Xiaohongshu REDtech Live: Presentation of Recent Top‑Conference Papers (Recruitment Session)

On August 24, 2024 (Saturday) from 14:00 to 18:00, Xiaohongshu’s technical team will host a live streaming event titled “REDtech Came” to share their latest research results published at top conferences such as ACL, CVPR, SIGIR, IJCV, ICLR, AAAI, and CIKM. The live stream will be broadcast simultaneously on WeChat Channels (Xiaohongshu Tech REDtech, Xiaohongshu Recruitment), Xiaohongshu account (Recruitment Shu), and Bilibili (Xiaohongshu Tech REDtech). Attendees can interact via live comments and receive surprise Xiaohongshu merchandise.

Live Stream Details

Time: 2024‑08‑24 14:00‑18:00 Platforms: WeChat Channels (Xiaohongshu Tech REDtech, Xiaohongshu Recruitment), Xiaohongshu account (Recruitment Shu), Bilibili (Xiaohongshu Tech REDtech)

Paper Presentations

Yang Dongjie – ACL 2024 Proposes PyramidInfer , a method that compresses KV cache during the pre‑fill stage of large‑language‑model inference by extracting key‑value pairs layer‑wise based on attention‑weight consistency. Experiments show a 2.2× speedup over Accelerate and a 54% reduction in GPU memory usage. Paper: https://arxiv.org/abs/2405.12532

Yu Jinpeng – CVPR 2024 Introduces the SSR‑Encoder module that extracts salient subject features from one or multiple reference images to guide zero‑shot image generation. The method works with ControlNet, LoRA and other consistency‑control techniques, improving subject selection and consistency. Paper: https://arxiv.org/abs/2312.16272

Li Yiwei – ICLR 2024 Presents Early‑Stopping Self‑Consistency (ESC) , a sampling strategy that stops self‑consistency decoding early without sacrificing performance, dramatically lowering the number of samples required. An ESC control scheme dynamically balances performance‑cost trade‑offs across tasks (math, commonsense, symbolic reasoning). Paper: https://openreview.net/pdf?id=ndR8Ytrzhh

Li Yiwei – AAAI 2024 (oral) Proposes a negative‑sample‑aware model distillation framework that leverages both positive and negative synthetic data to specialize large language models. The framework consists of Negative‑Assisted Training (NAT), Negative‑Calibrated Enhancement (NCE), and Adaptive Self‑Consistency (ASC). Paper: https://arxiv.org/pdf/2312.12832

Yuan Peiwen – ACL 2024 (oral) Introduces BatchEval , a low‑cost, human‑inspired text‑evaluation method that improves robustness by leveraging diverse evaluation perspectives. Compared with state‑of‑the‑art baselines, BatchEval achieves superior evaluation quality with lower computational overhead. Paper: https://arxiv.org/abs/2401.00437

Yuan Peiwen – ACL 2024 (findings) Proposes PEEM , a method that uses inter‑model consistency as an evaluation signal for super‑human LLMs. By modeling consistency between a reference model and the target model, PEEM provides an accurate capability metric even when human judgments become infeasible. Code: https://github.com/ypw0102/PEEM

Wang Xinglin – ACL 2024 (main) Presents Fine‑Grained Self‑Consistency (FSC) , which refines self‑consistency decoding by selecting fine‑grained commonalities among generated samples. FSC improves code generation, summarization, and mathematical reasoning while keeping computational cost comparable. Code: https://github.com/WangXinglin/FSC

Sun Yiping – CIKM 2024 Introduces RTAMS‑GANNS , a real‑time multi‑stream GPU‑accelerated approximate nearest‑neighbor search system designed for online insertion of new vectors, crucial for retrieval‑augmented generation in LLM applications. The system uses dynamic memory‑block insertion and stream‑level resource pooling to achieve high QPS without sacrificing latency. Paper: https://arxiv.org/abs/2408.02937

Li Shanglin – CVPR 2024 Proposes a zero‑shot instruction‑guided local editing technique that leverages cross‑attention “edit‑aware” signals in diffusion models to achieve precise, single‑instruction image editing, outperforming existing methods in both human evaluation and quantitative metrics. Paper: https://openaccess.thecvf.com/content/CVPR2024/papers/Li_ZONE_Zero-Shot_Instruction-Guided_Local_Editing_CVPR_2024_paper.pdf

Wang Haocen – ICCV 2023 (oral) & IJCV Introduces Open Vocabulary Video Instance Segmentation (OV‑VIS) , enabling segmentation, tracking, and classification of arbitrary object categories in videos, expanding the capabilities of video instance segmentation for short‑video editing and labeling. Paper: https://link.springer.com/article/10.1007/s11263-024-02076-w

The event aims to explore new opportunities and challenges in AI research, allowing participants to directly interact with the authors and gain insights into cutting‑edge techniques.