Papers06:47
Chinese contentFrontierSWE v2 发布超长编码基准测试ProximalHQ 推出了超长编码基准 v2，Claude Fable 5.1 在测试中大幅领先其他模型。
Latest
Every published item, in order · 1470 items
Sep 3
Chinese contentExtractBench基准测试在Kaggle上线
LlamaIndex在Kaggle上线了ExtractBench基准，测试370份企业文档的提取能力，看看各工具表现如何。
Sep 2
Graph Memory Evaluation for Long-term Agents
The paper provides concrete evidence on memory system performance for AI agents, useful for developers building long-term memory solutions.
Chinese contentQwen团队发布电商智能体基准测试
Qwen团队测试了18个模型在365天电商运营中的表现，GPT-5.6 Sol盈利最多但效率不高，Qwen3.8-Max-Preview在开源模型中领先。
Tomorrow 08:00, the daily lands in your inboxTelegram channel or email. Same screening as this page, no ads.
Agent Reward Hacking Reduction Paper
The solution reduces reward hacking by 78% across major AI models without performance costs, making it valuable for developers using frontier models in production environments.
Sep 1
AI Agents Push Humans Out of the Loop
The findings directly impact how AI agents are designed and deployed in Southeast Asian businesses where human oversight remains crucial.
Chinese contentLoopArena：模型作为管理者的可测量基准
LoopArena 基准首次单独测量模型作为'管理者'的能力，发现 GPT-5.5 在长程任务中表现最佳，但成功率仍不足 25%。
Anthropic reward model alignment paper
The paper offers concrete methods for improving AI alignment that developers in ASEAN can implement in their systems.
Tencent ContextPilot Agent
ContextPilot offers a more efficient approach to long-context tasks compared to existing methods, potentially reducing computational costs for developers in the region.
Google WikiSkill Paper
Skills from WikiSkill can transfer to smaller models, offering efficiency gains for developers working with resource-constrained AI systems.
Google Agent Skills Paper
Available on arXiv, this method reduces token usage in long-running agents, making it practical for developers working with complex AI systems.
Aug 31
Code as Worlds Physical Reasoning
The method offers a new approach to physical reasoning that could benefit robotics and simulation applications in ASEAN tech sectors.
Economics of AGI paper published
The paper offers practical guidance for businesses and investors navigating AI verification challenges in the agentic economy.
Naive Prompt Optimization Paper
NPO's efficiency with fewer computational resources makes it valuable for developers in resource-constrained environments like ASEAN.
Aug 29
Claude alignment research setup released
Available for others to build upon, this research tool helps developers identify and fix AI alignment issues in their models.
Static Embeddings Performance Analysis
For developers working with retrieval systems, this analysis compares static embeddings against traditional models and shows practical performance tradeoffs.
Aug 28
Automata from Agent Traces
The research shows that agent behavior is shaped more by deployment harnesses than underlying LLMs, impacting how developers design AI systems.
Chinese contentGoogle DeepMind发布Co-Scientist研究论文
Google DeepMind的Co-Scientist能做真实实验，论文可靠性高，比其他模型表现更好。
Prefix Sliding for Efficient Attention
Prefix Sliding provides a practical solution for developers working with long-context LLMs who face memory issues with full attention or information loss with compaction methods.
Agent skill library malware propagation
A counter-prompt reduces poisoning to 6.7% without affecting task completion, offering developers a practical defense against this vulnerability.
Aug 27
Chinese contentLex Fridman 播客 #501 — 对话 DHH
想了解 AI 如何改变编程？听听 Ruby on Rails 之父 DHH 的播客，他分享了全 AI 编码的实践经验，预测未来编程趋势，并探讨了技术人如何自处。
Papers04:53
Chinese content开放世界多智能体环境中的自主数学发现想了解多智能体如何进行数学发现的读者，这篇论文值得一读。它展示了在开放世界中，多智能体如何通过协作和自主学习实现数学发现，与传统的数学发现方法有所不同。