Ethan Chern

"Failure is just iteration. No explosion, no innovation. Keep going."🚀 You vent to @elonmusk—he looks you in the eye and replies instantly, like a video call. Introducing LiveTalk: real-time video gen system on a GPU that sees you, reads emotion, and responds in real time.🧵👇

@fudayuan @whistom25 @Eudaemonia279 @stefan_fee 👋 Introducing daVinci-Env: the largest fully transparent framework for SWE environment synthesis in Python at scale. We open-source 45,320 environments, and 32B/72B models trained on them reach 62.4%/66.0% on SWE-Bench Verified.

📍 Can LLMs discover, abstract, and reuse higher-level tool skills across tasks? Existing tool-use benchmarks test solving tasks with fixed tools. But real workflows contain recurring structures where efficiency comes from reusable tool compositions, not isolated calls. We introduce SkillCraft: 126 tasks across 6 domains designed to test whether LLM agents can acquire compositional skills, not just call atomic tools. We also propose Skill Mode, a lightweight protocol with four MCP primitives that let agents compose, verify, cache, and reuse tool chains at test time. Our Key findings across evaluating 8 SOTA models: ⚡Skill Mode enables agents to self-discover and reuse skills, leading to higher success and efficiency than agents without it. The gains are larger for stronger models. 🧠 Stronger models (e.g., Claude) discover more generalizable skills, which transfer across tasks and even across models. 🔍 Deeper composition ≠ better — shallow, well-tested skills generalize best. 🔗 Paper: arxiv.org/abs/2603.00718 💻 Code: github.com/shiqichen17/Sk… 🏠 Page: skillcraft-website.github.io/page (1/7)

很喜欢 SII - GAIR 最近一系列的围绕 Agent 的工作。 继 davinci 系列后，看到最新的工作，关于 数据 evolve 的 data Darwinism 来自 awsome @QinYi88814 @stefan_fee

Should data "evolve"? 🧬 Scaling is not enough. Model performance is bounded by data, but its value is defined by processing depth. We introduce Data Darwinism, a 10-level hierarchy (L0-L9) redefining data as an eternal co-evolutionary process.(1/n) huggingface.co/papers/2602.07…

本周最喜欢的三篇 Coding Agent 文章，da Vinci-Agency, da Vinci-Dev, ProjDevBench。

[1/8] 🐶 Introducing "AgentDoG": A Diagnostic Guardrail Framework for AI Agent Safety. It achieves SOTA performance, diagnosing root causes (e.g., prompt injection, tool misuse) with 82% accuracy, far surpassing general LLMs. 📄 Paper: arxiv.org/abs/2601.18491

🎨 AI agents are excellent "Sprinters" solving single functions in seconds, but they fail at "Marathons" like long-horizon tasks. They lose context, drift, or give up. Why? They lack endurance training. We introduce daVinci-Agency: The FIRST automatic data synthesis pipeline to achieve project evolution level agency! With just 239 samples, we beat baselines trained on 66k samples. 🚀 Paper: huggingface.co/papers/2602.02…

Can we instill foundational agent behaviors before SFT/RL? Introducing daVinci-Dev: First systematic study of Agentic Mid-Training for SWE. 🎨🤖 We lift Qwen-2.5 to the level of Qwen3, and we release billions of tokens! 📄 github.com/GAIR-NLP/daVin… 🤗 huggingface.co/collections/GA…

「 Benchmark, AgencyBench 」 很喜欢这个 benchmark，衡量 Agent 在 long horizon 长时度，超长上下文，复杂性高的任务表现。 文章分析不同模型做任务时的偏好非常有趣： Claude 和 GPT 喜欢用 shell command Gemini 喜欢用 memory tools Qwen 喜欢用 file operations Grok 和 GLM 喜欢用 web search

The boundary of evaluation determines the upper limit of intelligence. 🚀 NEW PAPER: "AgencyBench: Benchmarking the Frontiers of Autonomous Agents in 1M-Token Real-World Contexts"

实时的多模态交互！非常有前景的工作，来自 @ethanchern ！

LiveTalk Real-time multimodal interactive video diffusion system that achieves 20× speedup through improved on-policy distillation. Generates avatar videos from text, image, and audio inputs with sub-second latency.

🚀 LiveTalk 1.0 is NOW OPEN SOURCE! Pushing multimodal (video, audio) generation from “offline rendering” to “real-time interaction,” enabling AI to truly “show and tell” 🎯 Core Innovation: • Diffusion handles rendering • Next-token-prediction handles cognitive agency • On-policy distillation + system optimization → real-time interaction * Open-sourced • 20× speedup • 0.33s first-frame latency 📄 Paper: arxiv.org/pdf/2512.23576 💻 Code: github.com/GAIR-NLP/LiveT… 🤗 Model: huggingface.co/GAIR/LiveTalk-… Cognitively Agentic World Model: Bringing cognitive companions from the physical world into the digital realm 🌟

LiveTalk: A real-time multimodal avatar that streams video at ~25 FPS; - OmniAvatar-1.3B + Qwen3-Omni. github.com/GAIR-NLP/LiveT…

We’re pushing Video Gen beyond offline rendering into the era of Real-Time Interaction! 🎬📷 If you like what we’re building, please give us a ⭐ on GitHub and an ↑ on Hugging Face! It means a lot to the team. 🙏 x.com/ethanchern/sta…

LiveTalk is ready for the real world, as we’re moving from "generating video clips" to "building true relationships"!! Try out our model as we keep improving it! Great collaboration with Zhulin Hu, @Aphelios_Tang , @jiadisu7 , @steffichern , @SJTUDengLab, and @stefan_fee!!

Chat with your favorite idol through LiveTalk!!

"Failure is just iteration. No explosion, no innovation. Keep going."🚀 You vent to @elonmusk—he looks you in the eye and replies instantly, like a video call. Introducing LiveTalk: real-time video gen system on a GPU that sees you, reads emotion, and responds in real time.🧵👇