ML researcher in sparse attention, LLM agents, and reasoning post-training, with steeply rising recent output.
3
papers here
3
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
19
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—
Not yet scored — the snapshot lacks a verified aggregate of total works and venue spread. This axis would measure cumulative publication volume and its distribution across outlets and years.
Impact—
Not yet scored — citation aggregates for this axis are absent from the snapshot beyond the headline h-index (19) and i10-index (28). It would measure citation accumulation, per-paper influence, and uptake over time.
Focus0/100
Scores 0 because activity spreads evenly across three fields: the field distribution assigns 0.33 each to efficient long-context language models / attention sparsification, LLM agents and reinforcement learning, and LLM reasoning post-training. Topic counts reinforce the breadth, spanning topic modeling (113), NLP techniques (99), multimodal ML (31), reinforcement learning in robotics (30), and speech recognition and synthesis (26).
Collaboration56/100
Scores 56. The recent three-year window lists 16 distinct coauthors against 103 works — a small, heavily reused pool relative to volume. Twelve frequent coauthors anchor the network, and on all three Evolve-cataloged 2026 papers Mi is a middle author rather than corresponding, indicating contribution within larger team efforts.
Activity100/100
Perfect 100. The countsByYear record shows 20 works in 2024, 34 in 2025, and 49 in 2026 — 103 works in the recent three-year window, far above earlier phases (a peak of 11 in 2016 and 1 in 2006). The curve also shows a recorded 2017–2020 gap before restarting with 4 works in 2021.
Reproducibility55/100
Scores 55, computed from 3 assessed papers (reproMean 55). All three Evolve-cataloged 2026 preprints are rated medium reproducibility readiness. The verdicts consistently note rigor gaps: tight-budget wins within evaluation noise for SAS, no confidence intervals plus test-set checkpoint selection for T1, and small margins without significance tests for FlowBalance.
Research summary
Haitao Mi is a machine-learning researcher at Bellevue College (US) whose OpenAlex-resolved profile shows an h-index of 19 and an i10-index of 28. The snapshot records 103 works over the recent three-year window — 20 in 2024, 34 in 2025, and 49 in 2026 — a sharp acceleration after a recorded 2017–2020 gap, earning a perfect activity score of 100. Topical footprints concentrate on topic modeling (113 records) and natural language processing techniques (99), with secondary clusters in multimodal machine learning (31), reinforcement learning in robotics (30), and speech recognition and synthesis (26). Evolve has reviewed three 2026 arXiv preprints on which Mi is a middle author: SAS on end-to-end sparse attention, T1 on terminal-agent reinforcement learning for long-horizon tasks, and FlowBalance on verifier-grounded self-improvement for LLM reasoning. Evolve's verdicts credit workable recipes — log-space gates beating attention distillation, dense verifier rewards lifting a 122B MoE to 64.0% on Terminal-Bench 2.1, and verifier-gated gains over FlowRL/GRPO — while flagging tight margins without confidence intervals or significance tests. All three carry medium reproducibility readiness, consistent with a reproducibility score of 55 across 3 assessed papers. Collaboration scores 56, resting on 16 distinct recent coauthors.
Collaboration pattern
The recent three-year window pairs 16 distinct coauthors with 103 works, so the collaborator pool is small and heavily reused. Twelve frequent coauthors anchor the network (Leowei Liang, Zhiwei Li, Lei Zhu, Hao Gu, Xiang Hu, Yan Wang, Sirui Han, Leo Liang, Zhijiang Guo, Junyao Yang, Yucheng Shi, Zhongzhi Li). On all three cataloged 2026 papers Mi is a middle author and not corresponding, consistent with team-based deep-learning projects.
Research keywords
Highlights
SAS: training sparse-attention selectors end-to-end
T1: terminal-agent RL for long-horizon tasks
FlowBalance: verifier-grounded self-improvement
Timeline
- 2006Earliest recorded output in the snapshot: 1 work.
- 201611 works recorded — the high point of the first recorded phase.
- 2021Recorded output resumes after the 2017–2020 gap, with 4 works.
- 2024Output accelerates to 20 recorded works.
- 202534 works recorded, continuing the upward climb.
- 202649 works recorded — the highest year in the snapshot; three arXiv preprints (SAS, T1, FlowBalance) enter the Evolve catalog, all with middle authorship.
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 are OpenAlex-derived only; no ORCID is linked, so the profile is name-resolved from public bibliographic data and may include homonymous records.
- Output and impact dimensions are unscored: the snapshot lacks the citation and volume aggregates those axes require.
- All three cataloged papers are 2026 arXiv preprints; peer-reviewed venue status is outside the reviewed evidence.
- countsByYear omits 2012 and 2017–2020; those gaps reflect snapshot coverage, not verified inactivity.