Pengfei Liu

AI can now design better AI: ASI-Evolve discovers SOTA architectures, curates pretraining data, and designs RL algorithms, outperforming human-designed baselines by significant margins.

A 7B model trained on a new dataset of Chinese porcelain outperforms GPT-4 by 12% on expert connoisseurship tasks, demonstrating the power of domain-specific training and tool integration.

LLMs can't even reproduce published physics papers end-to-end, with the best model scoring only 34% on a new benchmark designed for this purpose.

Pretraining isn't just about scaling data volume; daVinci-LLM's ablations reveal that data processing depth, domain-specific strategies, and compositional balance are equally critical for unlocking LLM capabilities.

Achieve high-fidelity transparent text animations from image-to-video models without retraining the VAE, sidestepping data scarcity and latent pattern mixing issues.

Forget toy datasets: OpenSWE delivers 45K+ real-world, executable Python environments for leveling up your SWE agent, and it's all open-sourced.

Today's best AI agents fail at realistic software engineering tasks, stalling before even reaching 30% completion, revealing the urgent need for better long-horizon planning and human-AI collaboration.

Frontier LLMs can unlock substantial performance gains in scientific domains by refining and completing raw scientific text, leading to a +8.40 point improvement on domain-aligned tasks.