Official account of the Xiaohongshu tech team, sharing tech innovations and problem insights, advancing together.
Xiaohongshu’s middleware team migrated thousands of Java services from JDK 8 to RedJDK 11/17, achieving over 10% performance gains, 50% GC pause reduction, and eliminating OOM crashes through systematic JDK upgrades, GC tuning, native‑memory improvements, and standardized deployment pipelines.
The article introduces the Think When You Need (TWYN) method, a reinforcement‑learning approach that dynamically adapts chain‑of‑thought length, dramatically cuts redundant token generation in large language models, and maintains or improves accuracy across diverse reasoning benchmarks.
dots.llm1, an open‑source 142‑billion‑parameter Mixture‑of‑Experts language model from hi lab, achieves Qwen2.5‑72B‑level performance after training on 11.2 T high‑quality tokens, and the release includes full models, intermediate checkpoints, and detailed training pipelines for the research community.
MarkerGen introduces a novel, plug‑and‑play framework that decomposes length‑controllable text generation into four sub‑abilities—identifying, counting, planning, and aligning—integrates external tokenizers and dynamic markers, and achieves significantly lower length errors and higher quality across diverse models, tasks, and languages.
The TailoredBench framework dramatically reduces large‑language‑model evaluation cost and error by using a global probe set, model‑specific source selection, extensible K‑Medoids clustering, and calibration, achieving up to 300× speedup and a 31.4% MAE reduction across diverse benchmarks.
The paper introduces SOGAIC, a scalable overload‑aware graph‑based index construction system for billion‑scale vector similarity search that uses adaptive overlapping partitioning and load‑balanced distributed scheduling to cut construction time by 47.3% while maintaining high recall.
Facing billions of daily logs and the need for minute‑level experiment metrics, Xiaohongshu partnered with Yunqi Tech to design a generic incremental‑compute solution that delivers near‑real‑time data warehousing with lower cost, higher accuracy, simplified pipelines, and improved query performance.
UniCBE introduces a unified multi‑objective optimization framework for contrastive‑based evaluation that mitigates sampling bias, unbalanced uncertainty reduction, and inefficient resource allocation by combining three decoupled probability matrices through a greedy and Hadamard‑product strategy, achieving Pearson correlations above 0.995 with only 83 % of the annotation budget and cutting evaluation costs by more than 50 % across diverse LLM evaluators.
ROFF is a Rust‑implemented, seven‑layer gateway that delivers high‑throughput load balancing with memory‑safe performance, TLS hardware offload, native QUIC/HTTP3 support, a hot‑reload/upgrade mechanism, and an extensible module system allowing over thirty built‑in filters and custom Rust macros.
SAFE is a lightweight AI‑image detection framework using only 1.44 M parameters and 2.30 B FLOPs that preserves fine‑grained artifacts through crop‑based preprocessing, invariant augmentations, and high‑frequency wavelet features, achieving an average 96.7 % accuracy across 33 test subsets and strong generalization to unseen GAN and diffusion generators.
