When Question Banks Can't Keep Up, AI Starts Setting Its Own Problems: The Endless Frontier Team Runs Through the Data-Level RSI

The Endless Frontier team, composed of researchers from Shanghai Jiao Tong University's AI Institute, DP Technology, and the Shanghai Institute of Algorithm Innovation, has launched BigBang-V1, the first base model trained natively using recursive self-improving (RSI). In this approach, AI handles problem generation, solving, and verification, producing high-quality self-evolving synthetic data without human involvement per task. The model, with 35B parameters and about 3B active during inference, supports a 262K long context and is entirely trained on AI-synthesized data focused on frontier science tasks. It achieved 10 first places among 35B models in benchmarks including long-horizon search, coding, scientific research, and AI research, and even surpassed the 1T-parameter DeepSeek V4 Pro Preview on hard scientific tasks like FrontierScience Research and PaperBench. Examples of its capability include precisely locating a transposon insertion site by using junction evidence, and replicating a paper into runnable code by self-correcting during smoke tests. The team emphasizes that the data production system itself can be optimized, requiring tasks that are both frontier and verifiable, and they implement a dual-loop pipeline: an inner loop with Generator and Critic agents that produce and vet data, and an outer loop driven by real training results to calibrate the system. This data-level RSI loop prevents synthetic data collapse, while humans still define goals, budgets, risks, and acceptance criteria. The model and code are open-sourced on Hugging Face and GitHub.