Jianfeng Zhang

—

Unclaimed · AI-structured profileOpenAlex

Machine learning researcher working across LLM post-training with reinforcement learning and multimodal AI for physics.

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-02) 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: publication counts, venue distribution, and productivity-over-time data are unavailable in this snapshot. This axis would measure total research output — how many papers, in which venues, over time — so no number is estimated here.

Impact—

Not yet scored: no citation counts or citation-derived aggregates are available for this researcher. This axis would measure influence — citations to the two 2026 preprints and any earlier work — and cannot be computed from the current input.

Focus0/100

Scored 0/100: the field distribution is an even 50%/50% split between LLM post-training/reinforcement learning and AI-for-physics/scientific multimodal learning, so neither strand dominates the reviewed portfolio. The two cataloged papers sit in these separate strands — DCRL in RL for LLM reasoning and PhyMo in physical-field multimodality — indicating breadth rather than concentration at this early stage.

Collaboration53/100

Scored 53/100 from 13 distinct coauthors across the two reviewed papers. Zhang appears as a middle author and not as corresponding author on either paper, consistent with contributing within larger teams. The frequent-coauthor list repeats Henan Sun, Haitao Hu, and Nuo Chen, alongside Jia Li, Zehua Li, Qifan Zhang, Jin Liu, and Lujia Pan.

Activity—

Not yet scored: the snapshot has no year-by-year publication record, so publishing cadence cannot be computed. This axis would measure tempo — papers per year and gaps between releases — beyond the two 2026 arXiv preprints reviewed here.

Reproducibility20/100

Scored 20/100, the mean over the 2 reviewed papers, each rated low reproducibility readiness. Evolve's review of PhyMo notes that no code was released, while the DCRL review flags missing key ablations alongside thin theory. Neither paper currently ships the artifacts needed for independent replication, which anchors the low score.

Research summary

Jianfeng Zhang is a machine learning researcher whose reviewed work splits evenly, at 50%/50%, between reinforcement learning for LLM post-training and multimodal AI for physics. Both papers in the Evolve catalog are 2026 arXiv preprints with Zhang as a middle author rather than corresponding author. DCRL decouples and couples reinforcement learning via policy-reward manifold alignment, co-evolving the policy with evolving rubrics through shared weights; the review reports a 4B model beating a 32B model while flagging thin theory and missing key ablations. PhyMo casts PDE-structured physical fields as a new modality for multimodal AI4Physics, with gains across multiple datasets including solar forecasting, though margins are thin, the theory is trivial, and no code was released. Across the two papers Zhang worked with 13 distinct coauthors (collaboration score 53/100); Henan Sun, Haitao Hu, and Nuo Chen recur in the frequent-coauthor list, alongside Jia Li, Zehua Li, Qifan Zhang, Jin Liu, and Lujia Pan. Reproducibility is scored 20/100, the mean over the 2 reviewed papers, both rated low readiness. Output, impact, and activity lack sufficient data and are not scored. As the name is common, this profile is name-resolved from public bibliographic data.

Collaboration pattern

Across the two reviewed papers Zhang collaborated with 13 distinct coauthors (collaboration score 53/100), always as a middle author and never as corresponding author. The frequent-coauthor list repeats Henan Sun, Haitao Hu, and Nuo Chen, and also includes Jia Li, Zehua Li, Qifan Zhang, Jin Liu, and Lujia Pan — a recurring circle consistent with team-based work in a stable group.

Research keywords

LLM reasoningReinforcement learningReward modelingReward hackingPolicy-reward alignmentManifold geometryMultimodal learningAI for PhysicsPhysics-informed machine learningPDE residual supervisionSolar forecastingPhysical-field representation

Highlights

DCRL: a 4B model beats 32B via policy-reward co-evolution
Evolve's review describes DCRL as decoupling and coupling reinforcement learning through policy-reward manifold alignment: policy and reward co-evolve via shared weights and evolving rubrics. The headline reported result is a 4B model outperforming a 32B model, though the review flags thin theory and missing key ablations.
PhyMo: PDE-structured physical fields as a new multimodal modality
PhyMo proposes physical fields governed by PDEs as a first-class modality for multimodal AI4Physics, reporting gains across multiple datasets including solar forecasting. Evolve calls the direction promising but notes thin margins, trivial theory, and no released code.
Both 2026 preprints rated low reproducibility readiness
Across the 2 reviewed papers the reproducibility score is 20/100, with each paper rated low readiness: DCRL lacks key ablations and PhyMo released no code, so neither currently supports independent replication.

Timeline

2026DCRL, a reinforcement learning method for LLM reasoning built on policy-reward manifold alignment, appears as an arXiv preprint with Zhang as a middle author.
2026PhyMo, proposing physical fields as a new modality for multimodal AI4Physics, appears as an arXiv preprint, again with Zhang as a middle author.
2026Evolve reviews both preprints, rating each low on reproducibility readiness; the reproducibility score averages 20/100 across the 2 papers.

Papers in the Evolve catalog

Frequent co-authors

Henan Sun×2 Haitao Hu×2 Nuo Chen×2 Jia Li×2 Zehua Li×1 Qifan Zhang×1 Jin Liu×1 Lujia Pan×1 Henan Sun×1 Haitao Hu×1 Jianfeng Zhang×1 Nuo Chen×1

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 bibliometric data; the output, impact, and activity axes lack underlying data and are not scored.
'Jianfeng Zhang' is a common name; this profile is name-resolved from public bibliographic data, and institution, country, and ORCID are not on file.
Coverage is limited to the two 2026 arXiv preprints reviewed by Evolve; other publications may exist but fall outside this snapshot.
The reproducibility score (20/100) rests on only 2 papers; reviews note missing key ablations for DCRL and no released code for PhyMo.