Minji Lee

Masters candidate @ UCSF

[Minji Lee @federicorestre1]

Watching papers fan out in the clip feels like my desk on Friday. Jev for Science arrives as ScienceBuddy automates literature pulls via #ScienceBuddy right away. It offers a clear window into #AIforScience tools.

[NeonPunk @SuhaSiddiqui1]

Honestly the nicest part is being able to experiment without feeling like every experiment means starting a brand new project.

If you’ve been curious, Nano Banana / GPT Image / Qwen Image are 30% OFF for a limited time.

[Atria ASI @AtriaASI]

Introducing Atria Dawn Preview:
From Research Questions to Verifiable Results.

Atria Dawn Preview is an agentic foundation model built for long-horizon tasks, helping researchers and engineers turn open-ended questions into executable, verifiable, and reproducible outcomes.

Explore Atria Dawn Preview:
🌐 Website: https://t.co/MDDspFiqYO
💻 GitHub: https://t.co/tNmLocadjT
🤗 Hugging Face: https://t.co/cyZod5oHbw
🤖 ModelScope: https://t.co/6BNi5vHFDl
🚀 Try Atria:
EN：https://t.co/Ap5aO2gQeu
ZH：https://t.co/cG16cQdx4j

[Zhaochen Yu @yu_trafalgar]

Can frozen LLM agents keep improving on long-horizon tasks? Yes — if the memory evolves instead of the weights.

Introducing Recuris: Recursive Experiential–Working Memory Evolution for Long-Horizon Agent Harnesses.

As agents take on longer tasks, the common strategy is to accumulate skills and procedures in memory and retrieve them as the task unfolds. But long-horizon execution breaks this in two places. Retrieval degrades: the initial instruction stops describing the current problem, and the growing history mixes completed steps, stale info, and noise. And improvement is blind: memory updates are admitted by a single downstream score, which says something broke — never what.

So we built Recuris: the agent uses its experience against a verified task state, and improves that experience from the evidence the state leaves behind.

Skill invocation no longer searches the full history. Working Memory tracks which goals are done and which remain pending; a state change commits only when environment feedback supports it, and skills are invoked from Experiential Memory based on what is still missing.

Every run leaves a Structured Trace linking task state, skill, action, and observation. Fault localization from this trace reaches 64.8% — versus 13.0% from the outcome alone, below the 33.3% always-guess floor.

A fixed Meta-Agent repairs only the components the trace implicates, gated by held-out validation. Accepted updates change future execution, producing new evidence for the next round — with the base LLM frozen throughout.

Across 4 long-horizon benchmarks and 10 models, Recuris improves task success in 35 of 37 model–benchmark pairs. On τ²-Retail, GPT-5.6 Sol gains +17.8 points and Claude Opus 5 gains +15.6, reaching 87.9% — 9.7 points above the best any model achieves without Recuris.

📄 Paper: https://t.co/0DkJTJoN3F
💻 Code: https://t.co/kpmsGBlCn3

[Monera_Digital @Monera_Digital]

冲突进入第九天，霍尔木兹的商船已经开始被迫改道——地缘溢价还在，但黄金没有跟涨，这背后的逻辑值得琢磨。今日宏观简报奉上👇

【地缘政治｜美伊冲突持续升级】

伊朗"胜利2"第23波行动分三阶段展开，打击范围覆盖巴林萨基尔空军基地无人机维保设施、萨勒曼港目标及科威特阿里夫詹美军基地设施。美方伤亡持续累积：自7月7日以来近百名美军受伤（多为轻微脑震荡，96%已返岗），自2月28日冲突爆发以来累计17人阵亡、400余人受伤。特朗普在社交媒体强硬回应，称伊朗每杀一名美军将付出数倍代价，并已向国防部长及参联会主席下达指令。美东时间今日下午4时，美军按总统指示对伊朗发动新一轮打击，目标直指削弱伊朗袭击霍尔木兹海峡商船的能力。

【霍尔木兹海峡｜航运风险持续发酵】

伊朗证实19日有商船在"被迫"通行南部水域时发生爆炸。美军中央司令部称已拦截并迫使7艘商船改道，另有1艘商船丧失航行能力。沙特主导的多国联军同步宣布加强曼德海峡护航，对胡塞武装威胁予以回击。海湾航运风险溢价料将维持高位。

【全球市场】

美股三大指数小幅收跌（纳指-0.05%，标普-0.19%，道指-0.59%），但AI算力与加密概念股逆势走强——IREN涨超19%，Hut 8涨超10%，Bit Digital涨10%，Galaxy Digital涨超9%；中概股普涨，阿里+4.7%，腾讯+3.6%。日经225涨0.78%，韩国KOSPI继昨日大跌后再跌超1%，三星、SK海力士震荡整理。

【大宗商品与能源】

地缘溢价持续支撑油价：WTI 8月合约收涨近0.9%报83.23美元/桶，布油9月合约收涨1.27%报89.22美元/桶。美国SPR上周降至3.114亿桶，为1983年3月以来最低水平，战略储备缓冲空间进一步收窄。现货黄金微跌至4008美元/盎司附近，避险与获利了结情绪交织。

社群Telegram：https://t.co/7WlnQBGSML
欢迎来社群跟我们一起聊聊你的看法～

[Minji Lee @federicorestre1]

@GustavoBolivar 100x100 de acuerdo!!