- Reproducibility (mean 55) is computed from just 2 papers, and the verdicts note single-run results, no seeds or confidence intervals, and narrow test coverage — treat as preliminary.