Zeyu Zheng

Can frontier AI agents achieve Recursive Self-Improvement by turning a weak method into one that performs better on hidden data? Introducing RSI-Exam 🧵: 88 executable research tasks across 6 domains, now open for task contributions from every field🧑‍🔬. Its task bank spans virtual cells, TPU kernels, chip design, quantitative finance, agent harnesses, model distillation, and much more. Across these tasks, every rollout follows the same core protocol: 🔁 An agent inherits a working method and iterates using visible data. 🔒 Only its final artifact enters a fresh-container evaluation on the hidden set. 📊 The result so far: Opus 5 leads the 88-task leaderboard with a mean hidden-set score of 0.464. 📢 Want to help build the next release? Author an executable task, review one, or audit a rollout. Substantial contributions qualify for paper authorship. Special thanks to @XinyeYee for the tremendous effort in leading this project. Explore: rsi-exam.ai/ Blog: rsi-exam.ai/blog.html Contribute: rsi-exam.ai/contribute.html GitHub: github.com/aiming-lab/RSI-Ex…

🧠 Multimodal SimpleMem is here (v0.2.0). SimpleMem now supports multimodal lifelong memory for LLM agents. - 🏆 LoCoMo F1: +47% - 🏆 Mem-Gallery F1: +51% 🔬 What makes this special: Omni-SimpleMem was built using AutoResearchClaw's Human-in-the-Loop Co-Pilot mode — human researchers and AI agents co-piloting research, experiments, and writing together. 💡 Three design principles: - 🔍 Selective Ingestion — entropy-driven filtering per modality - 🎯 Progressive Retrieval — hybrid FAISS + BM25 with pyramid token budgets - 🕸️ Knowledge Graph Augmentation — multi-hop cross-modal reasoning 📄 Paper: arxiv.org/abs/2604.01007 🔗 Code: github.com/aiming-lab/Simple… More detailed explanation of Omni-SimpleMem coming soon 👀 @JiaqiLiu835914 @YanqingLiu83931 @StephenQS0710 @lillianwei423 @richardxp888 @HaoqinT @ZhengBerkeley @cihangxie @dingmyu

🚀 Introducing AutoHarness (「Aha」) — automated harness engineering for AI agents. In LLM training, the aha moment is when a model learns to reason. For agents, it's when a better harness makes the same model shine. Agent = Model + Harness. The model reasons. The harness does everything else: 🧠 Context management 🛡️ Tool governance 💰 Cost control 👁️ Observability 💾 Session persistence These are the patterns that separate a toy from a system. AutoHarness automates this entire layer. 🔧 What's inside: - 6-step tool pipeline: parse → classify → permit → execute → sanitize → audit - 3 modes (Core / Standard / Enhanced) — from lightweight to full-featured - Smart context management with token budgeting and multi-layer compression - Full observability: per-call cost tracking, JSONL audit trail, trace diagnostics - Multi-agent profiles with role-based permissions - Any LLM provider Every agent deserves its aha moment. Led by @JiaqiLiu835914, and Kudos to the team @XinyeYee, @richardxp888, @lillianwei423, @HaoqinT @Xinyu2ML, @yuyinzhou_cs, @ZhengBerkeley, @dingmyu, @cihangxie, etc.

🦞MetaClaw v0.4.0 is here. Introducing the Evolving Contexture Layer: skills and memory now co-evolve. +18% over OpenClaw + GPT-5.2 on deployed CLI agent benchmarks. Your agent remembers across sessions. User preferences, project context, and interaction history are automatically retrieved and injected into every prompt. Skills adapt based on what it remembers. Memory sharpens based on what it learns. No GPU required. Just talk. It meta-learns and evolves.🧠 Built with @openclaw and @thinkymachines Kudos to the team @richardxp888, @JimChenjw, @HaoqinT, @lillianwei423, @StephenQS0710, @Xinyu2ML, @JiaqiLiu835914, @ZhengBerkeley, @cihangxie.

Agents can remember. Agents can learn. But can they learn how to learn? 🤔 MetaClaw tech report is out 📄 — a continual meta-learning framework that teaches deployed agents not just to learn, but to learn how to learn, evolving 24/7 through normal usage. 📊 On 588 continual CLI tasks over 14 simulated workdays: 🦞 Kimi-K2.5 accuracy: 21.1% → 39.6% (+88%) 🚀 Task completion: 18.2% → 51.9% (+185%) 💪 GPT-5.2 also benefits: 44.9% → 49.1% acc, 58.4% → 67.5% completion Even the strongest models get better with MetaClaw. Kudos to the team @richardxp888, @JimChenjw, @Xinyu2ML, @lillianwei423, @StephenQS0710, @HaoqinT, @JiaqiLiu835914, @yuyinzhou_cs, @ZhengBerkeley, @cihangxie

Day 3 building MetaClaw🦞 🚀 One-click deploy just dropped. No config; no setup. Just talk to your🦞 and watch it evolve. Try it 👉: github.com/aiming-lab/MetaCl… Kudos to the team, @richardxp888, Jianwen Chen, @Xinyu2ML, @lillianwei423, @StephenQS0710, @HaoqinT, Zeyu Zheng, @cihangxie, @HuaxiuYaoML

We’ve officially brought the power of SkillRL to the CLI agent world! 🔧 Built on @openclaw and powered by Tinker @thinkymachines, MetaClaw lets agents learn new skills from conversations and evolve live — no GPU clusters needed. Our original SkillRL outperformed GPT-4o using a 7B model. Now Kimi-2.5 can hot-swap weights and inject new skills on the fly. 🧠🦾 Try it here: github.com/aiming-lab/MetaCl…