Fanqing Meng

I am so confused that some says research and engineer separately To be a Good Engineer , Then learn to become Researcher

ARC-AGI-3 is very funny because somewhere along the way the benchmark design converged to "bespoke puzzle video games"

@kalomaze it all comes full circle

We scored 36.08% on ARC-AGI-3 in one day using the Agentica SDK.

为什么我的cc最近总是不遵循plan的步骤，总是跳步。。。

cron / 给目标（autoresearch）

罗姐不亏是雷总从 DS 挖来的，霸气直接没接杨总的话哈哈哈，再说一个冷知识， Kimi 的商标是小米转让给了月之暗面，具体交易金额不晓得。😄

Now I will use apple watch which i buy it 3 years ago but never use 😂

@DLKFZWilliam2 这个产品名字是什么

受不了了，太赛博朋克了。 一边在现实世界度假，一边在混合现实里跟朋友打球。 不说别的，Meta的这些头戴设备的那个摄像头，真的特别像赛博朋克里面的那种改装的眼睛

superpower会经常向我qa，让我有种还在掌控他的感觉qwq

2016年我提出90后不买房不买车不结婚，把所有时间用于自我提升与科技创新，2026年我提出，能和AI聊天就不要和人类聊天，删除所有90后之前出生人的联系方式，千万不要沾染任何老登气息，时间宝贵！全力拥抱未来！

right

MetaBot 现在支持微信了！通过 ClawBot 插件，直接在微信里和 Claude Code Agent 对话——写代码、读文档、跑命令，手机上就能搞定。 飞书、Telegram、微信三端打通，同一个 AI 团队随时随地协作。 一行命令安装，扫码即用： curl -fsSLhttps://raw.githubusercontent.com/xvirobotics/metabot/main/install.sh GitHub: github.com/xvirobotics/me…

Today, we are officially opening the capability to integrate #OpenClaw into #Weixin. With the launch of the #WeixinClawBot, users can use Weixin as a dedicated messaging channel for OpenClaw. Now, you can send and receive messages with OpenClaw just like texting a friend. #AIAutomation #AI

Attention Residuals Revisited kexue.fm/archives/11664

this is why i still use cursor 😂😂

Gym-V is fully open-sourced. 5 lines of code to get started: env = gym_v.make("Task-v0") obs = env.reset() action = agent(obs) obs, reward, done, _ = env.step(action) 📄 Paper: arxiv.org/abs/2603.15432 💻 Code: github.com/ModalMinds/gym… Let's build the Gym for vision agents, together!

Text agents have their Gym. Vision agents? Not until now. Introducing Gym-V — a unified gym-style platform for agentic vision research, with 179 procedurally generated environments across 10 domains. One API to rule them all: 📦 Offline dataset 🤖 Agentic RL training 🔧 Tool-use training 👥 Multi-agent training 📊 VLM & T2I model evaluation All under the same reset/step interface. Key findings: 1. Observation scaffolding matters MORE than RL algorithm choice 2. Broad curricula transfer well; narrow training causes negative transfer 3. Multi-turn interaction amplifies everything 📄 Paper: arxiv.org/abs/2603.15432 💻 Code: github.com/ModalMinds/gym… Open the thread for a deep dive! 🧵

Does RL training on one domain help others? ✅ Broad curricula (Cognition, Puzzles) transfer broadly — covering diverse sub-skills pays off ❌ Narrow curricula (Geometry) can cause NEGATIVE transfer — domain-specific shortcuts actively hurt on new tasks Transfer is asymmetric: Logic → Cognition yields +11.0, but Cognition → Logic only +5.8. Some competencies act as prerequisites rather than interchangeable skills. Multi-turn amplifies everything — both the gains AND the damage.

Some finding: Observation scaffolding is the most decisive factor for RL training success — more than algorithm choice. ✅ Adding captions to images → consistent improvement across ALL environments ❌ Removing game rules → can kill learning entirely ⚖️ GRPO vs GSPO vs SAPO? All improve, but no single algorithm dominates HOW you present the task to the agent matters more than HOW you optimize it.

We evaluated 9 VLMs zero-shot across all categories. 🏆 Gemini-3-Pro dominates (73.1 avg) 🥈 Best open model Qwen3-VL-32B reaches only 36.2 📊 Newer 32B beats older 72B by 1.8× — training recipe > raw scale The "difficulty cliff" is striking: on some tasks, accuracy drops to near-zero when complexity increases just one level. Even frontier models collapse — Gym-V is far from saturated.

Gym-V spans 10 categories: 📐 Single-Turn (105 envs): Algorithmic, ARC, Cognition, Geometry, Graphs, Logic, Puzzles 🎮 Multi-Turn (74 envs): Games, Spatial (2D/3D), Temporal (retro arcade) All environments are procedurally generated with deterministic seeding and parametric difficulty levels (0, 1, 2). From Sudoku to Sokoban, from Chess to Streets of Rage — vision agents face real visual reasoning challenges.