LLM agent skills and continual-learning fine-tuning researcher; early-stage profile built from two 2026 arXiv preprints
2
papers here
0
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-09-28) 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 bibliometric snapshot does not provide publication counts or output-volume measures for this researcher. This axis would measure the quantity and cadence of papers over time, and no reliable number can be quoted today.
Impact—
Insufficient data: no citation counts, h-index, or venue-level impact metrics are available in the input snapshot. This axis would measure citations and scholarly influence; two brand-new arXiv preprints give no basis for an estimate.
Focus—
Not yet scored: the snapshot lacks the publication history and topic-distribution data this axis needs. It would measure how concentrated or dispersed the researcher's topics are across the two visible lines: LLM agents and continual learning.
Collaboration58/100
The collaboration axis scores 58/100, computed from a record of 17 distinct coauthors, with twelve recurring partners including Zhilong Zheng, Letian Tao, Yang Guan, and Shengbo Eben Li. Both catalog papers place this researcher in a middle author position on team-based arXiv preprints rather than as single or corresponding author, indicating steady but not yet leading collaborative activity.
Activity—
Insufficient data: apart from two 2026 preprints, no yearly publication counts or recency indicators exist. This axis would measure publishing pace and sustained engagement in recent years, which cannot be assessed from the current snapshot.
Reproducibility—
Unscored at researcher level: the radar lacks the metadata this axis aggregates. Per-paper signals do exist — both reviewed papers carry a medium reproducibility-readiness rating — but no overall score can be computed yet.
Research summary
Bo Zhang is an early-stage machine learning researcher whose documented output centers on making large language models more capable and more stable around adaptation time. The Evolve catalog links this name to two 2026 arXiv preprints, each with a middle authorship position: SkillGym, which internalizes human agent skills into LLMs by converting them into 2,756 verifiable training environments supporting supervised fine-tuning with verifiable rewards, and JANUS, a post-hoc Jacobian null-space rectification method that mitigates the stability-plasticity dilemma — catastrophic forgetting — in parameter-efficient LLM fine-tuning. Together they sketch a coherent agenda: skill acquisition through environment-scale training data on one side, weight-level correction that preserves prior capability on the other. Bibliometric data identifies 17 distinct coauthors, twelve of whom recur frequently — including Zhilong Zheng, Letian Tao, Yang Guan, and Shengbo Eben Li — producing a collaboration score of 58/100, the only fully scored radar axis; output, impact, focus, activity, and reproducibility remain unscored for lack of data. Evolve's reviews are cautiously positive: SkillGym's gains are substantial but tempered by harness confounds and an abstract-body number mismatch, while JANUS's first-order theory is judged sound but narrowly tested, with replay-evaluation overlap. Both papers carry medium reproducibility readiness. No institution, country, or ORCID is on record; Bo Zhang is a common name, so this profile is name-resolved from public bibliographic data.
Collaboration pattern
The record shows 17 distinct coauthors, with twelve recurring partners — among them Zhilong Zheng, Letian Tao, Yang Guan, Yujie Yang, Wei Xiong, and Shengbo Eben Li. Both catalog papers are middle-authored team preprints rather than solo or corresponding-author work, consistent with contributing to larger LLM-systems projects. The 58/100 collaboration score is the only radar axis with sufficient data.
Research keywords
Highlights
SkillGym: human skills turned into 2,756 verifiable agent training environments
JANUS: post-hoc Jacobian null-space rectification against catastrophic forgetting
Two complementary tracks: train-time skill environments, post-hoc weight repair
Read together, the two 2026 preprints frame LLM adaptation from opposite ends — building verifiable environments to internalize new skills, and rectifying fine-tuned weights to protect old ones. Both are middle-authored team efforts with medium reproducibility readiness.
Timeline
- 2026SkillGym preprint appears on arXiv, internalizing human agent skills into LLMs via 2,756 verifiable training environments.
- 2026JANUS preprint appears on arXiv, proposing post-hoc Jacobian null-space rectification to mitigate forgetting in LLM fine-tuning.
- 2026Both preprints enter the Evolve catalog with medium reproducibility readiness, forming a two-track early profile in LLM agents and 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 signals come from OpenAlex-based public bibliometric data only; just the collaboration axis (58/100, 17 distinct coauthors) had enough data to score.
- The reviewed catalog contains only two 2026 arXiv preprints; any broader publication record is outside this snapshot.
- No institution, country, or ORCID is on record; Bo Zhang is a common name, so this profile is name-resolved from public bibliographic data.
- Researcher-level reproducibility is unscored; the medium repro-readiness ratings apply to the two individual reviewed papers only.