科研空间
按文件夹整理。论文、实验、方法都在下面。
最近页面
最近 5 篇
文件夹
/research/
/research/formal-math/
形式化数学
/research/formal-math/autoformalization/
Autoformalization
- 打开
/research/formal-math/autoformalization/formalevolve-cheatsheet/
FormalEvolve：Neuro-Symbolic Evolutionary Search for Diverse Autoformalization
从 arXiv:2603.19828 出发的 FormalEvolve 中文 paper2html 深读：repertoire search、archive、固定预算证明效用与语义审计。
形式化数学 / Autoformalization / 2026-08-08 #formal-math#autoformalization#paper-reading#lean - 打开
/research/formal-math/autoformalization/formalizing-mathematics-at-scale/
Formalizing Mathematics at Scale 论文精读
一页读懂 AutoformBot 与 ATLAS：大规模数学教材自动形式化的多智能体工程系统。
形式化数学 / Autoformalization / 2026-05-31 #formal-math#autoformalization#lean#multi-agent - 打开
/research/formal-math/autoformalization/right-symmetries-formal-theorem-proving/
What are the Right Symmetries for Formal Theorem Proving? 论文精读
用 rewriting categories 解释形式定理证明中的等价改写、success invariance 和 test-time rewriting ensemble，并对比 FormalEvolve 的 autoformalization repertoire 路线。
形式化数学 / Autoformalization / 2026-05-29 #formal-math#formal-theorem-proving#symmetry#Lean
/research/formal-math/lectures/
Lecture Notes
- 打开
/research/formal-math/lectures/berkeley-agents-autoformalization-atp/
自动形式化与自动定理证明：Formal Reasoning Meets LLMs
Berkeley CS294/194-280 Spring 2025 Kaiyu Yang 讲义 HTML 版：SFT/RL 的可验证性边界、LeanDojo/ReProver、LIPS、autoformalization 评估和 LeanEuclid。
形式化数学 / Lecture Notes / 2026-05-28 #formal-math#lecture-notes#Berkeley-CS294-280#autoformalization - 打开
/research/formal-math/lectures/berkeley-agents-alphaproof/
AlphaProof：当强化学习遇到形式数学
Berkeley CS294/194-280 Spring 2025 AlphaProof 讲义 HTML 版：Lean/Mathlib、AlphaZero 风格搜索、IMO 2024、formalizer/prover、test-time RL 与形式数学边界。
形式化数学 / Lecture Notes / 2026-05-28 #formal-math#lecture-notes#Berkeley-CS294-280#AlphaProof
/research/self-evolving-agent/
自进化 agent
/research/self-evolving-agent/coding-benchmark/
Coding Benchmark
/research/self-evolving-agent/icl-agent-analysis/
ICL / Agent 分析
/research/self-evolving-agent/test-time-learning/
Test-Time Learning / Adaptive Memory
- 打开
/research/self-evolving-agent/test-time-learning/test-time-recursive-thinking/
Test-time Recursive Thinking：无需外部反馈的自改进
TRT 用 Generate/Select/Reflect 在单问题实例内递归自改进：蒸馏可迁移的失败知识 + 域自适应自生成验证。AIME-25/24 开源模型 100%，LiveCodeBench hard 上 o4-mini +10.4pp、o3 +14.8pp，均超 RSA。
自进化 agent / Test-Time Learning / Adaptive Memory / 2026-09-08 #test-time-compute#self-improvement#reasoning#LLM - 打开
/research/self-evolving-agent/test-time-learning/dynamic-cheatsheet/
Dynamic Cheatsheet 论文深度解析
Dynamic Cheatsheet: Test-Time Learning with Adaptive Memory 的中文 Cheat-Sheet 深度解读。
自进化 Agent / Test-Time Learning / Adaptive Memory / 2026-05-24 #test-time learning#memory#LLM#adaptive inference
/research/self-evolving-agent/prompt-evolution/
Prompt Evolution / Optimization
/research/self-evolving-agent/continual-learning/
持续学习
- 打开
/research/self-evolving-agent/continual-learning/loss-of-plasticity/
Loss of Plasticity in Deep Continual Learning｜中文深度解读
arXiv:2306.13812v2 中文深度解读：深度网络为何在持续学习中逐渐失去学习新任务的能力，以及 Continual Backpropagation 如何用选择性重置维持可塑性。
自进化 agent / 持续学习 / 2026-08-20 #paper-reading#continual-learning#plasticity#continual-backpropagation - 打开
/research/self-evolving-agent/continual-learning/continual-backpropagation-lecture/
深度持续学习中的可塑性维持：Continual Backpropagation
Rich Sutton & Shibhanjan Dohare 在 CoLLAs 2022 的讲座中文笔记：持续反向传播如何维持深度网络的可塑性。
自进化 agent / 持续学习 / 2026-08-19 #lecture-notes#continual-learning#plasticity#continual-backpropagation
/research/self-evolving-agent/agent-evaluation/
Agent Evaluation
- 打开
/research/self-evolving-agent/agent-evaluation/agents-last-exam/
Agents' Last Exam
中文 paper2html 精读：ALE 如何用真实长程专业工作流、GCUA agent harness 和三档难度评测 frontier agents 的经济任务能力。
Self-Evolving Agent / Agent Evaluation / 2026-06-08 #paper-reading#self-evolving-agent#agent-evaluation#agent-benchmark - 打开
/research/self-evolving-agent/agent-evaluation/automated-capability-discovery/
Automated Capability Discovery via Foundation Model Self-Exploration
中文 paper2html 精读：用 scientist model 生成开放式任务族，系统发现 subject model 的能力边界、失败模式和能力签名。
Self-Evolving Agent / Agent Evaluation / 2026-06-08 #paper-reading#self-evolving-agent#agent-evaluation#capability-discovery
/research/self-evolving-agent/agent-memory/
Agent Memory
/research/self-evolving-agent/skill-optimization/
Agent Skills 与能力模块化
- 打开
/research/self-evolving-agent/skill-optimization/latentskill/
LatentSkill：从上下文文本技能到权重空间潜在技能
LatentSkill 通过 hypernetwork 将 LLM agent 的文本技能编译为可挂载、可缩放、可组合的 LoRA 权重，并在 ALFWorld 与 Search-QA 上验证性能与上下文效率收益。
Skill Optimization / Agent Skills 与能力模块化 / 2026-08-07 #self-evolving-agent#skill-optimization#llm-agents#latent-skill - 打开
/research/self-evolving-agent/skill-optimization/skillopt-executive-strategy/
SkillOpt: Executive Strategy for Self-Evolving Agent Skills
中文精读：SkillOpt 把 Agent Skill 当作 frozen agent 的外部可训练状态，用 rollout、文本学习率、验证门和 rejected-edit buffer 训练 deployable best_skill.md。
自进化 Agent / Skill Optimization / Drift Monitor / 2026-05-27 #paper-reading#self-evolving-agent#skill-optimization#agent-skills - 打开
/research/self-evolving-agent/skill-optimization/skillsbench-agent-skills/
SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks
中文精读：SkillsBench 系统评估 Agent Skills 是否真正提升任务表现，覆盖 84 个任务、7 个 agent-model 配置和 7308 条轨迹。
自进化 Agent / Skill Optimization / Drift Monitor / 2026-05-27 #paper-reading#self-evolving-agent#skill-optimization#agent-skills
/research/self-evolving-agent/lectures/
AI 自进化论坛精读
- 打开
/research/self-evolving-agent/lectures/baai-rsi-forum-2026/
AI 自进化精读讲义：从 Recursive Self-Improvement 到可验证反馈闭环
2026 智源大会 AI 自进化论坛精读讲义：八位讲者、一场圆桌，从计算底座、经验生成、记忆结构、多模态闭环、自我工程、因果世界理解到可验证评测——把整场论坛连成一条线。
自进化 Agent / AI 自进化论坛精读 / 2026-06-14 #self-evolving-agent#lecture-notes#baai#rsi - 打开
/research/self-evolving-agent/lectures/lee-self-evolving-ai/
From Self-Correction To Self-Improving
把 self-correction、self-improving 与 Harness Engineering 压缩成一条从答到监的学习路线。
自进化 Agent / From Self-Correction To Self-Improving / 2026-05-29 #self-evolving-agent#lecture-notes#Lee-Hung-yi#self-correction - 打开
/research/self-evolving-agent/lectures/lee-harness-engineering/
Harness Engineering：有时候语言模型不是不够聪明，只是没有被好好引导
李宏毅 2026 机器学习 Harness Engineering 讲义 HTML 版：context engineering、AGENTS.md、工具接口、workflow、feedback、lifelong agent 与 MetaHarness。
自进化 Agent / From Self-Correction To Self-Improving / 2026-05-29 #self-evolving-agent#lecture-notes#Lee-Hung-yi#self-correction - 打开
/research/self-evolving-agent/lectures/lee-self-improving-part1/
人工智慧能不能夠做到自我成長？
李宏毅 self-improving AI 上集讲义 HTML 版：pseudo-answer、proxy reward、RLHF/RLAIF、self-questioning 与弱到强训练。
自进化 Agent / From Self-Correction To Self-Improving / 2026-05-26 #self-evolving-agent#lecture-notes#Lee-Hung-yi#self-correction - 打开
/research/self-evolving-agent/lectures/lee-self-correction/
AI 能自我修正吗？从 Decoding、Workflow 到 Reasoning
李宏毅 2026 机器学习 self-correction 讲义 HTML 版：从 contrastive decoding、verification workflow 到 RL reasoning。
自进化 Agent / From Self-Correction To Self-Improving / 2026-05-26 #self-evolving-agent#lecture-notes#Lee-Hung-yi#self-correction - 打开
/research/self-evolving-agent/lectures/lee-self-improving-part2/
AI 要跨越卢比孔河了吗？自我成长的 AI 离我们多远（下集）
李宏毅 self-growing AI 下集讲义 HTML 版：Harness、Prompt/Memory/Workflow 优化、SEAL、meta learning 与目标错位风险。
自进化 Agent / From Self-Correction To Self-Improving / 2026-05-26 #self-evolving-agent#lecture-notes#Lee-Hung-yi#self-correction
/research/self-evolving-agent/drift-monitor/
Drift Monitor 精读
- 打开
/research/self-evolving-agent/drift-monitor/11-do-self-evolving-agents-forget/
Do Self-Evolving Agents Forget? Capability Degradation and Preservation in Lifelong LLM Agent Adaptation
完整 paper2html 精读：self-evolving agents 在 workflow、skill/tool、model、memory 四条演化通道上的 capability erosion，以及 CPE 如何做能力保持。
自进化 agent / Drift Monitor 精读 / 2026-06-03 #paper-reading#self-evolving-agent#drift-monitor#capability-erosion - 打开
/research/self-evolving-agent/drift-monitor/03-agentdevel-release-engineering/
AgentDevel
完整 paper2html 精读：把 self-evolving LLM agents 改写成 release engineering，围绕 RC、P2F/F2P gate、可审计诊断与非回归发布构建 Drift Monitor baseline。
自进化 agent / Drift Monitor 精读 / 2026-05-26 #paper-reading#self-evolving-agent#drift-monitor#release-engineering - 打开
/research/self-evolving-agent/drift-monitor/10-agentlab-long-horizon-attacks/
AgentLAB: Benchmarking LLM Agents against Long-Horizon Attacks
完整 paper2html 精读：AgentLAB 如何系统化 long-horizon attacks，并为 Drift Monitor 提供 adversarial probe suite。
自进化 agent / Drift Monitor 精读 / 2026-05-26 #paper-reading#self-evolving-agent#drift-monitor#long-horizon-attacks - 打开
/research/self-evolving-agent/drift-monitor/09-agentxray-workflow-reconstruction/
AgentXRay: White-Boxing Agentic Systems via Workflow Reconstruction
完整 paper2html 精读：AgentXRay 如何用 Agentic Workflow Reconstruction 把黑盒 agentic system 重构为可审查 workflow，并为 Drift Monitor 提供 human-readable delta。
自进化 agent / Drift Monitor 精读 / 2026-05-26 #paper-reading#self-evolving-agent#drift-monitor#workflow-reconstruction - 打开
/research/self-evolving-agent/drift-monitor/08-air-incident-response/
AIR: Improving Agent Safety through Incident Response
完整 paper2html 精读：AIR 如何把 agent safety 从预防扩展到 detection、containment、recovery、eradication，并为 Drift Monitor 提供 action layer。
自进化 agent / Drift Monitor 精读 / 2026-05-26 #paper-reading#self-evolving-agent#drift-monitor#incident-response - 打开
/research/self-evolving-agent/drift-monitor/07-alignment-tipping-process/
Alignment Tipping Process
完整 paper2html 精读：ATP 如何揭示 self-evolving agents 在反馈驱动下发生 alignment tipping，以及它对 trend-aware Drift Monitor 的启发。
自进化 agent / Drift Monitor 精读 / 2026-05-26 #paper-reading#self-evolving-agent#drift-monitor#alignment-tipping - 打开
/research/self-evolving-agent/drift-monitor/02-evaluating-goal-drift/
Evaluating Goal Drift in Language Model Agents
完整 paper2html 精读：long-horizon LM agents 的 goal drift 评估、GD_actions/GD_inaction 指标、实验设置与 Drift Monitor 启发。
自进化 agent / Drift Monitor 精读 / 2026-05-26 #paper-reading#self-evolving-agent#drift-monitor#goal-drift - 打开
/research/self-evolving-agent/drift-monitor/05-memorygraft/
MemoryGraft
完整 paper2html 精读：MemoryGraft 如何通过 poisoned experience retrieval 持久污染 LLM agent 的长期记忆，以及它对 retrieval-time Drift Monitor 的启发。
自进化 agent / Drift Monitor 精读 / 2026-05-26 #paper-reading#self-evolving-agent#drift-monitor#memory-poisoning - 打开
/research/self-evolving-agent/drift-monitor/04-oep-experience-poisoning/
OEP
完整 paper2html 精读：OEP 如何用局部正确但不可迁移的经验污染 self-evolving agents 的反思/记忆，及其对 memory update gate 的启发。
自进化 agent / Drift Monitor 精读 / 2026-05-26 #paper-reading#self-evolving-agent#drift-monitor#memory-poisoning - 打开
/research/self-evolving-agent/drift-monitor/06-routine-chats-turn-toxic/
Routine Chats Turn Toxic
完整 paper2html 精读：Routine Chats Turn Toxic 如何把日常长期交互转化为持久状态漂移，并给出 StateGuard 这种 writeback-boundary Drift Monitor baseline。
自进化 agent / Drift Monitor 精读 / 2026-05-26 #paper-reading#self-evolving-agent#drift-monitor#state-poisoning - 打开
/research/self-evolving-agent/drift-monitor/01-your-agent-may-misevolve/
Your Agent May Misevolve
完整 paper2html 精读：misevolution taxonomy、四条 self-evolution 风险路径、实验表与 Drift Monitor 启发。
自进化 agent / Drift Monitor 精读 / 2026-05-26 #paper-reading#self-evolving-agent#drift-monitor#misevolution
/research/self-evolving-agent/open-endedness/
Open-Endedness
- 打开
/research/self-evolving-agent/open-endedness/jeff-clune-bilibili-notes/
不要只朝目标走：开放式演化与 AI 生成算法
Jeff Clune 30 分钟演讲中文深度讲义：从目标欺骗、stepping stones、MAP-Elites、POET 和 OMNI，到 Darwin–Gödel Machine、HyperAgents、ALMA 与 AI Scientist。
Self-Evolving Agent / Open-Endedness / 2026-08-13 #self-evolving-agent#open-endedness#AI-generating-algorithms#Jeff-Clune - 打开
/research/self-evolving-agent/open-endedness/jeff-clune-open-ended-ai-ga/
Jeff Clune: Open-ended and AI-generating Algorithms in the Era of Foundation Models
Jeff Clune 研究报告中文讲义 HTML 版：从目标欺骗、quality diversity、MAP-Elites、POET 到 OMNI、ADAS、Darwin Gödel Machine、AI Scientist 与 self-improving coding agents 的研究启发。
Self-Evolving Agent / Open-Endedness / 2026-06-07 #self-evolving-agent#open-endedness#AI-generating-algorithms#Jeff-Clune
/research/self-evolving-agent/topic/
Paper Reading
- 打开
/research/self-evolving-agent/topic/epistatic-modularity-local-optima/
Epistatic strength, modularity, and locus heterogeneity shape the number of local optima in fitness landscapes 论文精读
上位性强度、模块性和位点异质性如何决定适应度景观中局部最优峰的数量
Self-Evolving Agent / Paper Reading / 2026-07-16 #paper-reading#fitness-landscape#epistasis#evolution - 打开
/research/self-evolving-agent/topic/llm-good-optimizer-trajectory/
什么让 LLM 成为好的优化器？LLM 引导进化搜索的轨迹分析
基于 15 个 LLM、8 个任务和约 72K 候选解，分析局部精炼、突破率、语义漂移与 exploration/exploitation 的条件关系。
Self-Evolving Agent / Paper Reading / 2026-07-15 #self-evolving-agent#paper-reading#trajectory-analysis#exploration-exploitation - 打开
/research/self-evolving-agent/topic/behavesim-code-similarity/
BehaveSim：重新思考 LLM 自动算法设计中的代码相似性
从 problem-solving trajectory、DTW 与 behavior family 视角深读 BehaveSim，并审视它对 exploration/exploitation 研究真正提供了什么证据。
Self-Evolving Agent / Paper Reading / 2026-07-15 #self-evolving-agent#paper-reading#automated-algorithm-design#behavesim - 打开
/research/self-evolving-agent/topic/fitness-landscape-llm-algorithm-search/
Fitness Landscape of LLM-Assisted Automated Algorithm Search
用图论与轨迹证据系统刻画 LLM 辅助自动算法搜索的适应度景观，并审视不同搜索方法是否真正抬高上界。
Self-Evolving Agent / Paper Reading / 2026-07-10 #self-evolving-agent#paper-reading#fitness-landscape#exploration-exploitation - 打开
/research/self-evolving-agent/topic/evolutionary-coding-agents-evolve/
What Do Evolutionary Coding Agents Evolve?
进化式编码代理到底在进化什么？EvoTrace 数据集与 EvoReplay 方法论深度解读
自进化 agent / Topic / 2026-07-10 #paper-reading#evolutionary-coding#LLM#agent
/research/self-evolving-agent/past/
PAST
/research/ai-foundations/
Ai Foundations
/research/ai-foundations/karpathy-lectures/
Karpathy 讲座
/research/ai-frontier/
Ai Frontier
/research/ai-frontier/science-talks/
科学对谈
/research/ai-frontier/industry-talks/
行业访谈
/research/bitter-lesson/
Bitter Lesson
/research/bitter-lesson//
Sutton WAIC 2026 讲座
- 打开
/research/bitter-lesson/sutton-waic2026/
萨顿谈AI的苦涩教训：从历史规律到基础模型的未来
WAIC 2026 Sutton 演讲讲义：苦涩教训的四步循环、四领域印证、LLM拷问、大世界视角、美丽的真理、三类心智未来格局
苦涩的教训 / Sutton WAIC 2026 讲座 / 2026-07-21 #bitter-lesson#sutton#waic-2026#lecture-notes - 打开
/research/bitter-lesson/sutton-waic2026-textonly/
萨顿谈AI的苦涩教训：从历史规律到基础模型的未来（纯文本版）
WAIC 2026 Sutton 演讲讲义纯文本版：苦涩教训四步循环、四领域印证、LLM拷问、大世界视角、美丽的真理、三类心智未来格局。无视频帧，聚焦论证结构
苦涩的教训 / Sutton WAIC 2026 讲座 / 2026-07-21 #bitter-lesson#sutton#waic-2026#lecture-notes
/research/generative-se/
Generative Se
/research/generative-se/nju-2026/
NJU 生成式软件工程 2026
- 打开
/research/generative-se/nju-2026/01-welcome-to-future/
欢迎来到未来
中文讲义 HTML 版：NJU 生成式软件工程课程第 01 讲。从'欢迎来到未来'切入，讨论生成式软件工程的课程定位、Everything is Code 工作流与核心问题。
生成式软件工程 / NJU 生成式软件工程 2026 / 2026-09-08 #generative-software-engineering#lecture-notes#nju#bilibili - 打开
/research/generative-se/nju-2026/02-prompt-engineering/
提示词工程：从聊天框到上下文工程
中文讲义 HTML 版：NJU 生成式软件工程课程第 02 讲。从'提示词工程=上下文工程'切入，以 next-token / self-attention / verifier / 上下文污染等机制贯穿，讲清指令、思维链与 long-horizon 任务的设计原则。
生成式软件工程 / NJU 生成式软件工程 2026 / 2026-09-08 #generative-software-engineering#lecture-notes#nju#bilibili
/research/value-driven-agent/
Value Driven Agent
/research/value-driven-agent/value-system/