Haifeng Gong

Research

Current: AI Agents for Ranking & Content

At Meta Threads I work on AI agents in the recommendation funnel: agent based retrieval parameter personalization, agent based list ranking for nearline applications, agent based content understanding and feature engineering, and agents as teachers for ranking model distillation. I also build mixed vendor agent swarms for data analysis, model auditing and accident investigation.

At Google I built image agents for Display Ads — multimodal agent workflows chaining Text2Text, Text2Image and Image2Text models with traditional computer vision tools for image ad creation (2025 Google Tech Impact Award).

Image agent orchestration pipeline: ad brief to Text2Text LLM, Text2Image model, Image2Text LLM and CV tools, with critique-and-refine loop, RAG and training data treatment
Image agent orchestration for ad creation: a multimodal pipeline with a critique-and-refine loop, supported by few-shot/CoT RAG & embedding and training data treatment.
Ad assets table with business name, slogan, logo, marketing text, calls to action and images, composed into three different ad designs
Asset Based Creative: advertiser assets automatically composed into diverse ad designs with irregular cropping, decoration and asset combinations.

Generative Recommender

User interaction history modeling, HSTU based dual encoder and cross encoder, generative retrieval; retrieval / early stage / LTV ranking models with two-tower, multi-task multi-tower learning and knowledge distillation.

Earlier: Video Analysis

Video Manifold

Spatio-temporal Patches based Background Modeling

Tracking by Planning

Trackability

Technology Review Report

Pedestrian tracking frames and motion planning diagrams with plans, goals, obstacles, tracklets and occlusion handling
Tracking by planning: multi-hypothesis motion planning for pedestrian tracking by autonomous vehicles (ICCV 2011).
Night-time background modeling comparison: input frames, GMM, LBP and our method
Night-time video surveillance that beat the best commercial system of its day (MIT Technology Review; ICPR 2008).
Blob scale constraints imposed by ground plane and vertical surfaces with camera calibration
Scene-geometry constraints for indoor background modeling (CVPR 2008).

Earlier: Object Recognition

Active Basis

Mixed Template

Active Basis with EM