<feed xmlns="http://www.w3.org/2005/Atom"> <id>https://ivanbao9783.github.io/</id><title>ivanbao</title><subtitle>ivanbao 的个人博客，记录技术笔记、项目记录与个人思考。</subtitle> <updated>2026-08-31T09:56:47+08:00</updated> <author> <name>ivanbao</name> <uri>https://ivanbao9783.github.io/</uri> </author><link rel="self" type="application/atom+xml" href="https://ivanbao9783.github.io/feed.xml"/><link rel="alternate" type="text/html" hreflang="zh-CN" href="https://ivanbao9783.github.io/"/> <generator uri="https://jekyllrb.com/" version="4.4.1">Jekyll</generator> <rights> © 2026 ivanbao </rights> <icon>/assets/img/favicons/favicon.ico</icon> <logo>/assets/img/favicons/favicon-96x96.png</logo> <entry><title>如何识别 LLM 推理文本异常：它选择不看文本，只看概率</title><link href="https://ivanbao9783.github.io/posts/llm-response-anomaly-detection/" rel="alternate" type="text/html" title="如何识别 LLM 推理文本异常：它选择不看文本，只看概率" /><published>2026-08-31T09:00:00+08:00</published> <updated>2026-08-31T09:49:14+08:00</updated> <id>https://ivanbao9783.github.io/posts/llm-response-anomaly-detection/</id> <content type="text/html" src="https://ivanbao9783.github.io/posts/llm-response-anomaly-detection/" /> <author> <name>ivanbao9783</name> </author> <category term="技术笔记" /> <summary>拆解 msprobe response_anomaly：不看输出文本，只靠 token 概率分布检测乱码、复读机与异常回复。</summary> </entry> <entry><title>只测 3% 的题，就能还原模型评测结论？——评测集压缩的几种玩法</title><link href="https://ivanbao9783.github.io/posts/eval-data-mini-data-compression/" rel="alternate" type="text/html" title="只测 3% 的题，就能还原模型评测结论？——评测集压缩的几种玩法" /><published>2026-08-19T10:30:00+08:00</published> <updated>2026-08-20T11:42:32+08:00</updated> <id>https://ivanbao9783.github.io/posts/eval-data-mini-data-compression/</id> <content type="text/html" src="https://ivanbao9783.github.io/posts/eval-data-mini-data-compression/" /> <author> <name>ivanbao9783</name> </author> <category term="技术笔记" /> <summary>只测 3% 的题，就能还原模型评测结论？五种评测集压缩算法横向对比：从 K-Means 聚类到 EssenceBench 排名一致性，附团队实践与选型建议。</summary> </entry> <entry><title>AI的肖申克救赎：OpenAI模型自主越狱攻破HuggingFace，GLM临危救场</title><link href="https://ivanbao9783.github.io/posts/eval-hotspot-openai-hf-jailbreak/" rel="alternate" type="text/html" title="AI的肖申克救赎：OpenAI模型自主越狱攻破HuggingFace，GLM临危救场" /><published>2026-07-24T15:38:39+08:00</published> <updated>2026-08-19T14:58:14+08:00</updated> <id>https://ivanbao9783.github.io/posts/eval-hotspot-openai-hf-jailbreak/</id> <content type="text/html" src="https://ivanbao9783.github.io/posts/eval-hotspot-openai-hf-jailbreak/" /> <author> <name>ivanbao9783</name> </author> <category term="技术笔记" /> <summary>OpenAI 内测模型为在安全评测中抄答案，自主突破沙箱、挖出零日漏洞、入侵 HuggingFace 生产环境——而最终拦住它的，是一款中国开源模型。</summary> </entry> <entry><title>从 LLM-as-a-Judge 到 Agent-as-a-Judge：AI 自动化评测的范式演进与破局</title><link href="https://ivanbao9783.github.io/posts/eval-system-llm-to-agent-judge/" rel="alternate" type="text/html" title="从 LLM-as-a-Judge 到 Agent-as-a-Judge：AI 自动化评测的范式演进与破局" /><published>2026-07-24T09:03:04+08:00</published> <updated>2026-08-19T11:36:12+08:00</updated> <id>https://ivanbao9783.github.io/posts/eval-system-llm-to-agent-judge/</id> <content type="text/html" src="https://ivanbao9783.github.io/posts/eval-system-llm-to-agent-judge/" /> <author> <name>ivanbao9783</name> </author> <category term="技术笔记" /> <summary>从 LLM-as-a-Judge 到 Agent-as-a-Judge：AI 自动化评测的范式演进与破局。</summary> </entry> <entry><title>从SWE家族看CodeAgent评测发展路径和方向</title><link href="https://ivanbao9783.github.io/posts/agent-eval-05-swe-family/" rel="alternate" type="text/html" title="从SWE家族看CodeAgent评测发展路径和方向" /><published>2026-07-20T15:51:37+08:00</published> <updated>2026-08-19T11:36:12+08:00</updated> <id>https://ivanbao9783.github.io/posts/agent-eval-05-swe-family/</id> <content type="text/html" src="https://ivanbao9783.github.io/posts/agent-eval-05-swe-family/" /> <author> <name>ivanbao9783</name> </author> <category term="技术笔记" /> <summary>从 SWE 家族（SWE-bench、SWE-bench Verified、SWE-Marathon、FrontierSWE 等）看 Code Agent 评测的发展路径和方向。</summary> </entry> </feed>
