Pretraining has scaling laws to guide compute allocation. But for RL on LLMs, we lack a practical guide on how to spend compute wisely. We show the optimal compute allocation in LLM RL scales predictably. ↓ Key takeaways below

Really BIG MODEL SMELL

Moonshot’s Kimi K2.5 is the new leading open weights model, now closer than ever to the frontier - with only OpenAI, Anthropic and Google models ahead Key takeaways: ➤ Impressive performance on agentic tasks: Kimi.ai's Kimi K2.5 achieves an Elo of 1309 on our GDPval-AA

👀 🥝Full push for K3

🚤Real-Time Streaming VLA for Dynamic Manipulation🚤 #DynamicVLA is a 0.4B vision-language-action model that manipulates *moving* objects in real-time, with continuous inference and latent-aware action streaming - Project: infinitescript.com/project/dynami… - Code: github.com/hzxie/DynamicV…

𝗧𝗵𝗲 𝗔𝗜 𝗧𝗮𝗹𝗸𝘀 will be hosting SAM 3D (weiyaow ) and SAM 3D Body (Xitong Yang) from @MetaAI. 🕐 Feb 3 (Tue) - 13:00 SGT | Feb 2 (Mon) - 21:00 PST 📩 PM me for the Zoom link 🔔 Get notified of future talks The AI Talks: theaitalks.org/subscribe/

We're introducing WorldVQA, a new benchmark to measure atomic vision-centric world knowledge in Multimodal Large Language Models. Current evaluations often conflate visual knowledge retrieval with reasoning. In contrast, WorldVQA decouples these capabilities to strictly measure

✨Moving beyond static knowledge in Video Understanding! We are thrilled to unveil Demo-ICL🧠, a new framework that challenges current models to do more than just "remember"—we want them to learn from context. We introduce: 1️⃣ Demo-ICL-Bench📜: A massive, challenging

Can MLLMs learn from video demonstrations just like humans do? 🤔 Introducing 𝗗𝗲𝗺𝗼-𝗜𝗖𝗟: 𝗜𝗻-𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝗳𝗼𝗿 𝗣𝗿𝗼𝗰𝗲𝗱𝘂𝗿𝗮𝗹 𝗩𝗶𝗱𝗲𝗼 𝗞𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 𝗔𝗰𝗾𝘂𝗶𝘀𝗶𝘁𝗶𝗼𝗻 Most video MLLMs rely on static internal knowledge. This work

🤔In-Context Learning (ICL) in Video LLMs🤔 🎞️Demo-ICL🎞️ equips video LLMs with the ability to learn and adapt from *dynamic novel contexts from few examples*, rather than relying on the static internal knowledge. - Paper: arxiv.org/pdf/2602.08439 - Code: github.com/dongyh20/Demo-…

🚀Codec-Aligned Sparsity for Multimodal Intelligence🚀 lmms-lab presents 🎇OneVision-Encoder🎇, a *scalable, efficient & powerful vision encoder* for next-gen LMMs with streaming inputs 🧠Insight: codec-aligned, patch-level sparsity as a foundational principle 📊Performance:

Introducing GLM-5: From Vibe Coding to Agentic Engineering GLM-5 is built for complex systems engineering and long-horizon agentic tasks. Compared to GLM-4.5, it scales from 355B params (32B active) to 744B (40B active), with pre-training data growing from 23T to 28.5T tokens.

x.com/i/article/2021…

(1/n) Introduce On-Policy Context Distillation (OPCD), a framework to internalize transient in-context knowledge into model parameters via on-policy learning. This also launches our series, Experiential Learning -- Part I: On-Policy Context Distillation for Experiential Learning

On-Policy Context Distillation for Experiential Learning: learning from experience (consolidated from trajectories) at test time.

Looks like they are serious about +0.1 on the version 🤔 How about K2.6 👀