@JustJerry121 I didn't "harness" it. Just let the model do this.
English
Yaowei Zheng
68 posts
Yaowei Zheng
@hiyouga_dev
Founder @llamafactory_ai, CTO @prismshadow_ai Building platform for next-generation AI. Opinions are my own.
Millennium Science School Katılım Temmuz 2020
86 Takip Edilen40 Takipçiler
@hiyouga_dev Do you usually keep the CSS inline/tiny, or is plain semantic HTML already enough for most outputs?
English
A small product thought from today:
LLMs naturally default to Markdown, but it is not always the easiest format to read.
Ask for a minimal one-page HTML instead, and with only 1.1~1.2x token cost, the output feels much cleaner.
English
Kami is a great project built around the same idea:
github.com/tw93/kami
English
@DavidFSWD @teortaxesTex true, openrouter is really not reliable in production use
x.com/hiyouga_dev/st…
Yaowei Zheng@hiyouga_dev
Some providers on OpenRouter cannot even decode tool calls correctly. I ran into this issue while testing the Qwen3.6-35B-A3B on AgentHub, so I had to block these providers ☹️ The services we commonly use are a huge black box. I truly trust nothing but the official provider.
English
@teortaxesTex don't use openrouter for anything production use the actual original company, like deepseek. they cram scamming LLM providers that use really bad quants.
English
Man that's rough
what's going on at Openrouter? 87% is nothing, direct DeepSeek connection is reliably >95% hit rate
Geek@geekbb
还得是 DeepSeek，以后穷哥们就用 SA 吧。
English
@ericjang11 @tinkerapi agreed. post training is an essential way to build product moat
English
.@tinkerapi by TML is a highly underrated product. It should be part of the “starter post training” infra for any newco
English
We can basically think of an agent as a GitHub repo.
The LLM is an executable inside it. The context is a set of text files. The harness is the code that manages how context and the LLM interact.
Apply harness engineering to this repo, and you get a self-improving agent.
English
I built AgentHub, and I truly believe every AI agent developer needs this tool. github.com/Prism-Shadow/a…
Paul Graham@paulg
A startup idea that only works if there are already a significant number of people using it is not a valid startup idea. There has to be some subset of users who need what you're making so desperately that they'll use it even if no one else is.
English
Yaowei Zheng retweetledi
This is the new standard for engineering evals
Serena Ge (Datacurve)@serenaa_ge
Today we’re releasing DeepSWE, a new standard for agentic coding benchmarks. On public leaderboards, top models often look relatively close in capability. DeepSWE shows where they actually diverge, reflecting the realistic experience of developers in their day-to-day work.
English
Agent = LLM + Claude Code + Git Repo
Git Repo is all you need to build self-evolving agents
Shannon Shen@shannonzshen
"Evolve your repo, not just your agent." Self-evolving repository (SEPO) is a fun idea that I’ve been exploring recently. @sepoagent turns any GitHub repo into a shared workspace for humans and coding agents.
English
Some providers on OpenRouter cannot even decode tool calls correctly. I ran into this issue while testing the Qwen3.6-35B-A3B on AgentHub, so I had to block these providers ☹️
The services we commonly use are a huge black box. I truly trust nothing but the official provider.
English
Yaowei Zheng retweetledi
RLVR has become the recipe for agentic post-training. But for Computer-Use Agents, the bottleneck is not the algorithm, it is the data. 🐌
🚀 We introduce CUA-Gym: a scalable, lightweight synthesis engine that turns arbitrary task queries into verifiable RLVR data for computer-use agents. The largest open CUA RLVR dataset to date:
🎯 32,122 verifiable RLVR tasks with programmatic setup scripts + rewards
🌐 110 environments: 16 desktop apps + 94 synthesized mock web apps
🏆 Qwen3.5-based CUA models trained with GSPO reach 72.6% on OSWorld-Verified and 56.6% on WebArena
📄 Paper: huggingface.co/papers/2605.25…
🏠 Homepage: cua-gym.xlang.ai
🤗 Dataset: huggingface.co/datasets/xlang…
💻 Codebase: github.com/xlang-ai/CUA-G…
🧩 Environments: github.com/xlang-ai/CUA-G…
🧵[1/6]
English
Yaowei Zheng retweetledi
Introducing CHI-Bench on @huggingface: the world’s first long-horizon healthcare benchmark for AI agents.
75 real healthcare workflows + 20 apps + 200+ MCP tools + 1,290 skills + process / outcome rewards
huggingface.co/datasets/actav…
Any questions, lmk!
English
When building a project, choosing the right technology stack is very important. I suggest that you first clarify the product requirements, and then repeatedly communicate with AI about the choice of the technology stack. If the technology stack is wrong, the subsequent workload will be unimaginably amplified.
English
@JustJerry121 If you are building an AI app for a specialized use case, you can use statistics from other users to improve the initial context.
English
Biggest lesson from building an AI startup:
If you're building a full agent system, don't build a web app.
Local is the best interface.
That's why Claude Code and Codex chose local.
Web only makes sense when you're building one narrow, specialized feature.
English
Whale maxxing + deep drink 😆🍻
What a night at the Blue Whale event.
English
Yaowei Zheng retweetledi
🎉 Congrats to the VeRL-Omni team on the pre-release of a general RL post-training framework for multimodal generative models. Built on verl + vllm-omni.
vLLM-Omni handles the multimodal rollout with step-wise continuous batching and embedding caching; vLLM serves the VLM-as-judge / OCR reward model, overlapped with rollout and training. In the Qwen-Image OCR demo, moving the reward to its own GPU cuts per-step wall-clock by ~14%.
Released: Qwen-Image with FlowGRPO / MixGRPO / GRPO-Guard. BAGEL and Qwen3-Omni-Thinker PR-ready.
Excited to push multimodal generative RL forward together with VeRL-Omni and the broader community. 🙌
📖 vllm.ai/blog/2026-05-1…
🔗 github.com/verl-project/v…
English
When we say "meta harness" or "agentic harness engineering", we may miss what "harness" means.
It is not just code assets for building agents.
Harness is the top-level concept: leveraging intelligence inside a Ralph Loop until the human is satisfied.
ghuntley.com/loop/
English
Heuristic Learning is an RL loop over code assets.
trinkle23897.github.io/learning-beyon…
OpenAI's Harness Engineering is exactly the same idea in broader practice.
openai.com/index/harness-…
English