I'm co-organizing SimWorld: Human and Environment Simulation for Real Intelligence, a workshop proposed for ICLR 2027 on human and environment simulation for training and evaluating AI agents. We're recruiting Program Committee members — sign up to review.
Lingyun Wang is a researcher on Shopify's Agentic Foundation Modeling team, where he works on agentic AI systems that connect foundation model research, scalable infrastructure, and real-world production impact.
Before joining Shopify, he worked in Amazon Stores Foundational AI, Amazon Ads, Alexa AI, Hulu, and Microsoft, contributing to foundational AI research and large-scale machine-learning-powered products across shopping, advertising, entertainment, recommendation, and conversational AI.
Interests
News
New preprint: AMBER trains long-horizon web agents with append-only memory and outcome-based RL. We also built reusable asynchronous RL infrastructure with SGLang and Miles, designed for agent training beyond AMBER. It separates inference from environment execution, overlaps model generation with browser interactions, and isolates each rollout’s browser and website state for scalable training. Code and infrastructure are open source.
SimTrace has been accepted to two NeurIPS 2026 workshops: Evaluation of Interactive Agents and Who Verifies the Agents?. It generates grounded synthetic user trajectories for user modeling, matching real-data performance on purchase prediction and recommendation. Code is open source.
ShopGym: An Integrated Framework for Realistic Simulation and Scalable Benchmarking of E-Commerce Web Agents has been accepted to NeurIPS 2026 (Evaluations and Datasets Track). ShopGym turns live storefronts into resettable sandbox shops and auto-generates grounded benchmark tasks, closing the gap between realistic and controllable evaluation of e-commerce web agents. Explore the project page or the open-source code.
EARS: Explanatory Abstention for Reliable Sub-Agent Modeling in Large-scale Multi-Agent Systems has been accepted to the EMNLP 2026 Workshop for Research on Agent Language Models (REALM). EARS treats sub-agent abstention as an inter-agent communication protocol — instead of over-answering ambiguous or unsupported requests, a sub-agent exposes an actionable failure state to the coordinator for clarification, rerouting, or fallback. On a large-scale e-commerce assistant, EARS raises the response pass rate from 68.5% to 78.9%.
SimGym: A Framework for A/B Test Simulation in E-Commerce with Traffic-Grounded VLM Agents has been accepted to the COLM 2026 Workshop on Agent Behavior. SimGym grounds browser agents in real traffic to forecast how storefront changes affect buyer behavior, shortening experimentation from weeks to under an hour without exposing real shoppers to untested variants.
SimPersona: Learning Discrete Buyer Personas from Raw Clickstreams for Grounded E-Commerce Agents has been accepted to the COLM 2026 Workshop on Learning from Situated and Embodied Interaction. SimPersona learns compact buyer types directly from clickstreams, giving web agents grounded and diverse behavior while preserving each merchant's real buyer distribution.
New engineering blog: Teaching Sidekick to say no: automated data curation with LLM judge consensus. We share how consensus across LLM judges turns noisy production interactions into reliable training data, helping a domain-specific Sidekick model recognize impossible requests and refuse them well instead of producing an ungrounded answer.
Our paper, ShopGym: An Integrated Framework for Realistic Simulation and Scalable Benchmarking of E-Commerce Web Agents, has been accepted to the ICML 2026 RLxF Workshop, with open-source code released.
Research
Selected publications after 2021, grouped by research area.
Agentic modeling
AMBER: Training Long-Horizon Web Agents through Append-Only Memory
Append-only memory bank that guarantees evidence retention by construction, trained end-to-end with RL from outcome rewards for long-horizon web agents.
Code: GitHub
EARS: Explanatory Abstention for Reliable Sub-Agent Modeling in Large-scale Multi-Agent Systems
Abstention as an inter-agent communication protocol — sub-agents expose actionable failure states for clarification, rerouting, or fallback.
User and environment simulation
SimTrace: Grounded Multimodal User Trajectories Generation for Online User Modeling
Grounded computer-use agents generate faithful, fine-grained synthetic multimodal clickstreams as a shareable alternative to confidential user logs.
Code: GitHub
ShopGym: An Integrated Framework for Realistic Simulation and Scalable Benchmarking of E-Commerce Web Agents
Realistic, reproducible storefront simulation and benchmark generation for web agents.
Project page: shopgym-research.github.io · Code: GitHub
SimGym: A Framework for A/B Test Simulation in E-Commerce with Traffic-Grounded VLM Agents
Traffic-grounded browser agents for simulating e-commerce A/B tests.
SimPersona: Learning Discrete Buyer Personas from Raw Clickstreams for Grounded E-Commerce Agents
Buyer persona modeling from clickstreams for grounded synthetic shopping agents.
LLM modeling
LongLeader: A Comprehensive Leaderboard for Large Language Models in Long-context Scenarios
Benchmarking long-context comprehension across large language models.
An Exploration of Speech Conditioned Large Language Models (SLMs)
Design-space exploration for speech-conditioned LLMs and spoken instruction following.
Multimodal modeling
M-LLM Based Video Frame Selection for Efficient Video Understanding
Adaptive frame selection for efficient video reasoning with multimodal large language models.
Multimodal Instruction Tuning with Hybrid State Space Models
Hybrid transformer-Mamba modeling for long-context multimodal instruction tuning.
Invited Talks
Training Optimization for LLM with NVIDIA NeMo and AWS
A technical session on large language model training optimization with NVIDIA NeMo and AWS.
Watch the sessionBlog
Engineering perspectives on agentic modeling and training, inference optimization, and user and environment simulation.
ShopGym: Realistic, reproducible sandboxes for shopping agents
Turning live storefronts into resettable sandbox shops and generating grounded shopping tasks for reproducible agent evaluation.
Teaching Sidekick to say no: automated data curation with LLM judge consensus
Using consensus across LLM judges to curate training data at scale and teach domain-specific models when—and how—to refuse impossible requests.
Flow generation through natural language: an agentic modeling approach
Bringing agentic modeling from idea to production, with the lessons learned along the way.
2,000 robots walk into a shop...
An engineering deep dive into how SimGym was built, scaled, and applied to simulate e-commerce agent behavior, with a focus on inference optimization.
How Amazon scaled Rufus by building multi-node inference using AWS Trainium chips and vLLM
Scaling Rufus with multi-node LLM inference on AWS Trainium and vLLM, combining hybrid model parallelism with topology-aware deployment on Amazon ECS.
Products
Shipped products with media coverage.
Agentic Foundation Modeling
Sidekick — Shopify's AI-powered commerce assistant, helping merchants run and grow their businesses through natural-language guidance and agentic actions.
Sidekick Pulse — proactive, personalized recommendations that turn store data into actionable business insights, with context passed to Sidekick to help merchants take action.
SimGym — a framework for user behavior simulation in e-commerce with traffic-grounded VLM agents, enabling realistic experimentation before launch.
Coverage: Sidekick 2025 · Sidekick 2026 · SimGym · Pulse video demo · Pulse news
Foundation Modeling and Inference Optimization
Rufus — Amazon's generative AI–powered conversational shopping assistant, helping customers research products, compare options, and decide what to buy in natural language.
Coverage: Rufus Science · Rufus Product · Rufus Multi Modal
Recommendation
Alexa Brain — recommendation and personalization behind Amazon Alexa, powering skill discovery and proactive suggestions across hundreds of millions of devices.
Coverage: Amazon Alexa Developer Blog
Patents
Computer-Implemented Methods for Machine Learning Model Based Spatial-Temporal Adaptive Shift for End-to-End Text-Video Retrieval
Spatial-temporal shifting for text-video retrieval and video search matching.
Enhanced Shopping Based on Recognition of Objects Presented in Video
Object recognition in video for smart shopping and product discovery experiences.
Allocating a Bounded Query Budget Across Temporal and Interaction-Strength Evidence Tiers for a Generated-Query Retrieval
US 19/843,845 · Filed October 19, 2026.
Methods and Systems for Explanatory Abstention and Recovery in Multi-Agent Language Models
US 19/805,066 · Filed September 11, 2026.
Systems and Methods for Personalizing Large Language Model Prompts Using Discrete Category Tokens Assigned to Entities Based on Data Representations of Those Entities
US 19/772,631 · Filed August 11, 2026.
Methods and Systems for AI-Assisted Data Organization and Item Discovery in Two Dimensions
US 19/739,639 · Filed July 13, 2026.
Methods and Systems for Multi-Axis Organization of Data Using Spatial Semantic Context
US 19/739,712 · Filed July 13, 2026.
Computer System, Computer-Implemented Method, and Computer Readable Media for Generating Context Profiles for Large Language Models (LLMs)
US 19/739,553 · Filed July 13, 2026.
Computer System, Computer-Implemented Method, and Computer Readable Media for Utilizing Context Profiles for Large Language Models (LLMs)
US 19/739,458 · Filed July 13, 2026.
Tuning of LLM-Based Classifier Prompts
US 19/450,182 · Filed January 15, 2026.
Methods and Systems for Fine-Tuning Self-Healing Language Models with a Punctured Dataset
US 19/415,968 · Filed December 11, 2025.
Two additional patent applications related to generative recommendation are being prepared for filing with the U.S. Patent and Trademark Office.
Education
Virginia Tech
M.S., Computer Engineering, earned in the Ph.D. program
Advisor: Prof. Robert P. Broadwater