Li Wei

—

Unclaimed · AI-structured profileORCID OpenAlex

Machine learning researcher across production recommender systems and domain-specific LLM post-training

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—

Data insufficient: beyond two 2026 arXiv papers, no publication counts, venue history or per-year totals are available, so this axis is unscored. It would measure the volume and consistency of published output over time.

Impact—

Data insufficient: no citation counts, h-index or downstream adoption signals are captured for this profile yet, so the axis is unscored. It would measure how far the two reviewed works influence subsequent research and industrial practice.

Focus0/100

The focus score is 0, and the raw field distribution explains it: an exact 50/50 split between recommender systems (industrial machine learning systems) and natural language processing / large language models, with one 2026 catalog paper in each field. This reads as a deliberate two-track portfolio — production ranking systems on one side, domain-LLM post-training on the other — although two papers are too few to establish depth within either track.

Collaboration59/100

Collaboration scores 59, built on 18 distinct coauthors across the reviewed work. Eleven names recur on the frequent-coauthor list, with Sanjay Surendranath Girija the only one entered twice; other recurring partners include Aniruddh Nath, Yanhao Jiang, Shawn Andrews, Lukasz Heldt, Aditya Mahajan and Yi Wu. Both 2026 papers are middle-author, non-corresponding contributions, consistent with large-team industrial and LLM research rather than solo-led projects.

Activity—

Data insufficient: the catalog only spans 2026, so publication cadence and recent research tempo cannot be computed. This axis would measure how actively the researcher publishes and engages with the field year over year.

Reproducibility38/100

Reproducibility averages 38 across the 2 papers scored. 'Lightweight Ranking Heads' is rated low readiness, with Evolve's review calling its offline statistics thin; 'TelecomGPT-R1' reaches medium readiness on the strength of solid ablations, though contamination and evaluation-setting caveats persist. Neither paper attains high readiness, keeping this axis below the midpoint despite the telecom work's stronger ablation coverage.

Research summary

Li Wei (name-resolved from public bibliographic data via ORCID 0009-0008-9321-3983; institution and country are not recorded in the snapshot) is an early-visibility machine learning researcher whose Evolve catalog spans two 2026 arXiv papers in equally weighted fields: recommender systems (industrial machine learning systems) and natural language processing / large language models, each at 50% of the field distribution. In both papers Li Wei is a middle-author, non-corresponding contributor, consistent with large-team industrial research; the collaboration score is 59, built on 18 distinct coauthors with 11 recurring partners, Sanjay Surendranath Girija listed twice. On the systems track, 'Lightweight Ranking Heads' proposes frozen-backbone adapter heads that Evolve's review credits with cutting multi-task experiment cycles in a production recommender from 24 to 11 days, while noting the offline statistics are thin. On the language-model track, 'TelecomGPT-R1' applies unified post-training — reinforcement learning with verifiable rewards (DAPO) and chain-of-thought distillation — to heterogeneous telecom reasoning, with a claimed 89.6% mean on GSMA Open Telco that the review reports ahead of GPT-5 and Claude, alongside contamination and evaluation-setting caveats. Reproducibility averages 38 across the two papers (one low, one medium readiness). Output, impact and activity axes lack sufficient data and are unscored; the visible record is limited to two arXiv preprints dated 2026.

Collaboration pattern

Li Wei works in sizeable teams: 18 distinct coauthors appear across the reviewed papers, and 11 names recur as frequent collaborators, with Sanjay Surendranath Girija listed twice (the only repeat). Other recurring partners include Aniruddh Nath, Yanhao Jiang, Shawn Andrews, Lukasz Heldt, Yi Wu, Aditya Mahajan, Mohit Sharma, Bohao Wang, Chenwei Wu and Hang Zou. Both 2026 papers list Li Wei as a middle author without corresponding status, matching a contributor role in industrial-scale research groups rather than lab leadership.

Research keywords

Recommender systemsMulti-task learningRanking modelsAdapter headsStop-gradientsContinual learningLLM post-trainingReinforcement learning with verifiable rewardsDAPODomain-specific LLMsChain-of-thought distillationTelecommunications

Highlights

Cutting multi-task experiment cycles from 24 to 11 days in a production recommender
Evolve's review of 'Lightweight Ranking Heads' describes frozen-backbone adapter heads ('Light Heads') that compress multi-task experiment cycles in a production recommender from 24 to 11 days. The review calls it a compelling systems paper while noting the offline statistics are thin; reproducibility readiness is rated low.
Telecom reasoner claiming 89.6% mean on GSMA Open Telco
Evolve's review of 'TelecomGPT-R1' reports a unified telecom reasoner claiming an 89.6% mean on GSMA Open Telco, ahead of GPT-5 and Claude, built with unified post-training, reinforcement learning with verifiable rewards (DAPO) and chain-of-thought distillation. Ablations are solid, but contamination and evaluation-setting caveats persist; reproducibility readiness is rated medium.

Timeline

2026Co-authored 'Lightweight Ranking Heads: Accelerating Multi-Task Experimentation in Production Recommender Systems' (arXiv) as a middle author; Evolve's review credits the frozen-backbone adapter heads with cutting multi-task experiment cycles from 24 to 11 days.
2026Co-authored 'TelecomGPT-R1: Unified Post-Training for Reasoning Across Heterogeneous Telecom Tasks' (arXiv), unifying telecom reasoning via reinforcement learning with verifiable rewards (DAPO) and chain-of-thought distillation; Evolve cites a claimed 89.6% mean on GSMA Open Telco.
2026Both catalog entries fall within 2026, so the observable publication window is a single year; longer-term output tempo and career trajectory cannot yet be established from the data.

Papers in the Evolve catalog

Frequent co-authors

Sanjay Surendranath Girija×1 Aniruddh Nath×1 Yanhao Jiang×1 Shawn Andrews×1 Lukasz Heldt×1 Yi Wu×1 Aditya Mahajan×1 Mohit Sharma×1 Sanjay Surendranath Girija×1 Bohao Wang×1 Chenwei Wu×1 Hang Zou×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.

Identity is name-resolved from public bibliographic data (ORCID 0009-0008-9321-3983); 'Li Wei' is a very common name whose Chinese character form is not confirmed, and institution and country are not recorded in the snapshot.
All bibliometrics derive from OpenAlex-sourced records and the two Evolve-reviewed arXiv papers only; the output, impact and activity axes are unscored for lack of data.
The visible record covers only two 2026 arXiv preprints, both middle-author and non-corresponding; no peer-reviewed venue or citation information is in scope.
Quantitative claims (24→11 days; 89.6% mean on GSMA Open Telco) are quoted from Evolve review verdicts, not independently verified.