Countless times of iterations for cooking it, but the process is satisfying. I still believe we can pour more data in each stage if we have more hands so the potential is unlimited and scaling law hasn’t hit the wall yet! Towards Digital Agents🤖 We are already on the way.

Wrapped up a SWE-Perf website redesign using Qwen3-Coder on AnyCoder (huggingface.co/spaces/akhaliq…). The process was incredibly fast and great! One question for Qwen devs, though: did you pretrain a secret love for the color purple into the coder's persona? 😉

Proud to introduce Group Sequence Policy Optimization (GSPO), our stable, efficient, and performant RL algorithm that powers the large-scale RL training of the latest Qwen3 models (Instruct, Coder, Thinking) 🚀 📄 huggingface.co/papers/2507.18…

Failing on 𝐥𝐚𝐫𝐠𝐞-𝐬𝐜𝐚𝐥𝐞 𝐑𝐋 with VeRL? ⚠️ Mixing inference backend (𝐯𝐋𝐋𝐌/𝐒𝐆𝐋𝐚𝐧𝐠) with training backends (𝐅𝐒𝐃𝐏/𝐌𝐞𝐠𝐚𝐭𝐫𝐨𝐧) 𝐬𝐞𝐜𝐫𝐞𝐭𝐥𝐲 𝐭𝐮𝐫𝐧𝐬 𝐲𝐨𝐮𝐫 𝐑𝐋 𝐢𝐧𝐭𝐨 𝐨𝐟𝐟-𝐩𝐨𝐥𝐢𝐜𝐲 — even if they share the same weights! 📉 Blog:

Token crisis: solved. ✅ We pre-trained diffusion language models (DLMs) vs. autoregressive (AR) models from scratch — up to 8B params, 480B tokens, 480 epochs. Findings: > DLMs beat AR when tokens are limited, with >3× data potential. > A 1B DLM trained on just 1B tokens

🚀 OSWorld gets a major upgrade! OSWorld-Verified: 15 months community feedback → 300+ fixes (ambiguity, graders…), 50x faster eval through AWS parallelization More apple-to-apple comparison for reliable CUA evaluation ✨ 👇xlang.ai/blog/osworld-v…

⚡𝐅𝐏𝟖 makes RL faster — but at the cost of performance. We present 𝐅𝐥𝐚𝐬𝐡𝐑𝐋, the first 𝐨𝐩𝐞𝐧–𝐬𝐨𝐮𝐫𝐜𝐞 & 𝐰𝐨𝐫𝐤𝐢𝐧𝐠 𝐑𝐋 𝐫𝐞𝐜𝐢𝐩𝐞 that applies 𝐈𝐍𝐓𝟖/𝐅𝐏𝟖 for rollout 𝐰𝐢𝐭𝐡𝐨𝐮𝐭 𝐥𝐨𝐬𝐢𝐧𝐠 𝐩𝐞𝐫𝐟𝐨𝐫𝐦𝐚𝐧𝐜𝐞 compared to 𝐁𝐅𝟏𝟔! 📝 Blog:

Yes, brutally true. I tend to see LLM RL ≈ on-policy self-distilled SFT with reward re-weighting. The key differences between LLM SFT (rejection sampling) and RL are: 1. Negative examples, or more precisely, advantage-weighted samples 2. On-policyness: Even iterated SFT is more

[0/3] 🚀 Introducing Verlog – an open-source RL framework built specifically for training long-horizon, multi-turn LLM agents. 📊 Max episode length comparison: •VeRL / RAGEN → ~10 turns •verl-agent → ~50 turns •Verlog (ours) → 400+ turns 🔥 ⚙️ Technical foundation:

Where are our computer‑use agents (CUA) standing on OSWorld‑Verified? Potentially already ~80%. We made this analysis, which summarizes the latest OSWorld-Verified submissions with 27 models evaluated over 369 tasks, and conducted a case study on the o3+Jedi-7B approach to

We are super excited to release OpenCUA — the first from 0 to 1 computer-use agent foundation model framework and open-source SOTA model OpenCUA-32B, matching top proprietary models on OSWorld-Verified, with full infrastructure and data. 🔗 [Paper] arxiv.org/abs/2508.09123 📌

🔮 Introducing Prophet Arena — the AI benchmark for general predictive intelligence. That is, can AI truly predict the future by connecting today’s dots? 👉 What makes it special? - It can’t be hacked. Most benchmarks saturate over time, but here models face live, unseen

Try Mirage 2 now → dynamicslab.ai Here are a few worlds we’ve created—starting from images and prompts. More in the thread 👇 2/

🔥 Super excited to launch Mirage 2 A big leap toward a general-purpose world engine for live interactive play 🎮 Hard to believe how far we've come in just one month since Mirage 1 ⏩ If you’re impressed by Genie 3, come play with Mirage 2 — It’s live, offering an extended

Excited to share my 1st project as a Research Scientist Intern at Meta FAIR! Grateful to my mentor Jiawei Zhao for guidance, and to Yuandong Tian & Xuewei for their valuable advice and collaboration. Our work DeepConf explores local confidence for more accurate & efficient LLM reasoning!

We released DeepConf that can achieve 99.9% on AIME'25 with open source models with only 15% of the compute, compared to majority voting@512. The secret? Simple. Just to pruning the rollouts if they show a consecutive stream of low-confidence😀. Can be applied to any models

We are glad that TIS and FlashRL have received broad attention from the open-source community that they have been verified and supported (OpenRLHF Jian Hu, SkyRL NovaSky, REINFORCE++Jian Hu, OAT Zichen Liu)! A few updates on our blog and FlashRL package: (1) more in-depth

Router wasn't learning at first, we debugged it step-by-step and showed you how despite perfect load balancing, routing can be completely useless. We root caused it and fixed the problem. Papers skip the methodology, but you can find all details in our part 3 of MoE 101 series

One of the reasons academic science is so bad at producing or even accepting novelty is because the focus on hypothesis driven science has caused people to view exploratory studies as inherently unrigorous.