Biqing Qi

Found an interesting next model architecture exploration work from Shanghai AI Lab: SDAR, a new paradigm that converts trained AR models into blockwise diffusion models for FAST parallel decoding! ✅ AR's training efficiency ✅ Diffusion's inference speed The 30B MoE model even beats pure AR baselines on GPQA and ChemBench. HF Papers: huggingface.co/papers/2510.06… Model（1.7B/4B/8B/30B-A3B）：huggingface.co/collections/Je…

7/ More info & models: 🔗 Tech page: jetastra.github.io/SDAR/ 🔗 GitHub: github.com/JetAstra/SDAR 🔗 HuggingFace: huggingface.co/collections/Je… Technical report coming soon 📄

💥 Token-level AR is hitting its limits. We introduce SDAR — AR for pretraining, then switch to Diffusion+AR for SFT. ✅ 2×+ faster inference ✅ Equal/better accuracy ✅ Works for general & reasoning tasks 🔗 jetastra.github.io/SDAR/

TTRL Test-Time Reinforcement Learning

1/4 🔥 Just released: MARTI — a new open-source framework for Multi-Agent LLM Systems Reinforced Training and Inference! 📎 GitHub: github.com/TsinghuaC3I/MA… #AI #LLM #ReinforcementLearning #MultiAgent #OpenSource #AGI #TTRL

🧠 Again! Introducing: MARTI — Multi-Agent Reinforced Training and Inference A unified framework for LLM-based Multi-Agent Systems with centralized interaction & distributed policy training. Supports structured workflows (debate, MoA, chain), custom rewards, and 3rd-party MAS (e.g., AutoGen, CAMEL). 📈 Preliminary Highlights: Multi-agent RL > single-agent RL baselines under same inference budget MARTI-trained Qwen2.5-3B > standard RL & rivals instruct variants Strong results on AIME (66.7 score with TTRL + MAD, DeepScaleR-1.5B) 🧪 PPO | GRPO | REINFORCE++ | TTRL 🚀 vLLM V1 & Hybrid Engine compatible 🛠️ github.com/TsinghuaC3I/MA… 🤝 Collabs welcome! #LLM #MultiAgent #ReinforcementLearning #RLHF #AIResearch #OpenSource #MARTI

🎉Introducing our latest work: "ScienceBoard: Evaluating Multimodal Autonomous Agents in Realistic Scientific Workflows" 🤗 Huggingface: huggingface.co/papers/2505.19… 🏠Homepage: qiushisun.github.io/ScienceBoard-H… TLDR: We introduce ScienceBoard, featuring (1) a dynamic OS env with real scientific software (CLI + GUI), and (2) a human-validated benchmark spanning domains like biochem, astronomy, GIS, ATP, and more. 🧵[1/5]

Can 1B LLM Surpass 405B LLM? Rethinking Compute-Optimal Test-Time Scaling

@_akhaliq thank you for sharing our paper🎉

UltraMedical has been accepted as a spotlight at #NeurIPS2024 D&B Track! 🎉 Dive into our datasets, models and paper github.com/TsinghuaC3I/Ul…

🚀 Exciting news! Our paper on the adversarial robustness of SSMs has been accepted at NeurIPS 2024! 🎉 We have discovered some fascinating insights! Our paper: arxiv.org/pdf/2406.05532 and code is available at github.com/Biqing-Qi/Expl….

Thrilled to share our paper accepted by @COLM_conf and ranked the top 5% of all submissions! We firmly believe that LLM-generated hypotheses hold great promise for unlocking new paths to scientific discovery. Please checkout: arxiv.org/pdf/2407.08940 github.com/TsinghuaC3I/LL…

🚀 Exciting times in AI! As OpenAI unveils five new stages of AI development with a focus on autonomous discovery and collective intelligence, our latest paper echoes similar thoughts. 🧠💡 Read more here: hf.co/papers/2407.08… #AI #AGI #OpenAI #SpecializedGeneralist

🚀 Exciting news! Our paper has been accepted at CVPR 2024! 🎉 Dive into the future of machine learning with our groundbreaking work on Interactive Continual Learning based on system1 and system2 framework arxiv.org/abs/2403.02628. Code is available at: github.com/Biqing-Qi/Inte…