If you're looking to boost your model’s math reasoning ability, don’t miss out on MegaMath —— The Largest Open Math Pre-training in the world! 🧠 Need from-scratch training? Use MegaMath. 🔁 Need continual pre-training? Use MegaMath. 🧬 Need high-quality mid-training? Use

The Sparse Frontier: Sparse Attention Trade-offs in Transformer LLMs New benchmark from Cohere compares training-free sparse attention across: - 7B → 72B models - 16K → 128K tokens - up to 95% sparsity Findings: - larger, sparser models outperform smaller dense ones on long

Command A, our state-of-the-art generative model, is now the highest-scoring generalist LLM on the Bird Bench leaderboard for SQL! It outperforms other systems that rely on extensive scaffolding to tackle these SQL benchmarks, and instead delivers these results out-of-the-box,

We are kicking off a series of seminars at HKUNLP. Siyan Zhao will be giving a talk titled "d1: Scaling Reasoning in Diffusion Large Language Models via Reinforcement Learning" at ⏰Friday 5.9 11am HKT (Thursday 5.8 8pm PDT). Link to talk: hku.zoom.us/j/97925412724?…

🚀 Big news! Our collab w/ Snowflake, UCSD & UMD topped the BIRD leaderboard — beating prior SOTA by 2.8% in Text-to-SQL reasoning! RL was tough, but worth it. 📢 Best model coming soon. #AI #LLM #TextToSQL #ReinforcementLearning #Snowflake #UCSD #UMD #NLP #BIRDLeaderboard

As we all know large reasoning models are extremely overthinking even when answering 1+1. Many approaches have been proposed to mitigate this issue, but it doesn't need to be that complex -- we (and several concurrent works) found that just continuing standard RL training while

🤔 Have we really made great progress on software engineering tasks? 🚀 Introducing SWE-bench-Live, a live-updatable benchmark for real-world bug fixing. 😺 Even the best combo, OpenHands + Claude 3.7 Sonnet, sees a major performance drop! 👉 swe-bench-live.github.io 🧵 1/4

🚀 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…

Qwen3-VL is finally released and open-sourced, available in both Thinking and Instruct versions! This time, we’ve placed special emphasis on strengthening Visual Agent and Visual Coding, which are crucial steps toward building a true Digital Agent 🚀

Yeah, we do make efforts in benchmarking and setup-ing text-to-sql env towards real DBA cycle step by step😃~

How do you make LLMs both long-context capable and super fast? Meet InfLLM-V2 from Tsinghua x OpenBMB — a breakthrough dense-sparse switchable attention system that: 1⃣ Seamlessly adapts from short to long sequences 2⃣Runs 4× faster than dense attention 3⃣Keeps 99%+ accuracy

Muon saves optimizer memory overhead from 3 x model_size to approximately 2 x model_size, can we further reduce it to 1 - 1.5 model size?

Excited to share a major update to BIRD-SQL! Our team with efforts from global engineers, experts, and students has completed comprehensive quality control, releasing bird-sql-dev-1106: huggingface.co/datasets/birds… For new submissions using this cleaner split, please indicate this

Tried out Gemini-3-Pro on our BIRD-SQL Verfied Dev + Hidden Test sets. It’s the first general model which can break 70 🚀, getting surprisingly close to specialized SFT/RL models from Databricks and Snowflake . Really impressive generalist performance 🔍✨ BTW, Gemini-SQL

i'll be talking about qwen3-coder at 1530 at hall g. feel free to join!

XiYan-SQL is an innovative natural language–to–SQL conversion framework designed to address the performance challenges large language models face in SQL generation tasks. XiYan-SQL just hit #1 across all open BIRD-CRITIC (SWE-SQL) leaderboards — including BIRD-CRITIC-1.0-Open,

🎮Get a first look at Tencent HY World 1.5 (WorldPlay)! 🎮 Our newest world model with real-time interaction and long-term memory. It’s going *open-source* tomorrow.

Most recent LLM+RL work focuses on clipping for stability, there's another path → better baselines! Check out our new method which introduces 𝗼𝗽𝘁𝗶𝗺𝗮𝗹 𝘁𝗼𝗸𝗲𝗻 𝗯𝗮𝘀𝗲𝗹𝗶𝗻𝗲 for more stable RL training 🔥

Excited to share that our work "BIRD-Interact" has been accepted to ICLR 2026! 🎉 In this project, we introduce a dynamic, interactive Text-to-SQL environment, testing whether LLM agents can handle ambiguous, evolving database tasks like a real DBA. 🤖 🚀 Here is our project

Happy CNY! We are glad to introduce our latest language model Seed-2.0. We make great progress (agent, reasoning, vision understanding, etc.) since Seed-1.8 without any distillation Right now it's only available in CN now, and will soon be ready globally. seed.bytedance.com/en/seed2