How we rank sources
Peer-reviewed journal or conference paper — DOI + volume/page, or a proceedings record. Language-neutral: an English Nature article and a Chinese core journal (e.g. 计算机学报 / 自动化学报 / CSCD-listed) or a CCF-A conference paper are equally Tier 1.
Preprint with a pinned version. arXiv, but also Chinese preprint servers (ChinaXiv / arxiv.cn / CNKI preprint). We always cite the version we actually read.
Official repository, official lab announcement, or official release from the originating body. A lab's own site (.edu.cn, official lab page) announcing its own work is Tier 3 — as authoritative as a GitHub repo. A vendor's own press release is Tier 3 too. Media reposting the same news is NOT Tier 3 (it is Tier 4).
Press, blogs, social posts — leads only. A Chinese tech-media write-up (e.g. 机器之心 / 量子位) is Tier 4 exactly like an English one; only the originating body's own post is Tier 3.
Rules
Language is not a factor. Tier is decided by provenance and evidence strength, never by language. A Chinese lab announcement and an English GitHub repo are judged by the same standard.
Chinese-language sources must carry:
published— the date shown on the sourcearchive— a local snapshot file (under /data/archive/)archive_bytes+sha256— so the snapshot is verifiablequote_zh— the verbatim claim, andtranslation_en— its English rendering
Chinese-language sources must be archived locally (a snapshot file + sha256) and carry a verbatim quote (quote_zh) with an English translation (translation_en), so a reader who cannot read Chinese can still verify the claim. An entry missing any of these fields is rejected from the ledger.
All 29 sources
| Lang | Source | Type | Verified | Link |
|---|---|---|---|---|
| 中文 | 视觉–语言–动作模型综述：从前史到前沿 Vision-Language-Action Models: From the Early Foundations to the State-of-the-Art (Acta Automatica Sinica)
「VLA)模型作为具身智能发展的核心方向, 旨在构建统一的多模态表示与感知–决策–执行一体化架构, 以突破传统模块化系统在功能割裂、语义对齐不足及泛化能力有限等方面的瓶颈.」
Vision-Language-Action (VLA) models are a core direction for embodied intelligence, aiming to build a unified multimodal representation and a perception–decision–execution integrated architecture, to overcome the bottlenecks of traditional modular systems in functional fragmentation, insufficient semantic alignment, and limited generalization.
|
journal-survey | 2026-10-31 | open ↗ |
| 中文 | 资源受限的大模型高效迁移学习算法研究综述 A Survey on Efficient Transfer Learning Algorithms for Large Models under Resource Constraints (Chinese Journal of Computers)
「本文是高效迁移学习领域的首篇中文综述，系统总结了近5年来该领域的研究进展.」
This paper is the first Chinese-language survey in the field of efficient transfer learning, systematically summarizing research progress in the field over the past five years.
|
journal-survey | 2026-10-31 | open ↗ |
| 中文 | 大模型时代的混合专家系统优化综述 A Survey on Mixture-of-Experts System Optimization in the Era of Large Models (Journal of Computer Research and Development)
「混合专家（mixture of experts，MoE）凭借在模型参数扩展、计算成本控制和复杂任务处理等方面的独特优势成为大模型的主流架构之一.」
Mixture of Experts (MoE), with its unique advantages in model parameter scaling, compute-cost control, and complex task handling, has become one of the mainstream architectures of large models.
|
journal-survey | 2026-10-31 | open ↗ |
| 中文 | 混合专家大语言模型的系统与架构优化技术综述 A Survey on System and Architecture Optimization Techniques for Mixture-of-Experts Large Language Models (Journal of Electronics & Information Technology)
「混合专家已经成为当前进一步提升大语言模型推理能力的重要方法。」
Mixture of Experts has become an important method for further enhancing the inference capabilities of large language models today.
|
journal-survey | 2026-10-31 | open ↗ |
| 中文 | 深度学习框架测试研究综述 Survey on Testing of Deep Learning Frameworks (Journal of Software)
「深度学习框架作为连接应用场景和硬件平台的中间部件, 向上支撑深度学习应用的开发, 帮助用户快速构造不同的深度神经网络模型, 向下深度适配各类计算硬件, 满足不同算力架构和环境下的计算需求.」
A deep learning framework is the middleware connecting application scenarios with the hardware platform: upward it supports the development of deep learning applications and helps users quickly build diverse neural network models; downward it adapts deeply to various computing hardware to meet compute needs across architectures and environments.
|
journal-survey | 2026-10-31 | open ↗ |
| EN | DeepSeek-R1, Nature 645, 633-638 | journal | 2026-10-30 | open ↗ |
| EN | arXiv | index | 2026-10-30 | open ↗ |
| EN | BioMARS (paper) | preprint | 2026-10-30 | open ↗ |
| EN | BioMARS (PDF) | preprint | 2026-10-30 | open ↗ |
| EN | DeepSeek-R1 (paper, v2) | preprint | 2026-10-30 | open ↗ |
| EN | DeepSeek-R1 (PDF) | preprint | 2026-10-30 | open ↗ |
| EN | X-VLA (paper) | preprint | 2026-10-30 | open ↗ |
| EN | X-VLA (PDF) | preprint | 2026-10-30 | open ↗ |
| EN | Artificial Analysis — model leaderboard | benchmark | 2026-10-31 | open ↗ |
| 中文 | 通义Qwen3 全系列上架阿里云百炼（官方产品动态） Qwen3 full series on Alibaba Cloud Bailian (official product news)
「Qwen3全系模型实现了推理模式和非推理模式的有效融合，灵活切换两种模式，达到同规模业界顶尖水平。」
The full Qwen3 family fuses 'thinking mode' and 'non-thinking mode' and switches between them flexibly, reaching top-tier levels for its size class.
|
official-announcement | 2026-10-31 | open ↗ |
| EN | DeepSeek API documentation | official-docs | 2026-10-31 | open ↗ |
| EN | DeepSeek-R1 LICENSE (MIT) | license | 2026-10-30 | open ↗ |
| EN | deepseek-ai/DeepSeek-R1 | repository | 2026-10-30 | open ↗ |
| EN | deepseek-ai/DeepSeek-V3 | repository | 2026-10-30 | open ↗ |
| 中文 | DeepSeek-V3 正式发布（官方中文发布说明） DeepSeek-V3 official release note (Chinese)
「DeepSeek-V3 为自研 MoE 模型，671B 参数，激活 37B，在 14.8T token 上进行了预训练。」
DeepSeek-V3 is a self-developed MoE model with 671B parameters, 37B activated, pre-trained on 14.8T tokens.
|
official-release | 2026-10-31 | open ↗ |
| EN | InternLM/InternLM | repository | 2026-10-30 | open ↗ |
| EN | OpenBMB/MiniCPM | repository | 2026-10-30 | open ↗ |
| 中文 | 飞桨 PaddlePaddle 官方中文站 PaddlePaddle official Chinese site
「源于产业实践的开源深度学习平台」
An open-source deep learning platform born from industrial practice.
|
official-site | 2026-10-31 | open ↗ |
| EN | PaddlePaddle issue tracker | official-tracker | 2026-10-31 | open ↗ |
| EN | PaddlePaddle/Paddle | repository | 2026-10-30 | open ↗ |
| EN | Qwen3 official blog | official-blog | 2026-10-31 | open ↗ |
| EN | QwenLM/Qwen3 | repository | 2026-10-30 | open ↗ |
| 中文 | 阿里巴巴推出Qwen3，树立混合推理开源AI领域新标杆 Alibaba launches Qwen3, setting an open-source benchmark for hybrid reasoning
「Qwen3系列包含六个密集模型和两个混合专家（MoE）模型」
The Qwen3 series comprises six dense models and two mixture-of-experts (MoE) models.
|
official-press-release | 2026-10-30 | open ↗ |
| 中文 | 书生·浦语3.0 升级（上海AI实验室官方公告） InternLM3 upgrade — Shanghai AI Lab official announcement
「书生·浦语3.0通过精炼数据框架，大幅提升了数据效率，并实现思维密度的跃升。仅使用4T训练数据，其综合性能超过了同量级开源模型，节约训练成本75%以上；同时，书生·浦语3.0首次在通用模型中实现了常规对话与深度思考能力融合，可应对更多真实使用场景。」
InternLM3 raised data efficiency through a refined-data framework and achieved a leap in 'intelligence quality per token'. Using only 4T training tokens, InternLM3-8B-Instruct outperforms same-size open models and cuts training cost by over 75%; it is also the first general model to fuse ordinary dialogue with deep-thinking capability.
|
official-announcement | 2026-10-31 | open ↗ |
Chinese-language share is 10/29 (≈34%). Five are Tier-3 first-hand sources from originating bodies' own channels (paddle-cn, deepseek-v3-zh, shlab-internlm3-zh, aliyun-qwen3-zh, qwen3-zh-release). The other five are Tier-1 Chinese core-journal surveys — CCF-A / CSCD / 中文核心 — that directly back site items: 软件学报→PaddlePaddle, 自动化学报→X-VLA, 计算机研究与发展+电子与信息学报→DeepSeek-V3/Qwen3 (MoE), 计算机学报→MiniCPM (resource-constrained). Each carries archive + sha256 + verbatim quote_zh + translation_en. Harvesting is now automated (harvest-zh-sources.js) with a monthly gap-report automation. See /sources/. Tips welcome via the report page.
Spot a missing or wrong source? Report it →