Pursuing Open-Ended Intelligence through
Self-Evolving Synthesis of Verifiable Frontier Tasks
BigBang is a general-purpose large language model evolved from Qwen3.6-35B-A3B through efficient post-training, driven by an adversarial, self-evolving synthetic data framework grounded in verifiable frontier research tasks. Trained with simple SFT on only 10.6K high-difficulty examples, BigBang-v1 rivals agents trained with extensive continual pre-training and reinforcement learning.
Problems at the boundary of human knowledge — including those for which no known optimum exists.
Candidate solutions can be assessed using objective rules, formal methods, computational tools, simulators, or experimental feedback.
The strongest overall performance among ∼30B ReAct-based search agents — without continual pre-training, without RL.
Real frontier tasks expose capability gaps; a generator–critic loop turns them into high-quality synthetic data; the trained model is validated on real tasks — and the next round of evolution begins.
Strategy Update · Task Construction
A code agent running a continuous search over programs: it directly modifies, executes, and debugs the data-synthesis code, then refines its generation strategy from critic feedback. An agent harness records motivation, results, and failure causes each round, enabling long-horizon exploration.
Quality Evaluation · Constraint Checks
A two-level adversarial review. Basic constraints — trajectory information, tool-use validity, format integrity — quickly remove degraded outputs. Higher-level evaluation judges evidence, contradiction, verifiability, and research depth: a fast proxy for training value.
Outcome-Calibrated Co-Evolution
Model variants trained on sampled pipelines are evaluated on held-out real research tasks. The meta-critic compares the critic's assessments with observed training outcomes — distinguishing tasks that merely appear difficult from those that genuinely improve capability.