KK Blog

Menu

Archives

2026 / 10

Kimi K2.6 Inference Tuning Starts With Workload Shape

2026-10-04

最佳实践 #Kimi K2.6#LLM Inference#Coding Agents#KV Cache#Speculative Decoding#SGLang#GPU Performance
A workload-first playbook for Kimi K2.6 inference: measure request shape, find the current bottleneck, then tune decode, cache, and routing.
AI 研究 Agent 选对赢家，也可能不懂系统

2026-10-04

解读 #AI研究Agent#WhatWorkedBench#Agent评测#实验设计#响应面#科研AI
WhatWorkedBench 说明，找到最佳配置只能证明搜索成功，无法证明研究 Agent 理解了未观测配置、组件效应和交互关系。
AI Research Agents Can Pick Winners Without Understanding

2026-10-04

解读 #AI Research Agents#WhatWorkedBench#Agent Evaluation#Experimental Design#Response Surface#Scientific AI
WhatWorkedBench shows why finding the best configuration cannot prove that a research agent understands the experimental system.
自动化红队正在变成安全数据工厂

2026-10-03

解读 #自动化红队#LLM安全#对抗训练#提示注入#安全数据#AI评测
自动化红队的长期价值来自把攻击转成可验证训练数据、隔离回归集和可度量的防御改进，而非单纯增加攻击数量。
Automated Red Teaming Is Becoming a Security Data Factory

2026-10-03

解读 #Automated Red Teaming#LLM Security#Adversarial Training#Prompt Injection#Security Data#AI Evals
Automated red teaming creates lasting value when attacks become verified training data, isolated regression sets, and measurable defense improvements.
Failure Map：把边界错误变成可复现的 Coding Agent 评测

2026-10-02

最佳实践 #Failure Map#Coding Agent评测#程序修复#边界条件#隐藏测试#数据泄漏#AI编程
Failure Map 展示了如何把边界 Bug 转成可执行任务，同时避免把公开 checks 误当作隐藏评测。
模型的权威偏差：为什么已验证来源会推翻正确答案

2026-10-02

解读 #LLM权威偏差#RAG评测#来源验证#AI安全#检索增强#工具输出#NeurIPS 2026
已验证来源标签可能推翻模型原本正确的答案。RAG 系统需要 Claim 级证据与冲突测试，来源声望本身不等于正确。
Failure Map: Turn Boundary Bugs into a Reproducible Coding-Agent Evaluation

2026-10-02

最佳实践 #Failure Map#Coding Agent Evaluation#Program Repair#Boundary Cases#Hidden Tests#Data Leakage#AI Coding
Failure Map shows how to turn boundary-case bugs into executable coding-agent tasks without confusing public checks with hidden evaluation.
Authority Bias in LLMs: When a Verified Source Defeats a Correct Answer

2026-10-02

解读 #LLM Authority Bias#RAG Evaluation#Source Verification#AI Safety#Retrieval#Tool Outputs#NeurIPS 2026
A verified-source label can overturn correct LLM answers. RAG systems need claim-level evidence and conflict tests, not source prestige alone.
Claude-Shaped Science: Build the Research Harness Around Verifiability

2026-10-02

解读 #Claude-shaped science#BootLoops#Research Agents#Scientific AI#Harness Engineering#Verification#Anthropic
Claude-shaped science works when research problems, executable checks, and expert judgment fit the model's real capabilities.
Previous Next