Qianru Lao

Qianru Lao is a systems engineer and Member of Technical Staff on OpenAI’s Inference team, working on infrastructure for large-scale model serving. Her engineering experience spans distributed storage, application infrastructure, and contributions to the open-source Delta Lake project. Her research includes computer vision and memory-aware inference scheduling, with particular attention to requests that pause while calling external services.

From storage and application infrastructure to model serving

Lao studied computer science at Sun Yat-sen University before pursuing graduate study in Computational Science and Engineering at Harvard from 2023 to 2025. Her early engineering work included developing features for an internal data portal during a Google internship in 2021 and working on Google Pay infrastructure in 2022. In 2023, she joined Alibaba Cloud’s Yunxiao team as a research-and-development intern, optimizing a distributed storage system.

Alongside these infrastructure roles, Lao pursued computer-vision research. From late 2022 into 2023, she worked on person-centric visual grounding with Huaizu Jiang’s Visual Intelligence Lab at Northeastern University. Her co-authored research examines how models can identify objects from natural-language descriptions and learn from images whose labels are incomplete. At Harvard, she also served as a teaching assistant for operating systems and advanced computer networks.

Before OpenAI, Lao contributed to Delta Lake at Databricks. Her research also extended into language-model serving: she co-authored MARS, published at NeurIPS 2025, which addresses the scheduling and memory costs of model requests that call external APIs. This work connects her systems background to a practical problem in inference: deciding which requests should run when waiting on another service can keep GPU memory occupied.

Research contributions

• Scheduling around tool calls: In MARS, Lao and her collaborators address requests that stop generating while an external service responds but continue occupying scarce GPU memory. Even a request producing a short answer can take a long time to finish. The framework combines predictions about API calls with memory-aware request prioritization, accounting for costs that output length alone cannot capture. The team’s open implementation builds on vLLM.

• Visual grounding through relationships: Lao co-authored research on structural similarity between images and captions, published at CVPR 2024. Visual grounding means locating the object described in natural language. Recognizing the nouns may be insufficient when relationships distinguish otherwise similar objects. The team decomposes images and descriptions into subject–predicate–object triplets and matches their structure, while fine-tuning a vision-language model to improve its understanding of relationships. This approach to zero-shot visual grounding uses connections among entities and actions to identify the intended object.

• Learning from incomplete labels: Her co-authored research on partially labeled images addresses the risk of treating an unrecorded label as evidence that an object is absent. The method uses semantic similarities across images to recover missing positive labels and reject unreliable labels, with thresholds that adapt by category. It supports multi-label image recognition when exhaustive annotation is too costly.