Sitemap
A list of all the posts and pages found on the site. For you robots out there is an XML version available for digesting as well.
Pages
Posts
CICAI 2026 参会笔记：网安 & Agent 视角观察记录
Published:
从 CICAI 应用报告出发，记录可靠性、真实反馈、长尾与成本观察，并联想网安 Agent 和 FrontPost 的跨领域启发。
FrontPost 满两月：渐成其章，渐见远方
Published:
从 7 月中旬到现在，FrontPost 跑了两个月。两个月不长，但足够把一个概念推进到产品验证的门槛上。FrontPost 的假设是有一批人需要一个能精准解决信息筛选与判断问题的自动化产品——围绕长期关注的主题自动获取信息，过滤噪音，把值得看的变化组织好主动推过来。两个月下来，这个假设已经有了完整的产品载体：Web、账户体系、Focus、内容生产、邮件投递、后台任务、支付能力全部落地，ICP 备案已完成，支付中台已跑通。因此目前需要回答的问题已经从”系统能不能跑”变成了”这套产品关系能否长期成立”。
反本本主义：反经典崇拜、反榜单崇拜、反方法崇拜、反权威口味、反引用崇拜
Published:
上一篇写 FrontPost 的方法论 时，我想讨论的是如何看待期刊名次、引用数、机构声誉、社交热度和 GitHub star。这些指标都有一点信息量，但作用有限：它们可以帮助我们决定先看哪里，无法单独说明一项研究是否重要、严谨、可复现，或者是否真的带来了新东西。
FrontPost 的方法论
Published:
前沿信息的难点在于：及时发现变化，理解它的证据、背景和可能影响，并将这些判断送到需要它的人面前。FrontPost 希望建立一套长期工作的研究型信息系统：从开放世界中发现变化，判断哪些主张有依据，理解一个人正在解决什么问题，再把值得他现在知道的内容编成可以读完、可以追溯的一期邮报。
我想做一个 AI 邮报（FrontPost）
Published:
最近我又有了一个小想法，而且很想把它做出来。
小狗要好好休息好好睡觉！
Published:
隆重介绍一下，博客的新成员来了：小狗啾啾！🐶
为不确定性建立秩序：LLM API 网关的设计与实现
Published:
LLM API 网关不是“把几个模型接口包一层”这么简单。对 AI-native 产品来说，它更像一层执行控制面：在 MaaS（Model-as-a-Service）、对象存储、任务状态机、内容安全、成本统计和合规审计之间建立秩序。本文从一个 AI 漫画平台的自建 LLMAPI 模块实践出发，讨论为什么通用聚合层还不够、媒体生成任务会怎样放大网关价值，以及以后面向国内小微团队的低门槛大模型网关可能长成什么样🤔。
向前进，莫彷徨，黑暗尽处有曙光
Published:
第一篇博客文章，被考研折磨一年后的碎碎念🥱
portfolio
2024相册
大二下大三上
2025相册
大三下大四上
publications
Evaluating and Enhancing LLMs for Multi-turn Text-to-SQL with Multiple Question Types
Published in International Joint Conference on Neural Networks (IJCNN), 2025
Developed MMSQL, a test suite that evaluates how well LLMs manage different question types and multi-turn interactions. Additionally, created a multi-agent system to better identify question types and select appropriate strategies. Experiments show that this approach enhances the models' ability to navigate conversational complexities. For a more detailed presentation, refer to the Page. IEEE Arxiv
AdaCOS: adaptive differential privacy shuffle model based on cosine similarity
Published in Journal of King Saud University Computer and Information Sciences, 2026
AdaCOS makes federated learning training dynamic and smart. Instead of treating every device the same, it constantly evaluates their updates. For devices contributing valuable, on-target information, it intelligently increases their communication allowance while reducing protective noise.
QDA-SQL: Questions Enhanced Dialogue Augmentation for Multi-Turn Text-to-SQL
Published in CAAI International Conference on Artificial Intelligence (CICAI), 2026
Fine-tuning LLMs for Text-to-SQL tasks is effective, but they often struggle with multi-turn queries due to ambiguity. QDA-SQL, a data augmentation method that generates diverse multi-turn Q&A pairs using LLMs. Fine-tuning with QDA-SQL improves SQL statement accuracy and enhances the models' ability to manage challenging, unanswerable questions. The generation script and test set are released at Github. CICAI 2026 Poster (PDF).
talks
Conference Proceeding talk on MMSQL testsuit
Published:
presentation of paper “Evaluating and Enhancing LLMs for Multi-turn Text-to-SQL with Multiple Question Types” Bilibili
teaching
Teaching experience 2
Workshop, University 1, Department, 2015
This is a description of a teaching experience. You can use markdown like any other post.