new paper: RLVR Exploration-Exploitation🫨 we carefully studied several interesting open questions regarding RLVR dynamics in entropy and clipping mechanism... p.s. "uncontaminated" Llama could also improve under random reward, while Qwen-Math might also fail to improve.

Check our #NeurIPS2025 work ComPO, a new preference alignment scheme that 😋Effectively extracts signals from noisy pairs that DPO cannot utilize; 🤯"Micro-finetuning": Only updates 0.0002% of weight! 😎Resolves likelihood-displacement ✨Novel zeroth-order LLM training method:

Introducing The Most Advanced Memory System for LLM Agents MIRIX is by far the most advanced memory system in the world, designed to make AI truly remember, learn, and help you over time. Website: mirix.io/ Paper: arxiv.org/abs/2507.07957 Github: github.com/Mirix-AI/MIRIX 1️⃣ Performance Benchmarks On ScreenshotVQA: 410% improvements over Gemini with 93.3% storage reduction; 35% improvements over RAG with 99.9% storage reduction! On LOCOMO, we obtained 85.4% accuracy, the current state-of-the-art performance (better than Mem0, MemOS, Zep, ...) 2️⃣ Six Core Memory Types (Inspired by Human Cognition) 1. Core Memory: Stores the AI’s personality and your long-term preferences: tone, settings, identity. 2. Episodic Memory: Your personal “event log.” 3. Semantic Memory: Facts, concepts, and your social graph. 4. Procedural Memory: How-tos and workflows. 5. Resource Memory: Your raw materials, files, notes. 6. Knowledge Vault: Verbatim information, your addresses (less sensitive), ID numbers (highly sensitive), encrypted with multi-level access controls. 3️⃣ Multi-Agent Workflow We have eight different agents: ⚙️ Meta Memory Manager: Analyzes inputs, then routes them to the right memory modules. ⚙️ Memory Managers: Specialized for each memory type, insert, deduplicate, merge, index, embed data. ⚙️ Chat Agent: Interfaces with the user, synthesizes context-aware responses Then during memory update: (1) Checks if similar data already exists; (2) Let meta memory manager analyze and route to different agents, (3) After that, Each module structures, merges, and stores in parallel. (4) Finally, all modules report back when done. 💡 This is why we call it an AI “cognitive prototype.” It’s not just an index. It’s the first step toward a system that remembers, reasons, and grows with you.

A Personal AI Assistant that Sees What You See and Understands You - Powered by Our Most Advanced Memory System 🧠 Imagine having a personal AI that truly remembers you. Not just another chatbot. Not just a search box. A real assistant that sees what you see, understands what you care about, and helps you turn scattered information into clarity and action. Today, we’re introducing MIRIX, the first AI personal assistant powered by the world’s most advanced memory system. 📈 Why does this matter? We live in an era of overwhelming information. Most AI tools can only answer questions in the moment. But real productivity—and real understanding—comes from memory. That’s why we built a system that doesn’t just process your input but continuously learns, organizes, and evolves alongside you. 💡 What can you do with it? Instantly recall anything you’ve seen or worked on; Draft and polish applications, reports, or emails; Analyze documents and get concise summaries; Keep your knowledge organized automatically; Never lose important insights again... 🔒 Your data stays yours. Everything you share is stored securely and privately. You control what to record and what to delete—always. 🚀 We believe this is the future of AI: A true second brain that doesn’t just respond, but remembers and understands. Ready to try it? 👉 mirix.io/ 🛠️ Curious how this memory system works? In the next post, I’ll share benchmarks and the technical story behind it. #AI #PersonalAssistant

Yap early, yap only, yap often. @KaitoAI is connecting AI, attention and capital with Yaps. Just claimed my social card and I'm accumulating Yap points in real-time. Claim yours 👉 yaps.kaito.ai/referral/11323…

New research alert! 📝Spectral Policy Optimization" that helps LLMs learn from ALL responses, even when they're ALL wrong! Solves a major GRPO limitation and improves model reasoning across multiple sizes. Read it here: arxiv.org/abs/2505.11595 #AI #reasoning #LLM

Thrilled to attend the "AI Agents Meme Coins" panel at #CryptoAISummit2025! by Bitpush🚀 mp.weixin.qq.com/s/C90HFxQZy… Transparency and verifiability are key for the AI finance future! #AI #Agents #Meme #Coins #Bitpush

🚨 Get ready for CryptoAI Summit 2025 – where AI meets Web3 in NYC! 🗓️ May 10 | Full day | 6 powerful panels One of the most talked-about topics? 🎙️ Panel 6: “AI Agents & Meme Coins – The New Market Makers?”

At our latest CDS Seminar, CDS-affiliated @NYUStern's Prof. Xi Chen (@Prof_XBlock) discussed privacy-preserving dynamic pricing that protects customer data, and a new blockchain mechanism for decentralized AI training markets. His work bridges privacy, ML, and blockchain tech.

My recent work on Proof-of-Learning arxiv.org/abs/2404.09005 (A new version with substantially improved theoretical results is coming!)

What do #NFTs & Chat GPT-4 have in common? Attend #NFTNYC2023 and hear from @Prof_XBlock (Professor, @nyuniversity) on how AI & blockchain technologies intersect. Check out the program here: NFT.NYC/program

visiting louvre museum in abu dhabi

Teaching at Abu Dhabi, learn GPT-4 last night and teach in the morning

Our ZKP MOOC (zk-learning.org) has been really exciting with ~3500 enrolled globally! Today we're thrilled to announce a new ZKP/web3 hackathon, open to the public, March 1-May 2, virtual & in-person @UCBerkeley, with prizes up to $200K ! Sign up: rdi.berkeley.edu/zkp-web3-ha…

Check out this overview of DEX LP RISK-HEDGE strategy, focusing on the IL perspective and LP as options perspective.

I'm looking forward to seeing @Prof_XBlock and other Speakers on the Brands Track at #NFTNYC2023 nft.nyc/speakers/brands?trac…

My new paper on Transaction Packaging in High-throughput Blockchains: arxiv.org/abs/2301.10944. We propose a strategic coordination mechanism for the decentralized transaction packaging process by using a game-theoretic approach.

privacy: the next big thing in crypto a16zcrypto.com/big-ideas-202… #privacy #zkevm #Crypto