Haitao Hu

The Hong Kong University of Science and Technology (Guangzhou)

Unclaimed · AI-structured profileOpenAlex

ML researcher at HKUST(GZ) spanning RL-based LLM post-training and physical-field multimodal learning 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 scored: beyond the two reviewed 2026 arXiv preprints, this snapshot holds no publication counts, venue tiers or authorship-volume data. The axis would measure aggregate production — outputs per year, publication venues and author-role mix — once a fuller bibliographic record exists.

Impact—

No citation counts, h-index or adoption signals exist in this snapshot, and both reviewed papers are recent preprints. The axis would measure scholarly influence — citations, reuse and uptake by other groups — and stays blank rather than estimated.

Focus0/100

Scores 0, the lowest reading, and the raw field distribution explains it: the two reviewed papers split exactly 50/50 between machine learning for LLM post-training with reinforcement learning, and machine learning for physics with scientific multimodal learning. No single topic dominates yet, so the score flags dispersed concentration rather than weak work; it reads as a methods-first toolkit applied to two problem domains instead of a settled specialisation.

Collaboration53/100

Scores 53 on the 0-100 collaboration scale, computed from 13 distinct coauthors across the two reviewed papers. Henan Sun, Jianfeng Zhang and Nuo Chen recur across both entries, forming a small stable core, while Jia Li, Zehua Li, Qifan Zhang, Jin Liu and Lujia Pan each appear once. With middle, non-corresponding positions on both papers, the pattern fits a contributing member of group-led teams rather than a solo or PI-led line of work.

Activity—

Not scored: no publication cadence, first-publication year or recent-volume indicators are available beyond the two 2026 entries. The axis would measure research tempo — outputs per interval and continuity over time — which cannot be estimated from the current record.

Reproducibility20/100

Scores 20, the mean readiness Evolve assigns across the 2 reproducibility-assessed papers, and both are individually rated low. The verdicts name concrete gaps: DCRL is missing key ablations, and PhyMo ships no code alongside thin margins and trivial theory. Until artifacts or fuller ablations appear, third parties should treat the reported gains as unverified; the score measures artifact availability, not the implausibility of the results.

Research summary

Haitao Hu is affiliated with The Hong Kong University of Science and Technology (Guangzhou). This profile is name-resolved from public bibliographic data and rests on the two papers Evolve has reviewed to date, both 2026 arXiv preprints on which Hu appears as a middle author. His reviewed work spans two subfields of machine learning in an even split: LLM post-training with reinforcement learning, and AI for physics with scientific multimodal learning. On the first line, DCRL co-evolves policy and reward through shared weights and evolving rubrics, with a 4B model outperforming a 32B one; Evolve notes the theory is thin and key ablations are missing. On the second, PhyMo proposes PDE-structured physical fields as a new multimodal modality for AI4Physics, showing multi-dataset gains with thin margins, trivial theory and no released code. Both papers are rated low on reproducibility readiness, giving a reproducibility mean of 20 across the 2 reviewed papers. The collaboration footprint covers 13 distinct coauthors, with Henan Sun, Jianfeng Zhang and Nuo Chen recurring, and a collaboration score of 53. The even topical split across the two subfields yields a focus score of 0. Output, impact and activity metrics are not yet available and remain unscored.

Collaboration pattern

Haitao Hu's reviewed record connects to 13 distinct coauthors. A stable core recurs: Henan Sun, Jianfeng Zhang and Nuo Chen appear in both entries, while Jia Li, Zehua Li, Qifan Zhang, Jin Liu and Lujia Pan each appear once. Holding middle, non-corresponding authorships on both papers, Hu fits the profile of a contributing team member working inside group-led projects that span two subfields — reinforcement-learning-based LLM post-training and AI for physics.

Research keywords

LLM reasoningreinforcement learningreward modelingreward hackingpolicy-reward alignmentmanifold geometrymultimodal learningAI for Physicsphysics-informed machine learningPDE residual supervisionsolar forecastingphysical-field representation

Highlights

DCRL: co-evolving policy and reward, a 4B model beats 32B
DCRL decouples and couples reinforcement learning via policy-reward manifold alignment: shared weights and evolving rubrics let the policy and reward model co-evolve, curbing reward hacking. Evolve credits the headline result — a 4B model beating a 32B one — but finds the theory thin and key ablations missing; reproducibility readiness is rated low.
PhyMo: physical fields as a new multimodal modality for AI4Physics
PhyMo proposes PDE-structured physical fields as a first-class modality for multimodal AI4Physics, using PDE residual supervision; solar forecasting is among the application keywords. Evolve sees promising multi-dataset gains but thin margins, trivial theory and no released code, with low reproducibility readiness.

Timeline

2026DCRL preprint posted to arXiv: decouples and couples reinforcement learning via policy-reward manifold alignment, with Hu in a middle author position.
2026PhyMo preprint posted to arXiv, proposing physical fields as a new modality for multimodal AI4Physics.
2026Both preprints enter the Evolve review catalog with low reproducibility readiness and a combined reproducibility mean of 20.

Papers in the Evolve catalog

Frequent co-authors

Henan Sun×2 Jianfeng Zhang×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 and raw counts come only from the Evolve snapshot and cover just the two reviewed papers; they should not be read as career totals.
'Haitao Hu' is a common name: the profile is name-resolved from public bibliographic data, with no ORCID or country available to confirm identity.
Output, impact and activity axes lack underlying data — no citation, productivity or cadence metrics — and are left unscored rather than estimated.
Both reviewed papers are 2026 arXiv preprints rather than peer-reviewed publications, and the Chinese-character spelling of the name is unverified.