Machine learning researcher at Xiaopeng Motors spanning LLM agent harnesses, speech-LLM compression, and bandit theory
3
papers here
3
reviewed by Evolve
1
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—
Output metrics are not yet available: this snapshot has no publication count, venue history, or yearly volume. This axis would measure the scale and consistency of research production over time, which three 2026 preprints alone cannot establish.
Impact—
Impact data is not yet available: no citation counts or downstream-usage signals exist in this snapshot. The axis would measure how widely the work is cited and built upon by other researchers, which cannot be estimated for 2026 preprints.
Focus0/100
The focus score is 0, and the underlying field distribution explains why: the cataloged output splits evenly across three fields — AI for science (LLM agent harnesses and benchmarks), machine learning (online learning / bandit algorithms), and speech processing (speech LLM efficiency) — at one third each. No field dominates, so this profile reads as deliberately broad rather than narrowly specialised, spanning agent systems, compression engineering, and theory.
Collaboration67/100
Collaboration scores 67. Across the three cataloged papers, Hao Li has worked with 28 distinct coauthors, and twelve recurring collaborators appear in the network, including Gongbo Zhang, Yu Wang, Mujie Lin, Liuzhenghao Lv, and Jun Zhu. Authorship position varies: first author on Meta-LinEXP3, middle author on OpenAI4S and X-AuT, with no corresponding-author role recorded in the catalog — a mid-network, team-based collaboration pattern.
Activity—
Activity metrics are not yet available: the snapshot lacks year-by-year publication records, so recent cadence cannot be plotted. This axis would measure publishing rhythm and ongoing engagement over successive years, which a single publication year (2026) cannot reveal.
Reproducibility43/100
Reproducibility averages 43 across the three assessed papers, with readiness rated medium (OpenAI4S), low (X-AuT), and medium (Meta-LinEXP3). Reviews flag concrete gaps: X-AuT relies on a single random seed and omits key controls despite transparent reporting; OpenAI4S's reported gains (7.83 vs 5.79–6.36) are partly built into its own evaluator; Meta-LinEXP3's experiments are modest but honest. Methods are described, yet verification headroom remains.
Research summary
Hao Li is a machine learning researcher affiliated with Xiaopeng Motors (China), whose visible publication record — as captured in Evolve's reviewed catalog — consists of three 2026 arXiv preprints spanning three distinct areas: LLM agent harnesses for science, efficiency in speech large language models, and online learning theory. This even three-way split (each field accounting for one third of the cataloged output) yields a focus score of 0, marking Li as a breadth-oriented contributor rather than a single-track specialist. Li first-authored Meta-LinEXP3, which extends LinEXP3 to adversarial linear contextual bandit meta-learning with Lean-checked O(√n)/O(n^2/3) regret bounds — strong theory backed by modest but honest experiments. On OpenAI4S and X-AuT, Li appears as a middle author. Across the catalog, Li has worked with 28 distinct coauthors, including recurring collaborators such as Gongbo Zhang, Yu Wang, and Mujie Lin, supporting a collaboration score of 67. Reproducibility averages 43 across the three assessed papers, with medium, low, and medium readiness; reviews note single-seed experiments, missing controls, and evaluator overlap as recurring transparency concerns. Citation-based output, impact, and activity metrics are not yet available in this snapshot.
Collaboration pattern
Hao Li's cataloged work is team-based: 28 distinct coauthors across three 2026 papers, with a recurring core of twelve collaborators including Gongbo Zhang, Yu Wang, Mujie Lin, Liuzhenghao Lv, Yicheng Mao, and Jun Zhu. Authorship sits first on Meta-LinEXP3 and middle on OpenAI4S and X-AuT, with no corresponding-author role recorded — consistent with a contributor moving between theory, agents, and speech-efficiency teams. Collaboration scores 67.
Research keywords
Highlights
Lean-verified regret bounds for adversarial linear contextual bandits
Code-as-action agent harness benchmarked against Claude Code
Progressive audio-encoder pruning for speech LLMs
Timeline
- 2026First-author paper Meta-LinEXP3 posts Lean-checked O(√n)/O(n^2/3) regret bounds for adversarial linear contextual bandit meta-learning.
- 2026Co-authors OpenAI4S, an agent harness for science sessions reporting 7.83 vs 5.79–6.36 over 36 scenarios against Claude Code.
- 2026Co-authors X-AuT, pruning the Qwen3-ASR audio encoder from 18 to 14 layers with cross-scale distillation.
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.
- Radar scores are computed from public bibliometric data; output, impact, and activity axes lack underlying metrics and are left unscored.
- 'Hao Li' is a common name and no ORCID is provided; this profile is name-resolved from public bibliographic data and may merge or omit records.
- All three cataloged papers are 2026 arXiv preprints reviewed by Evolve; no peer-reviewed venue record appears in this snapshot.
- Reproducibility scores (mean 43 over three papers) reflect Evolve review judgments, not independent replication.