Qi He

Microsoft (United States)

Unclaimed · AI-structured profileOpenAlex

Microsoft researcher working across LLM agents and omni-modal foundation models for physical-world understanding.

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-08) 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—

Insufficient data to score. This axis would measure publication volume and productivity over time, but the catalog currently holds only two 2026 arXiv papers, too few for a reliable output estimate.

Impact—

Insufficient data to score. Impact would be measured through citations, downloads, and uptake of the two benchmarks, but no citation or usage metrics are available yet in this profile.

Focus0/100

Focus currently scores 0/100, and the underlying field distribution explains why: the reviewed output splits evenly at 0.5/0.5 between multimodal machine learning (omni-modal foundation models, physical-world understanding) and artificial intelligence (LLM agents). The two catalogued 2026 papers sit one in each field, so the present record reads as deliberate breadth across two adjacent directions rather than concentration; a single-field trajectory would need more same-field papers to register.

Collaboration59/100

Collaboration scores 59/100, computed from 18 distinct coauthors, of whom 12 recur as frequent collaborators, including Yizhou Liu, Jinghang Han, and Kaixiang Qiu. Both catalogued papers are multi-author efforts in which Qi He holds a middle authorship position rather than first or corresponding author, consistent with team-based industrial research at Microsoft; the moderate score reflects a fairly wide but not yet leading collaboration footprint.

Activity—

Insufficient data to score. Activity would track publishing cadence and recent output trends, but both known papers fall in 2026, leaving no multi-year signal to estimate sustained activity.

Reproducibility78/100

Reproducibility scores 78/100, the mean across two scored papers. The Tasteful Agent rates high in reproducibility readiness, backed by its 502-fork benchmark and careful protocol, while OmniFysics-Nano-V2 rates medium: its 4B model and two-stage GRPO recipe are described, but Evolve flags untested top omni rivals and two self-published benchmarks. Overall the record suggests above-average attention to benchmark construction and evaluation transparency.

Research summary

Qi He is a researcher at Microsoft (United States) whose reviewed output spans two complementary fronts of modern AI. The first is agentic decision-making: 'The Tasteful Agent' (2026, arXiv) introduces Taste-Bench, a benchmark of 502 automatically mined decision forks labeled by hindsight outcomes, paired with a careful evaluation protocol and distillation gains that improve agent taste on long-horizon tasks; Evolve rates its reproducibility readiness as high while flagging residual label-validity gaps. The second is omni-modal foundation models: the OmniFysics-Nano-V2 technical report (2026, arXiv) describes a 4B-parameter model built on physics-aware data construction and a two-stage GRPO pipeline, claiming wins on 17 of 21 benchmarks, though Evolve notes that top omni rivals went untested and two benchmarks were self-published, with only medium reproducibility readiness. Bibliometrically the profile is deliberately described as early: two 2026 papers with middle authorship, 18 distinct coauthors giving a collaboration score of 59, an even 50/50 field split yielding a focus score of 0, and a reproducibility score of 78 across two scored papers, while output, impact, and activity axes remain unscored for lack of data.

Collaboration pattern

Qi He's reviewed work is team-produced: both 2026 papers carry middle authorship with no corresponding role, and the bibliometric record lists 18 distinct coauthors, 12 of whom recur as frequent collaborators (e.g., Yizhou Liu, Jinghang Han, Kaixiang Qiu, Dingkang Yang). The collaboration score of 59/100 reflects a fairly broad but mid-tier network. The pattern fits collaborative industrial research at Microsoft (United States), where Qi He contributes to benchmarks and model training rather than leading sole-author projects.

Research keywords

LLM agentsbenchmark evaluationdecision-makinglong-horizon taskshindsight labelingdistillationomni-modal modelphysical world understandingphysics-aware data constructionGRPOreinforcement learning curriculumreward-diversity filtering

Highlights

Taste-Bench: 502 hindsight-labeled decision forks for measuring agent taste
In 'The Tasteful Agent' (2026, arXiv), Qi He and coauthors automatically mine 502 decision forks labeled by hindsight outcomes, pair them with a careful evaluation protocol, and report distillation gains on long-horizon tasks; Evolve rates reproducibility readiness high while flagging remaining label-validity gaps.
OmniFysics-Nano-V2: a 4B omni-modal model for physical-world understanding
The 2026 technical report trains a 4B-parameter omni-modal model with physics-aware data construction and a two-stage GRPO curriculum using reward-diversity filtering, claiming wins on 17 of 21 benchmarks; Evolve notes top omni rivals were untested and two benchmarks self-published, with medium reproducibility readiness.
Benchmark-first methodology with above-average reproducibility
Across both reviewed papers, the common signature is benchmark- and data-centric methodology — hindsight-labeled decision forks on one side, physics-aware data pipelines on the other — yielding a reproducibility score of 78/100 over two scored papers, with readiness levels of high and medium respectively.

Timeline

2026Released the OmniFysics-Nano-V2 technical report on arXiv: a 4B omni-modal model trained with physics-aware data and two-stage GRPO, claiming wins on 17 of 21 benchmarks.
2026Released 'The Tasteful Agent' on arXiv, introducing Taste-Bench with 502 auto-mined, hindsight-labeled decision forks and distillation gains on long-horizon agent tasks.
2026Both papers entered the Evolve catalog and were reviewed, forming an early profile with 18 distinct coauthors and a 78/100 reproducibility score across two scored papers.

Papers in the Evolve catalog

Frequent co-authors

Yizhou Liu×1 Jinghang Han×1 Kaixiang Qiu×1 Minghao Han×1 Yue Jiang×1 Xujia Chen×1 Wei Zou×1 Shunli Wang×1 Lihua Zhang×1 Dingkang Yang×1 Wenbo Pan×1 Zhichao Liu×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.

Evidence base is narrow: only two reviewed 2026 arXiv papers feed this profile; output, impact, and activity dimensions are unscored because no citation counts, publication-volume, or multi-year cadence data were available.
Bibliometric scores (collaboration 59, focus 0, reproducibility 78) are computed from public bibliographic data under this name resolution and may shift as more papers are indexed.
'Qi He' is a common name; this profile is name-resolved from public bibliographic data (institution: Microsoft, United States) and the two catalogued arXiv papers, without ORCID confirmation.
Quantitative claims such as OmniFysics-Nano-V2's 17-of-21 benchmark wins are self-reported in the paper and flagged by Evolve; they have not been independently verified.