Thomas Li

Principal Applied Scientist · Microsoft Edge
Architect, Edge Agent Memory
Email 250145013@qq.com Web thomasli.cn Location Suzhou, China
Experience
Microsoft · Edge
Apr 2022 – Present
Principal Applied Scientist (L66) Tech Lead Edge Architect Leader Team · AI browser / agentic experiences

Since 2022, led AI-browser and agentic experiences from 0→1: scaled Edge Copilot (the browser's side-panel AI assistant) from 0 to millions of DAU, and re-architected core features like Tabs with LLMs — driving 3× growth on a business line serving tens of millions of users.

2022 Shipped Tabs Auto Grouping — the industry's first LLM-based tab auto-grouping, ~12 months ahead of Google.
2023 Side-panel long-document summarization (contextual RAG).
2024 Built Browse-for-me, Edge's early agent prototype (search agent + automatic curation and summarization across the current page / Tabs / Favorites / History).
2025 Led end-to-end rebuild of Edge Copilot as a GUI / Vision / Voice agent (task success 76%→90%, action success 50%→90%, browser-use latency 10s→1s, first token 1s→241ms).
2026 Lead across three tracks — Intent / Journeys / Blueprints. Shipped Journeys (industry-first; turns long-term browsing history into visual task cards) and Blueprints (turns browsing activity into web apps / extensions).
Tencent Music
Sep 2018 – Apr 2022
Head of Social Recommendation Real-time social (live streaming / duet) + async social (relationship & family recommendation)
Owned algorithms, engineering, and team from 0→1.
Live-stream recsys DSSM softmax-in-batch + bitwise FM/AutoInt + CGC + PCGrad + GradNorm + ESMM. Follow penetration +17%, watch time +1.15%; doubled the number of active streamers.
User / family recsys Bipartite-graph GraphSAGE + GNN + Attention + SSL. Family-business D+1 retention +3.4%, follow penetration +100%.
Social infra Unified retrieval / feature / bandit / traffic-control services across 15 entry points and 4–5 verticals; overall penetration +10%.
NetEase Cloud Music
Nov 2017 – Sep 2018
Head of Social Recommendation Musician recommendation · Follow-feed ranking
Musician recsys Built "Musician Recommendation" from scratch (w2v similar-user retrieval + LR ranking); follow rate and follow count +100%.
Follow-feed ranking itemCF + GBDT ranking for the follow feed; play count +10%.
Alibaba · iDST
Apr 2015 – Nov 2017
Senior Algorithm Engineer Tianhe Program · Taobao Push
Tianhe Program Online matching of business logic (PID) with algorithms to guarantee delivery volume, holding CTR and display revenue flat.
Taobao Push End-to-end: data analysis, online delivery engineering, and ops-management platform; +100% vs. the prior audience-targeting strategy.
Huawei · 2012 Labs (Xi'an Institute)
Jan 2013 – Apr 2015
Backend Engineer Data collection · Web · C++ backend
Data collection Forwarded on-premise data to an HDFS cluster; 10× throughput.
Web JS-based web work for early Huawei Cloud.
C++ backend Design and development of network-element creation and maintenance software.
Selected Projects
Journeys
Industry First L3 · Episodic Memory Edge Copilot Mode flagship
Tech Lead · Backend Owner (C5 GA) · Microsoft Edge
Automatically clusters browsing history into themes (one group = one journey); each group renders a visual task card plus two suggested actions, so users can pick up where they left off right from the browser home. The industry's first browser-native experience that turns long-term browsing intent into visual task cards. Business impact: +3.9% conversion · +4.3% CTR · +24% coverage.
Tabs Auto Grouping
2022 · Industry First L1 · Activity Memory ~12 months ahead of Google
Sole owner of Client + Backend + Algorithm on this 0→1 project · Microsoft Edge
World's first LLM-based browser tab auto-grouping (Google's comparable Tab Organizer only shipped with Chrome 121 in Jan 2024). Algorithm: in-house Tab semantic-similarity + topic-clustering pipeline, with routing that evolved along the accuracy / latency / cost triangle — ① classic BERT semantic clustering + topic modeling → ② GPT 3 → 5 series → ③ an in-house SLM (fine-tuned on Mistral). Impact: AI + Tabs from 0 → millions of DAU, 3× growth on the Tabs business, and +1% BSOM (Browser Share of Minutes). Featured on the official Edge Tab Groups page.
Web Blueprints
Edge 148 Canary Browser-native
Lead (one of three 2026 core tracks) · Microsoft Edge
Lets the browser "grow" web capabilities for the user: auto-generate mini-apps from the current page or browsing history, or reshape a page on demand via natural language / templated scripts (hide elements, add buttons, extract data…).
Societas / Sociemate
L4 · Team Memory Multi-Agent 0→1 Digital-workforce lead
Tech foundation + digital-workforce lead · Microsoft EdgeML
Societas: built a Manus-class multi-agent stack from scratch, incubating agent teams that cover Office workflows (PPT / Doc). Sociemate: takes multi-agent further so it "lives inside the team's workspace" — a team-first, proactively-aware agent harness combining team-level memory (team-scale LLM-wiki) and proactive participation in discussions and long-horizon tasks (graph engineering for long-horizon tasks), shipped into Teams.
Browser Agent · Vision Action
L2 · Working Memory ReAct Loop
Copilot algorithms / engineering · Microsoft Edge
Intent decides when to act; a ReAct loop plus Vision Action delivers "just say it and it's done." Couples visual grounding with action execution so users complete complex UI tasks by voice alone. Contextual Chat task success 76% → 90%, action success 50% → 90%, action-chain latency 10s → 1s.
Research · Publications · Patents
2027

What Is a Skill Worth? Crediting the Units of an LLM Agent's Skill

First author · AAAI 2027 · Under review

2026

WebRouter: Query-Specific Router via Variational Information Bottleneck for Cost-Sensitive Web Agent

First author · ICASSP 2026 · IEEE Signal Processing Society

Compresses query encodings via a variational information bottleneck (VIB) to cut average per-query inference cost while preserving task quality; deployed for live intent routing in Edge Copilot.

2024

Mixture of Rationale: Multi-Modal Reasoning Mixture for Visual Question Answering

First author · ICONIP 2024

A mixture architecture for multi-modal reasoning that improves both interpretability and accuracy on VQA.

2026

Fara1.5 (Microsoft Research) — Edge × MSR collaboration

Microsoft Research · Computer-Use Agent models

As the Edge-side collaborator, contributed 38,384 multilingual task pairs (DE / ES / FR / PT / IT / JA / ZH — 7 non-English languages) to MSR's Fara1.5 (4B/9B/27B) Computer-Use Agent models. Fara1.5-27B reaches 72% success on Online-Mind2Web, ahead of Gemini 2.5 Computer Use, OpenAI Operator, and Yutori Navigator n1.

Education
Xi'an University of Technology
Pattern Recognition & Intelligent Systems · M.S.
Sep 2009 – Dec 2012
Air Force Engineering University
Electronic & Information Engineering · B.S.
Sep 2005 – Jul 2009