Xiaohongshu 2024 Large Model Frontier Paper Sharing Live Event

The Xiaohongshu technical team announces a live streaming session titled "Xiaohongshu 2024 Large Model Frontier Paper Sharing" scheduled for June 27, 2024, 19:00‑21:30 (UTC+8) on WeChat Channels, Bilibili, Douyin, and Xiaohongshu.

The session will feature six papers accepted at top international conferences (ICLR, ACL, CVPR, AAAI, SIGIR, WWW) that explore new opportunities and challenges at the intersection of large models and natural language processing, as well as effective evaluation methods and real‑world applications.

Early‑Stopping Self‑Consistency (ESC) – Speaker: Li Yiwei

Self‑Consistency (SC) is a widely used decoding strategy for chain‑of‑thought reasoning but incurs high computational cost due to multiple samples. The proposed Early‑Stopping Self‑Consistency (ESC) introduces a simple, scalable sampling process that stops early when sufficient consensus is reached, dramatically reducing the number of samples without sacrificing performance. Experiments on mathematical, commonsense, and symbolic reasoning tasks show a large reduction in average sampling steps while preserving accuracy. Paper: https://arxiv.org/abs/2401.10480

Fine‑Grained Self‑Consistency (FSC) – Speaker: Wang Xinglin

FSC addresses the limitation of coarse‑grained sample selection in existing self‑consistency methods for free‑form generation tasks. By leveraging fine‑grained commonality among sample fragments through a large‑model self‑fusion mechanism, FSC achieves significant improvements on code generation, summarization, and mathematical reasoning benchmarks while keeping computational overhead comparable. GitHub: https://github.com/WangXinglin/FSC

BatchEval – Speaker: Yuan Peiwen

BatchEval is a novel evaluation framework that attains near‑human performance on text assessment with lower cost. It analyzes the shortcomings of current metrics—uneven score distributions and limited perspective diversity—and introduces a batch‑wise comparison inspired by human evaluation practices. Empirical results show BatchEval outperforms state‑of‑the‑art methods in both efficiency and effectiveness. Paper: https://arxiv.org/abs/2401.00437

PEEM (Probabilistic EM‑based Evaluation Method) – Speaker: Yuan Peiwen

PEEM leverages inter‑model consistency to evaluate large language models that surpass human performance, where reliable human labels become scarce. The method derives a theoretical guarantee that, under independence between reference and target model predictions, consistency with the reference model serves as an accurate ability metric. An EM‑based algorithm mitigates practical violations of the independence assumption, enabling accurate assessment of super‑human LLMs. GitHub: https://github.com/ypw0102/PEEM

Negative‑Sample‑Assisted Distillation – Speaker: Li Yiwei

This work proposes a pioneering framework that incorporates negative (incorrect) samples into the knowledge distillation process for large language models. The three‑step pipeline—Negative‑Assisted Training (NAT), Negative‑Calibrated Enhancement (NCE), and Adaptive Self‑Consistency (ASC)—demonstrates that negative data can substantially improve the distilled model’s reasoning quality, especially on complex mathematical problems. Paper: https://arxiv.org/abs/2312.12832

LLM‑Based Note Content Representation for Recommendation – Speaker: Zhang Chao

To address the cold‑start problem of newly created notes on the Xiaohongshu app, this research builds a recommendation system that encodes note content using large language models. Two approaches are explored: generation‑enhanced representations and multimodal content embeddings. The system has been deployed in multiple product scenarios, delivering notable performance gains. Paper: https://arxiv.org/abs/2403.01744

The live stream will also include a Q&A session, a QR‑code for a community group, and recruitment information for the Xiaohongshu search team.