NEW RESEARCH: Approximating Language Model Training Data from Weights ever wonder how much information is available in an open-weights model? DeepSeek R1 weights are 1.2 TB... what can we learn from all those bits? our method reverses LLM finetuning to recover data: 🧵

Benchmarks aside, It thinks: → 23 reasoning steps per task (avg.) → 200+ URLs explored → Multi-turn tool use of search, browser, and code → Inline citations Beta access is rolling out at kimi.com — get on the waitlist 👉 [docs.google.com/forms/d/e/1FAI…]

notion drops are back

Moonshot does it again, nice deep research + RL work moonshotai.github.io/Kimi-Researche…

Our first End2End RL Trained Agent is out! Hope you like it!

had a blast talking with himanshu about all things RL + agents :) 2hr vid is up now!

Reasoning, Attention & Memory Workshop @ COLM Submission Deadline: June 23, 2025 -- Today!

For the past ~2 months we have been working on training reasoning models on TextArena games. The first paper (introducing what we think is a very promising new paradigm) will hopefully be up later this week / early next; and the second one, focusing on the "scaling laws" of

very bullish for generative RMs super cool to see people training them at scale :)

More story line for the next merch drop

:)

📢 Can LLMs really reason outside the box in math? Or are they just remixing familiar strategies? Remember DeepSeek R1, o1 have impressed us on Olympiad-level math but also they were failing at simple arithmetic 😬 We built a benchmark to find out → OMEGA Ω 📐 💥 We found

some things we should CoT more about

word

holy shit, it’s here! deepmind just released AlphaGenome. an AI model that reads 1 million bases of DNA and predicts how any mutation changes molecular function not just in single genes but across the entire regulatory genome. DNA is code, and you are software 1/

🔥明天（2025.06.26）北京时间10:30到11:30，我将线上直播讲解我们Qwen团队和清华Leaplab团队的最近比较火的文章《Beyond the 80/20 Rule: High-Entropy Minority Tokens Drive Effective Reinforcement Learning for LLM Reasoning》 🌟欢迎感兴趣的朋友参加！ 直播报名链接：event.baai.ac.cn/activities/935

the CIA is not ready for the RL era israeli intelligence guy just hacked into a live surveillance camera in front of me with an exploit generated by qwen vulnerable software is simulatable. penetration success is verifiable. hacking is RLable.

1/ Introducing RL Swarm’s new backend: GenRL. A modular reinforcement learning library built for distributed, fault-tolerant training - now powering RL Swarm from the ground up. 🧵

We did it — SYNTHETIC‑2 is complete. A planetary-scale decentralized inference run generating 4M verified reasoning samples. 1,250+ GPUs joined in 3 days — from 4090s to H200s — creating data for complex RL tasks. Full open-source release + technical report coming next week!

introducing Cluely. today is the start of a world where you never have to think again. we just killed 9 industries (thread):