Dynamic Parametric Evaluation Suite for LLM Reasoning Benchmarks

Hi everyone,

I’ve been tracking the recent engineering developments around reasoning capabilities and symbolic logic. I wanted to share a benchmarking suite I recently built natively on the Hub to explicitly stress-test frontier models on multi-step analytical math and physics execution.

It programmatically targets structural failures where model gradients collapse or spatial constraints fail:

I’d love to know if anyone in the research community is interested in using these parametric JSON suites to stress-test your upcoming model iterations or logic pipelines!