Apex Intelligence’s automated AI research system reports new official best results on three public benchmarks and a near-top result on a fourth.
Research across the training stack
The system is evaluated on problems that operate at different layers of modern AI systems: predicting scaling behavior before large training runs, improving a language model within a fixed compute budget and optimizing performance-critical H100 GPU kernels.
The four evaluations cover SLDBench, NanoChat Autoresearch, GPUMode TriMul and MLS-Bench. Each result is measured under its own fixed protocol, hardware setting and numerical tolerance.
- SLDBench: average score of 0.8846, up from 0.8613.
- NanoChat Autoresearch: 0.892426 val_bpb on one B200 in 300 seconds, ahead of Recursive SuperIntelligence (0.903891) and Tencent Hunyuan Hyra (0.901543).
- GPUMode TriMul on H100: 1,036.1 μs, 2.71% faster than the strongest human-designed solution in same-machine tests.
- MLS-Bench fused causal attention on H100: 430.7, 468.4 and 451.7 TFLOP/s, up 27.18%, 15.63% and 17.11% on the previous best.
Discovery must survive verification
The system does more than propose ideas. It writes and modifies code, runs experiments, compares outcomes with baselines and preserves the artifacts needed to audit each result. Public code and evaluation records make it possible to trace claims back to evidence.
AI improving AI only matters when each proposed improvement can be tested, reproduced and bounded by a clear evaluation protocol.
What the results do and do not show
The report demonstrates automated discovery within four well-specified tasks. It does not establish general superiority across all research problems, deployment settings or safety requirements. The value of the work lies in a repeatable loop from question to experiment to verified result.