Two Ant Group Papers Selected for AAAI 2025: Human‑Feedback Evaluation Framework for Product Image Background Inpainting and Bagging‑Expert Network for Multi‑Task Learning

With the continuous advancement of artificial intelligence technology, large language models and visual reasoning have become key forces driving the sustained development of intelligent technologies. At the upcoming AAAI 2025 conference, several Ant Group papers have been accepted, revealing the latest research achievements in the field.

In this issue, we introduce two of these papers, which showcase the cross‑disciplinary exploration of intelligent technologies.

The paper "An Evaluation Framework for Product Images Background Inpainting based on Human Feedback and Product Consistency" proposes a human‑feedback‑driven evaluation framework to measure the consistency between the main subject and background in product images. It also introduces an EfficientSAM‑based pipeline to improve the efficiency of subject‑consistency judgment, and constructs the HFPC‑44k dataset containing images with mismatched backgrounds and subject variations, balanced across product categories, providing valuable data resources for related research.

The second paper, "Bagging‑Expert Network for Multi‑Task Learning: A Depolarization Solution in Multi‑Gate Mixture‑of‑Experts," addresses the expert polarization problem in Multi‑Gate Mixture‑of‑Experts (MMoE) for multi‑task learning. It proposes a Bagging‑Expert Network solution, offering a new perspective for multi‑task learning that could play an important role in recommendation algorithms, large models, and other domains, further enhancing multi‑task learning performance.

These two papers demonstrate the cross‑domain integration of intelligent technologies, ranging from expert networks for multi‑task learning to a human‑feedback‑based evaluation framework for image background restoration, reflecting the diversity and practicality of AI. This integration not only showcases the broad application potential of intelligent technologies but also provides new ideas for solving real‑world problems.

We have invited the first authors of the two papers to share their work live: Jiankui Bi, Ant Group Algorithm Engineer, and Gongduo Zhang, Ant Group Application Algorithm Expert. They will present their research from 18:00 to 20:00 on January 16, 2025, via the "Paper Show Live#14" and "SecretFlow Live#28" streams, and welcome audience interaction to learn about their research ideas and experiments.

Live Viewing Guide

Time: 2025‑01‑16 18:00‑20:00 Platforms: WeChat Channels (Ant Technology Research Institute, AntTech), SecretFlow, Bilibili (Ant Technology Research Institute, SecretFlow) – simultaneous live broadcast. Please follow and set a reminder.