We release ParEVO, a new framework capable of synthesizing performant parallel algorithms for irregular data structures. By decoupling standard code generation from strict unit tests, our Evolutionary Coding Agent achieves an average 106x speedup on ParEval through performance tests.
Read full post →We introduce QEDBench, a 272-problem benchmark that decouples mathematical proof generation from verification to reveal systemic limitations in frontier LLM reasoning. Our evaluation of 5 solvers and 7 LLM judges against 1,000+ hours of expert grading reveals a dangerous Sycophancy Trap and the Discrete-Continuous Reasoning Gap.
Read full post →