Ant Group’s Selected NeurIPS 2024 Papers: Summaries and Highlights

This document provides a concise roundup of Ant Group’s contributions to the NeurIPS 2024 conference, listing fifteen papers across a broad spectrum of AI research areas. For each paper, the source, open‑access link, paper type (Spotlight, Poster, etc.), involved research fields, and the original abstract are reproduced.

1. MKGL: Mastery of a Three‑Word Language Source: Ant Group Joint Lab Link: PDF Type: Spotlight Fields: Large Language Models, Knowledge Graphs, Graph Completion Abstract: Large language models (LLMs) excel at many NLP tasks, yet their integration with knowledge graphs (KGs) remains under‑explored. This work introduces a dedicated KG language (KGL) composed of entity‑verb‑entity triples, equips LLMs with custom dictionaries and contextual retrieval, and demonstrates substantial error‑rate reductions on KG completion benchmarks.

2. Collaborative Refining for Learning from Inaccurate Labels Source: Ant Group Independent Link: PDF Type: Poster Fields: Machine Learning, Weak Supervision, Noisy Labels Abstract: In industrial settings, cheap automatic labeling yields noisy data. Existing methods ignore data refinement. The proposed CRL framework leverages multi‑annotator consistency to assess label reliability, introducing label‑refinement (LRD) and robust sample selection (RUS) modules that improve performance on both benchmark and real‑world datasets.

3. AMOR: A Recipe for Building Adaptable Modular Knowledge Agents Through Process Feedback Source: Industry‑Academic Collaboration Link: PDF Type: Poster Fields: Question Answering, Knowledge Reasoning, Large‑Model Agents Abstract: Proposes a modular knowledge agent that interacts with external KG via a finite‑state machine, enabling process‑level feedback. Two‑stage fine‑tuning (pre‑heat and adaptation) yields superior knowledge correctness over strong baselines.

4. Exploring Fixed Point in Image Editing: Theoretical Support and Convergence Optimization Source: Ant Group Targeted Collaboration Link: Poster Type: Poster Fields: Image Enhancement & Generation Abstract: Analyzes fixed‑point theory for DDIM inversion, proposes loss‑function refinements to accelerate convergence, and extends the approach to unsupervised image de‑hazing.

5. DeepITE: Designing Variational Graph Autoencoders for Intervention Target Estimation Source: Ant Group Independent Link: PDF Type: Poster Fields: Causal Inference Abstract: Introduces DeepITE, a VGAE‑based framework that jointly learns from labeled and unlabeled interventions, achieving state‑of‑the‑art recall@k while remaining efficient on large graphs.

6. End‑to‑end Learnable Clustering for Intent Learning in Recommendation Source: Ant Group Independent Link: PDF Type: Poster Fields: Recommendation, Intent Learning, Clustering Abstract: Proposes ELCRec, an end‑to‑end learnable clustering architecture that unifies behavior encoding and intent clustering, delivering significant NDCG gains and reduced computation cost in large‑scale industrial recommender systems.

7. Identify Then Recommend: Towards Unsupervised Group Recommendation Source: Ant Group Independent Link: PDF Type: Poster Fields: Recommendation, Intent Learning, Clustering Abstract: Introduces ITR, an unsupervised framework that first discovers user groups without predefined counts and then applies self‑supervised pre‑tasks for group recommendation, achieving notable NDCG improvements.

8. Nimbus: Secure and Efficient Two‑Party Inference for Transformers Source: Ant Group Research Intern Link: PDF Type: Poster Fields: Large Models, Secure Computation Abstract: Presents Nimbus, a two‑party privacy‑preserving inference framework built on Secretflow‑SPU, accelerating linear and non‑linear layers of Transformers by up to 4.7× compared with prior work.

9. Fine‑Grained Dynamic Framework for Bias‑Variance Joint Optimization on Data Missing Not at Random Source: Ant Group Independent Link: PDF Type: Poster Fields: Recommendation, Online Advertising, Causal Inference Abstract: Theoretical analysis reveals limitations of existing regularizers for MNAR data; proposes a dynamic learning framework that adaptively selects estimators per user‑item pair, guaranteeing bounded variance and improved generalization.

10. Rethinking Memory and Communication Costs for Efficient Large Language Model Training Source: Ant Group Independent Link: PDF Type: Poster Fields: Large Language Models, Distributed Training, Performance Optimization Abstract: Introduces PaRO, a set of base strategies (PaRO‑DP, PaRO‑CC) that refine data‑parallel sharding and collective communication, achieving up to 266% speedup over ZeRO‑3 in certain LLM training scenarios.

11. Accelerating Pre‑training of Multimodal LLMs via Chain‑of‑Sight Source: Industry‑Academic Collaboration Link: PDF Type: Poster Fields: Multimodal Large Models, Multi‑Scale Vision, Training Acceleration Abstract: Proposes Chain‑of‑Sight, a visual multi‑scale token resampling and compound token scaling technique that reduces visual token count during pre‑training, cutting training time by ~73% while preserving or improving downstream performance.

12. A Layer‑Wise Natural Gradient Optimizer for Training Deep Neural Networks Source: Ant Group Independent Link: PDF Type: Poster Fields: Deep Neural Network Optimization Abstract: Proposes LNGD, a hierarchical natural gradient descent method that approximates the Fisher matrix via block‑wise Kronecker products and adaptive learning rates, delivering SOTA results on image classification and machine translation.

13. PSL: Rethinking and Improving Softmax Loss from Pairwise Perspective for Recommendation Source: Industry‑Academic Collaboration Link: PDF Type: Poster Fields: AI, Data Mining, Recommendation Abstract: Analyzes limitations of standard Softmax loss, introduces Pairwise Softmax Loss (PSL) with alternative activations that better align with DCG, balance data contribution, and act as a DRO‑enhanced BPR loss.

14. On Provable Privacy Vulnerabilities of Graph Representations Source: Industry‑Academic Collaboration Link: PDF Type: Poster Fields: Graph Neural Networks, Privacy Abstract: Shows that under sparse random graphs, node embeddings can be exploited by the SERA algorithm to reconstruct edge information, and discusses noisy aggregation defenses.

15. PSL (Pairwise Softmax Loss) – detailed above (see entry 13).

16. LLMDFA: Analyzing Dataflow in Code with Large Language Models Source: Industry‑Academic Collaboration Link: PDF Type: Poster Fields: Machine Learning Applications, Code Analysis, Reasoning Abstract: Introduces LLMDFA, a three‑stage LLM‑driven data‑flow analysis pipeline (source‑sink extraction, summary generation, reachability checking) that mitigates hallucination via tree‑sitter parsing and Z3 verification, achieving competitive precision/recall on Juliet and TaintBench.

17. LotCLIP: Improving Language‑Image Pre‑training for Long Text Understanding Source: Ant Group Research Intern Link: PDF Type: Poster Fields: Long‑Text Vision‑Language Abstract: Augments LIP models with corner‑token aggregation to better capture long captions, demonstrating an 11.1% gain on long‑text image retrieval while retaining short‑text performance.

18. Zero‑shot Image Editing with Reference Imitation (MimicBrush) Source: Industry‑Academic Collaboration Link: PDF Type: Poster Fields: Image Editing, Diffusion Models Abstract: Proposes MimicBrush, a self‑supervised diffusion framework that learns semantic correspondence between two video frames to enable reference‑based, zero‑shot image editing without explicit masks.

Overall, these papers illustrate Ant Group’s active research across foundational AI, large‑scale model training, privacy, multimodal learning, and practical system innovations.