Zonghan Yang

💪🦾Agentless Training as Skill Prior for SWE-Agents We recently released the technical report for Kimi-Dev. Here is the story we’d love to share behind it: (1/7)

Millions of users now have months-long conversation histories with AI assistants💬 But this data is proprietary and unavailable to the academic community for research, training, or benchmarking. We introduce HorizonBench🌅, a benchmark and data generator for long-horizon personalization: tracking a user's current preferences across a history where life events have silently changed them.

Grateful for my experiences in labs at two top Chinese tech firms, and special thanks to @JustinLin610 for recommending me to the early Qwen team. These shaped my career profoundly, both technically and beyond. In the fast-evolving field of LLMs/Agents/Claws, org challenges are complex yet simple. Complexity arises from balancing intricate goals under constraints (GPU, time, infra, inference costs), rapid iteration, and scaling teams amid limited resources. But ultimately, it's simple: Trust, inclusion, respect, and empathy are the most effective—and timeless—solutions.

AI agents are tackling more and more "human work" But are they benchmarked on the work people actually do? tl;dr: Not really Most benchmarks focus on math & coding, while most human labor and capital lie elsewhere. 📒 We built a database linking agent benchmarks & real-world work Submit new tasks + agent trajectories today 🧵

Dimitris has been demonstrating the new way to do AI research in this agent era: find a neat problem, reason about why it’s interesting and tractable, offload the execution work to agents, analyze the results and write up a fun post about it. It’s still important for us human researchers to have the expertise to be able to identify the problem and judge the findings (eg, the background knowledge on matrix completion and SVD, and making the connection to LLM benchmarking); but even just automating the execution alone is already a massive acceleration and I think as a community we should really embrace this new form of AI-assisted research.

Cursor now shows you demos, not diffs. Agents can use the software they build and send you videos of their work.

Many people are using RL to make models smarter. We used RL to pull training data out of the models themselves. Our results show that models know a lot more about their training data than most people think. We develop Active Data Reconstruction Attack (ADRA) — a data detection method that uses RL to induce models to reconstruct data seen during training. ADRA beats existing methods by an average of >10% across pre-training, post-training, and distillation. Our paper, with @uwnlp, @Cornell, and @BerkeleyNLP @Berkeleyai, is now available. Arxiv: arxiv.org/pdf/2602.19020 Joint work with @jxmnop @shmatikov @sewon__min @HannaHajishirzi

🔭 We’re releasing Hodoscope: an open-source tool for unsupervised behavior discovery. It lets you visually explore and compare agent behaviors at scale. It helped us discover a novel reward hacking vulnerability in Commit0 - with just a couple minutes of human effort.

Agent Skills are everywhere - Claude Code, Gemini CLI, Codex all support them. But do they actually work? 105 domain experts from Stanford, CMU, Berkeley, Oxford, Amazon, ByteDance & more built SkillsBench to find that out. 86 tasks. 11 domains. 7,308 trajectories. 🧵👇

One prompt = A publication-ready research paper with academic charts. If you need pro academic formatting, latex formulas and figures, Kimi is all you need. You can get a word file just by chatting.

We've been working on very long-running coding agents. In a recent week-long run, our system peaked at over 1,000 commits per hour across hundreds of agents. We're sharing our findings and an early research preview inside Cursor.

A lot of people quote tweeted this as 1 year anniversary of vibe coding. Some retrospective - I've had a Twitter account for 17 years now (omg) and I still can't predict my tweet engagement basically at all. This was a shower of thoughts throwaway tweet that I just fired off without thinking but somehow it minted a fitting name at the right moment for something that a lot of people were feeling at the same time, so here we are: vibe coding is now mentioned on my Wikipedia as a major memetic "contribution" and even its article is longer. lol The one thing I'd add is that at the time, LLM capability was low enough that you'd mostly use vibe coding for fun throwaway projects, demos and explorations. It was good fun and it almost worked. Today (1 year later), programming via LLM agents is increasingly becoming a default workflow for professionals, except with more oversight and scrutiny. The goal is to claim the leverage from the use of agents but without any compromise on the quality of the software. Many people have tried to come up with a better name for this to differentiate it from vibe coding, personally my current favorite "agentic engineering": - "agentic" because the new default is that you are not writing the code directly 99% of the time, you are orchestrating agents who do and acting as oversight. - "engineering" to emphasize that there is an art & science and expertise to it. It's something you can learn and become better at, with its own depth of a different kind. In 2026, we're likely to see continued improvements on both the model layer and the new agent layer. I feel excited about the product of the two and another year of progress.

New Engineering blog: We tasked Opus 4.6 using agent teams to build a C compiler. Then we (mostly) walked away. Two weeks later, it worked on the Linux kernel. Here's what it taught us about the future of autonomous software development. Read more: anthropic.com/engineering/bu…

Congrats to @stefan_fee on this excellent work! Really impressed by the daVinci-Agency paper. It shows how chain-of-PRs (global PR context) substantially boosts SWE performance—a natural yet underexplored signal. Even more surprising: agentic trajectories from these PRs yield big gains on Toolathlon (toolathlon.xyz/introduction), a tough real-world tool-use benchmark beyond SWE. Definitely worth diving deeper!

Proud to support @stanfordnlp CS224N this year. 🌲 Students are building their final projects using the Kimi K2.5 API. We can’t wait to see what this next generation of NLP researchers creates by the March 16 poster session. web.stanford.edu/class/cs224n/ Enjoy building with Kimi 💫

Excited to share our latest work: "Semi-Autonomous Mathematics Discovery with Gemini." We used Gemini to systematically evaluate 700 "open" conjectures in the Erdős Problems database. The result? We addressed 13 problems marked as open—finding 5 novel autonomous solutions and identifying 8 existing solutions missed by previous literature. Read the full case study here: arxiv.org/abs/2601.22401

RL Anything! Your environment, reward model and policy can be improved in a closed-loop optimization. They provide feedback for each other to enhance the training signals and benefit the whole system. Check this out.

STAT has been accepted to ICLR 2026! See you in Brazil 🇧🇷 Skill-Targeted Adaptive Training (STAT) is a continual learning method that squeezes out 🚨 7~10% more performance on extensively trained models like Qwen. It constructs a 🧩 Missing-Skill-Profile for each model based on what skills the model lacks in their responses, and adaptively curates post-training data accordingly. Check out our Blog Post 👉 ying-hui-he.github.io/Skill-Targeted… 🔗arXiv : arxiv.org/abs/2510.10023 💻GitHub: github.com/princeton-pli/…

People keep saying 2026 will be the year of continual learning. But there are still major technical challenges to making it a reality. Today we take the next step towards that goal — a new on-policy learning algorithm, suitable for continual learning! (1/n)

Training LLMs with verifiable rewards uses 1bit signal per generated response. This hides why the model failed. Today, we introduce a simple algorithm that enables the model to learn from any rich feedback! And then turns it into dense supervision. (1/n)