China's Endless Frontier Team Releases BigBang-V1, First Open-Source Foundation Model Natively Trained With Recursive Self-Improvement

The Endless Frontier team — a collaboration of Shanghai Jiao Tong University's School of Artificial Intelligence, DP Technology and the Shanghai Algorithm Innovation Research Institute — has released BigBang-V1, which it calls the first foundation model trained natively with recursive self-improving (RSI). During training, generating questions, solving them and verifying answers were all handed to AI, which continuously produced high-quality self-evolving synthetic data from verifiable frontier tasks without humans writing questions one by one.

BigBang-V1 has 35B parameters with about 3B activated at inference and supports a 262K-token context window. Its post-training data is 100% AI-synthesized around frontier scientific tasks. On evaluations spanning long-horizon search, code, scientific research and AI research, it took first place in 10 categories among all 35B-class models, and on hard research benchmarks such as FrontierScience Research and PaperBench it even outperformed the 1T-parameter DeepSeek V4 Pro Preview.

The model and its technical report have been open-sourced. In a transposon localization task, BigBang used junction-evidence-based constraint mapping to lock a 47bp insertion site to coordinate 4144431; in a paper-reproduction task, it rejected three candidate fixes in a row and restored a detached gradient term itself, finishing with a reproduction score of 0.7657, passing 331 of 485 checkpoints.

The team frames the work around a harder question: as models get stronger, fewer humans can write, solve and verify problems for them — so where will high-quality training data come from? That is the core of recursive self-improving, the concept Silicon Valley has been fixated on this year, with Jeff Dean's Discovery Loop and AlphaGo creator David Silver pursuing the same direction.

In implementation, BigBang pairs two kinds of agents: a Generator Agent that keeps proposing and solving hard tasks in science, technology and AI research, and a Critic Agent that reviews candidate data adversarially. On top of that fast inner loop sits an outer loop driven by real training results, which recalibrates the Critic using held-out real research tasks.

The system embeds RSI into the training pipeline itself: results from the previous round of models and experiments shape how the next round of training data is generated and filtered. The team says this data-layer RSI loop breaks the “synthetic data collapse” curse — wrong answers are never fed back into training as if they were correct, because the verification chain keeps catching them.

Humans have not left the loop, though: research goals, budgets, risk boundaries and final acceptance criteria are still defined by people, and AI cannot change evaluation standards on its own. The Endless Frontier team mixes academia and industry — founding members include SJTU associate professor Chen Siheng, SJTU assistant professor and DP Technology founder Zhang Linfeng, and CAS academician E Weinan.

When training-data production no longer depends on humans writing questions one by one, and AI starts helping decide what the next generation of AI should learn, the ceiling on AI progress gets lifted by AI itself. The next competitive edge may lie less in stacking compute and more in data intelligence.