Pin Qian

Research Engineer at Meta

Post-Training · Agent Evaluation · Training Infrastructure

I work on agent post-training and training infrastructure at Meta. My work spans post-training computer-use agents for Ads and building an in-house, Tinker-like Post-Training as a Service (PTaaS) API for teams adapting Meta models to their use cases.

I’m interested in agent evaluation and how evaluation can guide improvements in model behavior through better training data, reward signals, and post-training methods.

Previously, I worked on game-agent preference optimization at Tencent and LLM inference research at Carnegie Mellon University.

Industry Experience

Meta Feb 2025 - Present

Research Engineer

Post-Training Infrastructure

Developed an in-house, Tinker-like Post-Training as a Service (PTaaS) API, providing post-training capabilities to internal teams adapting Meta models to their use cases.

Computer-Use Agent Post-Training · Ads

Worked on post-training computer-use agents for Ads workflows.

Software Engineer Intern, Machine Learning, Creative Delivery, Core Ads Growth May 2024 - Aug 2024

Worked on multimodal LLMs for audience targeting in early-stage delivery systems.

Tencent Jun 2022 - Oct 2022

Machine Learning R&D Intern, Game AI Research Center

Worked on PUBG game AI agent preference optimization with SFT and RL.

Selected Projects

Dynasurge: Dynamic Tree Speculation for Prompt-Specific Decoding

LLM Inference · ML Systems · Course Project

A prototype exploring dynamic token-tree construction and enhanced tree verification for speculative decoding. Includes evaluation scripts comparing autoregressive, static-tree, and dynamic-tree decoding on C4 and CNN/DailyMail.

Research Experience

Infini-AI Lab, Carnegie Mellon University Jul 2024 - Dec 2024

Research Intern, ML Systems

Advised by Prof. Chen Beidi.
Conducted research on accelerating LLM inference using speculative decoding.

University of Liverpool Jan 2021 - Aug 2022

Research Intern, Reinforcement Learning

Conducted reinforcement learning research for portfolio optimization, with a focus on stochastic policies, robust sequential decision-making, and interpretable agent behavior.
Co-authored a paper on interpretable stochastic model-free RL for portfolio optimization, published in Applied Intelligence [paper].

Publications

ForceBench overview: contrastive evaluation of evidence-calibrated claims and stronger overclaims

Relevant Is Not Warranted: Evidence-Force Calibration for Cited RAG

Pin Qian, Su Wang, Xiaoyuan Wang, Yihang Chen, Wenxuan Xu, Qiaolin Yu, Shuhuai Lin, Sipeng Zhang, Junxian You, Xinpeng Wei.

Findings of EMNLP 2026. [paper]

Introduces FORCEBENCH, a benchmark for evaluating whether cited-RAG evaluators correctly calibrate claim strength to supporting evidence.

From deterministic to stochastic: an interpretable stochastic model-free reinforcement learning framework for portfolio optimization

Zitao Song, Yining Wang, Pin Qian, Sifan Song, Frans Coenen, Zhengyong Jiang, Jionglong Su.

Applied Intelligence, 2023. [paper]

Education

Carnegie Mellon University Aug 2023 - Dec 2024

MS, AI Engineering, Electrical and Computer Engineering

University of Liverpool Sep 2019 - Jul 2023

BS, Computer Science