First-author researcher at the intersection of LLM post-training (RL) and autonomous agent systems engineering
2
papers here
2
reviewed by Evolve
2
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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-volume metrics are available yet. This axis would measure total and per-year publication counts and venue spread; the two cataloged 2026 arXiv preprints are too small a base to score it.
Impact—
No citation or influence data exists in the current record. This axis would measure citations, downstream adoption, and references to the work; both cataloged papers are recent 2026 preprints, so no impact signal can be scored yet.
Focus0/100
The focus dimension scores 0, and the raw field distribution explains why: the cataloged work splits evenly — 0.5 machine learning (LLM post-training and reinforcement learning) and 0.5 AI systems engineering (autonomous agent orchestration). This is a two-track profile rather than single-field concentration; both 2026 first-author preprints sit on opposite sides of that split, so no single-field dominance is measurable yet.
Collaboration32/100
The collaboration dimension scores 32, drawn from a raw count of 4 distinct coauthors across the cataloged record — Jiexiong Liu, Yixuan Chen, Chi Hong, and Ziheng Huang. Both 2026 papers carry Cheng Li as first author without corresponding-author status, consistent with compact-team work in which this researcher leads the writing and technical core within a small, stable collaborator set.
Activity—
No activity metrics (publication cadence over time, per-year volume, or recent-venue record) are available in this snapshot. The axis would measure research tempo and continuity; with only two 2026 preprints cataloged, there is not enough time-series data to score it.
Reproducibility20/100
The reproducibility dimension scores 20, a reproMean of 20 averaged over 2 scored papers — both flagged low reproducibility readiness in Evolve's reviews. The incremental-execution paper reports n=1 case studies with no released code, and the post-training paper's headline +3.0 gain rests on single runs below the noise floor; neither ships an artifact supporting independent replication yet.
Research summary
Cheng Li is an early-stage researcher whose Evolve-cataloged record consists of two first-author arXiv preprints posted in 2026, split evenly — 0.5 and 0.5 in the raw field distribution — between machine learning (LLM post-training and reinforcement learning) and AI systems engineering (autonomous agent orchestration). The first, "Train What You Deploy: Token-Faithful Post-Training of a Production Coding", proposes C-DPPO, a certified trust-region post-training method built on total variation divergence, keeping training aligned with how the deployed model emits tokens; Evolve judges the certification theory sound but notes the headline +3.0 gain rests on single runs below the noise floor. The second, "Incremental Consistency Execution for Autonomous Intelligent Systems", introduces perturbation-built invariant domains, task fact contracts, and field-level dependency masks to bound change propagation in LLM agent pipelines; Evolve calls the design promising but flags n=1 case studies, no released code, and misaligned citations. Both papers are rated low reproducibility readiness, yielding a reproducibility score of 20 (reproMean 20 over 2 papers). Collaboration is compact: 4 distinct coauthors — Jiexiong Liu, Yixuan Chen, Chi Hong, Ziheng Huang — for a collaboration score of 32, with Cheng Li first author on both preprints and no corresponding-author record. This is a promising but early-stage, name-resolved profile.
Collaboration pattern
Evolve's catalog shows a compact collaboration pattern: 4 distinct coauthors — Jiexiong Liu, Yixuan Chen, Chi Hong, and Ziheng Huang — behind both 2026 papers, giving a collaboration dimension score of 32. Cheng Li takes first-author position on both preprints without corresponding-author status, consistent with leading the writing and technical core of a small, stable team rather than supervising a large lab network.
Research keywords
Highlights
Certified trust-region post-training for production coding agents (C-DPPO)
Invariant-domain incremental re-execution for agent pipelines
Two-track first-author output in a single year
Both cataloged 2026 papers are first-author arXiv preprints splitting evenly across two fields — LLM post-training/RL and autonomous agent orchestration (field distribution 0.5/0.5) — though both are currently rated low reproducibility readiness.
Timeline
- 2026Published first-author arXiv preprint "Train What You Deploy: Token-Faithful Post-Training of a Production Coding" in machine learning — LLM post-training and reinforcement learning.
- 2026Published first-author arXiv preprint "Incremental Consistency Execution for Autonomous Intelligent Systems" in AI systems engineering (autonomous agent orchestration).
- 2026Cataloged record reaches two first-author papers split evenly across two fields (0.5/0.5), backed by 4 distinct coauthors.
- 2026Both papers reviewed by Evolve and rated low reproducibility readiness, giving a reproMean of 20 over 2 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.
- Bibliometrics and dimension scores are computed from the public bibliometric snapshot; no citation, venue, or output metrics are available for this profile yet.
- "Cheng Li" is a common name — this profile is name-resolved from public bibliographic records and may merge or omit same-name researchers.
- Only two 2026 arXiv preprints are in the Evolve catalog; institution, country, and ORCID are unrecorded, and the reproducibility average (reproMean 20) covers just 2 papers.
- Evolve verdicts reflect review-time assessments (single-run results, n=1 case studies, no released code) and may change if the authors release artifacts or revised versions.