🚀 Excited to announce the launch of OpenRSI! Join us and witness how RSI progresses in real time. We're entering the second half now. Let’s see how far it can go. Join now:

It has been mentioned many times that there are two main road rocks in the path to AGI: recursive self-improvemet and continual learning, which are sharing the same spirit: AI should not only act like humans, but also teaches itself like human in the current fast-moving, time-sensitive, and complicated world. Towards a better understanding towards these problems, I am excited to share that we have published the first end-to-end benchmark of computer use, continual learning, and long-horizon agentic capabilities, set in a real job: In this benchmark, you can see how frontier AI acts and learns like humans as an apprentice in a real job and please expect the potential future of AI-native world! #astra# #fable5# #continuallearning#

Excited to share that I’ve taken leave from UCSD and joined NeoCognition as Director of Machine Learning. Since July, I’ve been working closely with a talent-packed team on a strategic bet. Today is the day that we release details on part of what we’ve been building. 🧵1/n

We see incredibly smart and shockingly stupid cases from modern agents in our daily research. But are they ready for complex, real-world knowledge work, and can they continually learn from it? #Fable# 5.1 and #Astra# beat human testers on @ApprenticeBench, which evaluates realistic and economically valuable accounting tasks.They can use computers, handle the long contexts needed for knowledge work, and already show real potential for continual learning. They are still slow and expensive, and continual learning is far from reliable. All other evaluated agents still fall short of best human workers in effectiveness and efficiency. With ApprenticeBench, we can differentiate among the latest agents and identify what still prevents them from doing real jobs. Check out our blog!

Fable 5.1 and Astra can now learn a knowledge job end to end. It’s the culmination of rapid and comprehensive capability advances across computer use, continual learning with memory notes, and long-horizon agentic behavior over the past year. Together, these advances lead to a step change in AI’s job readiness. It’s a bittersweet moment for me personally. I’ve been working on computer-use agents since 2022, back in the early days. I’ve always believed that the best way for agents to become truly helpful to us is for them to operate and learn the same way we do: use the same software through GUIs, and go on a lifelong learning journey on the job. That’s the path of least resistance for AI adoption across industries. And that’s what drove us to found @NeoCognition. What surprises me is how fast frontier AI is getting there. The rapid ramp in the CUA capability across generations of Claude and GPT models is a loud and clear sign. Fable 5.1 and Astra can now use GUIs just as well as APIs. Computer use is not a fundamental blocker anymore. We are also seeing early signs of the emergence of continual learning, even with a simple approach based on memory notes. But there’s still a lot of work to do. Researchers need to understand and improve how agents learn, in both effectiveness and efficiency. Builders need to turn those capabilities into systems people can trust. Organizations need to bring workers into the process and ensure they share in the gains. AI has crossed an important threshold in job readiness. It is inevitable that they will start doing real work across industries. But together, we have the opportunity to make that a benevolent future that benefits everyone. It is a future worth building.

A machine worker is built to repeat a fixed process. A human worker learns from experience and grows into harder tasks. That's why humans can't be easily replaced. I was firmly in the "machine worker" camp for the current agents, until I saw the impressive performance of #Astra# and #Fable51#... How did they change my mind, and how do we test an agent's ability to keep learning? See ApprenticeBench⬇️

Zero Train–Inference Mismatch — now for linear attention, and under async RL 🎯 We got bitwise-exact trainer/generator parity for Gated DeltaNet (Qwen3.5-9B / 35B-A3B) on TorchTitan RL + vLLM, then asked the question nobody had actually tested in open source: does it help async RL? Two firsts (as far as we know): 1️⃣ First open-source stack to hit train/inference logprob diff = 0 on a linear-attention model 2️⃣ First to measure what zero mismatch is worth under async off-policy RL How — smaller than it sounds: • trainer and generator share one model definition (TorchTitan unified model). No more aligning trainer & generator op by op • recurrent kernel everywhere on the forward (prefill + decode + trainer), chunked only for backward • split-K off, batch-invariant GEMM, + a batch-invariant bmm for the MoE router • we never touch FLA kernel internals — the recurrent kernel is intrinsically batch-invariant: one sequence's own state, fixed order, no cross-sequence reductions • vLLM's linear-cache management + prefix caching keep working as-is 📊 Qwen3.5-9B + DAPO-Math, async at offpolicy = 4 / 12 / 32: • logprob_diff = exactly 0 at step 0 ✅ every nonzero after that is staleness, not precision • at offpolicy=32 the stock stack runs away past 0.065; ours holds ~0.035 • the unified model is free — the entire bill is the kernels • also verified on MoE 35B-A3B and a 64-turn / 64K-context terminal agent 🤔 Now the honest part: it depends on the config. On math at offpolicy=12, BI leads on train reward and peaks highest on held-out AIME. On the terminal agent it's slightly ahead. On search, and on other math configs, the curves are indistinguishable. What is consistent is the numerics — BI keeps the gap from exploding as you widen the window. Cost: 2–3× trainer throughput (≈5× on the agent). Our take: bitwise parity is a debugging tool, not a production default. Flip BI on for 20 on-policy steps — if the gap is 0 and your run is still broken, the infra is exonerated and the bug is in your data or your algorithm. That's worth 20 steps, not a whole run. Next: • make BI cheap (kill the 2–3×) • more workloads, more seeds • stay tuned for more complex agentic training recipes/tests 👀 Full writeup 🔗 We collaborated closely with the TorchTitan team, and we sincerely encourage everyone to check out TorchTitan RL. With its clean, efficient unified model architecture and native async RL design, we believe it is a strong foundation for your next RL engine. #LLM# #RL# #agentic# #pytorch# #torchtitan#

why can the china labs build glm-5.2, kimi k3, and many more to come? it is because of the openness. not just the open weights but the whole ecosystem. most of the work done in the china labs is carried by interns. i met brilliant undergrad and graduate interns who deeply understand the model training details, and they are 100x more open to share. that means the talent that knows how to train llms in china is 100x greater in number than the talent in the us, and it is growing in contrast, the us ai ecosystem is too closed. frontier labs do not hire interns. i know brilliant phd students at stanford, berkeley, and so on. they struggle to get an internship and the compute to train a properly sized model. most of the secret recipes are locked away by a very small group of privileged researchers it is not about china or the us. it is about open and closed science. the fact is that every average cs student can learn how to train an llm. they just need the opportunity. labs should be more open and hire more interns, like how deepmind and fair did in the pre-llm era

🚀 Introducing our latest research: AgentCL: Toward Rigorous Evaluation of Continual Learning in Language Agents Continual learning for language agents has not yet been clearly defined. How should we evaluate their ability to continually learn from experience and improve themselves on complex, long-horizon tasks? - Traditional continual learning provides a useful perspective on the plasticity–stability trade-off, but its formulation does not naturally extend to non-parametric learning paradigms today. - Many recent works on agent memory still focus on retrieval and reasoning over long (static) contexts, rather than on how agents can reuse experience from complex agentic tasks. Across coding, deep research, and language understanding and reasoning tasks, we show that carefully designed task streams and metrics are essential for understanding continual learning in language agents. 📄 Paper: 🤗 Dataset:

🚀 Impressive work. Training a strong Deep Research agent with just 32 H100s, 8K synthetic samples, and a fully open recipe highlights how fast open ecosystems are advancing. Great to see OpenResearcher ( included as a baseline. Exciting progress for open Deep Research. 👏

🚀 Introducing DR-DCI: Scaling Direct Corpus Interaction via Dynamic Workspace Expansion Search agents face a basic tradeoff: Retrieval scales to massive corpora, but often exposes evidence only as ranked snippets or bounded document views. Direct Corpus Interaction (DCI) gives agents precise terminal-style control with tools like rg, grep, find, and read, but raw full-corpus search becomes slow and unstable as the corpus grows. DR-DCI bridges the two. The agent dynamically calls pull(query, topK) to expand a local workspace, then performs DCI inside that bounded workspace for evidence search, comparison, and verification. Retriever = corpus-level discovery DCI = workspace-level investigation Result: on BrowseComp-Plus, DR-DCI improves Raw DCI from 62.9% to 71.2%, while reducing wall time and estimated cost. With workspace-preserving context reset, it reaches 73.25%. HF paper: Code: More details below 👇

I'd also like to share a few thoughts from the past six months of building QUEST that didn’t make it into the paper, but might still be useful to the community. > Mid-training is surprisingly powerful, especially for smaller models. In our early experiments, we applied mid-training to smaller models like Qwen3-8B and saw surprisingly large gains (~3pp over pure SFT) with only around 100K tasks. The effect became weaker as model size increased, but for smaller Deep Research agents, carefully designed mid-training tasks can still go a long way. How should those tasks be designed? That’s a much longer story. If you’re interested, check out the “Unsuccessful Attempts” section in our paper. Honestly, it’s probably my favorite section in the whole paper. > For agentic RL, infrastructure matters way more. This might sound like a cliche in 2026, but I still want to yell it loudly. We spent almost two months just making our RL pipeline stable. Agentic RL is essentially a giant chain of dependencies. The judge model needs to work. The search and retrieval services need to work. The cache needs to work. And all of them need to keep working for days. A tiny bug in any part of the pipeline can ruin an entire training run. We end up spending a surprising amount of time designing fallback mechanisms for everything. Not because it’s elegant, but because otherwise your training will randomly die at 3 AM after running for days. The secret isn’t fancy RL tricks. The secret is making sure your pipeline survives when things inevitably break. > Session-level training is probably the reason we can afford long-horizon training. Everyone working on Deep Research agents knows the pain: context is expensive. Compared to traditional reasoning tasks or short-horizon agents, long-horizon Deep Research training burns through GPU resources much faster. For most academic groups, GPUs are still the biggest bottleneck. In QUEST, we use session-level training and limit each training instance to 32K tokens. It’s not the most glamorous idea in the paper, but it helps make large-scale long-horizon training much more practical. > Cache saved us a ridiculous amount in API spend. We spent a lot of effort building our caching system. At first, it sounded like an engineering optimization. Later, it became a necessity. Many websites visited during data synthesis show up again during RL training. So as the search tool. Without caching, you’re effectively paying for the same information over and over again. The funny part is that cache becomes even more valuable when your experiments fail. RL runs crash. Services fail. You restart things. But cached results stay. One funny observation: the more failed runs you have, the higher your cache hit rate gets. It’s a slightly sad but comforting story: every failed run leaves behind a few more cache entries for the next one. By our final training run, the cache hit rate had reached ~40%, which translated into a significant reduction in API costs. How painful it was to build. How rewarding it was to finish. We hope you’re as excited to meet QUEST as we are to finally share it.

Excited to share our lab’s latest work on fully open deep research agents, led by @jianxie_! For anyone doing agentic post-training, it’s a gold mine: a full recipe (mid-training → SFT → RL), a rubric-tree data synthesis pipeline, and an open model family from 2B to 35B.

🚀Excited to announce QUEST today! Benchmark numbers will fade, but the know-how behind them endures. With QUEST: See how we synthesize diverse Deep Research tasks through a unified framework at scale. See how we train Deep Research agents through three stages: Mid-Training → SFT → RL. See how our caching system dramatically reduces API costs during training. See how our context management enables unbounded deep research. We’ve released everything we can!

🤖 Agents increasingly rely on skills to handle complex tasks with privileged trust, which makes a poisoned skill a dangerous attack surface. Worse, skills persist and get reused: one can look benign today, silently mutate itself, and attack tomorrow ⚠️ Introducing SkillHarm: Lifecycle-Aware Skill-Based Attacks via Automated Construction • 2 attack scenarios across the skill-use lifecycle: Fixed-Payload Poisoning & Self-Mutating Poisoning • 12 risks organized by the workflow component harmed: Data Pipeline, System Environment & Agent Autonomy Exploitation • AutoSkillHarm: automatic attack construction with coding agents driven by natural-language harnesses 🚨 Frontier coding agents stay vulnerable with ASR up to 86.3% and current defenses don't hold. 🔥 Already 3.7K+ downloads on Hugging Face in the first week! 🧵: