Meta-optimization in scientific discovery: 67× 3-SAT speedup

Yuan-Hang Zhang, Chesson Sipling and Massimiliano Di Ventra submitted a 35-page paper titled "Scientific discovery as meta-optimization: a combinatorial optimization case study" to arXiv on 25 Jun 2026 (arXiv:2606.26728). The authors formalize research as meta-optimization and introduce a technique they name "consensus objective aggregation," which combines LLM-generated objective functions via correlation-weighted voting. They apply this to algorithm discovery for 3-SAT instances built on digital MemComputing machines and report that the baseline scaling with problem size N falls from ~N^{2.51} to ~N^{1.33}, delivering a ~67× speedup on the largest instances tested.

What did the authors propose and how does it work?

The paper proposes treating scientific discovery as meta-optimization, meaning the evaluation objective itself is also optimized; the concrete mechanism is "consensus objective aggregation," where multiple LLM-generated objective functions are combined using correlation-weighted voting to form a single evolving evaluation criterion. The opening description frames the research space as a vast set of theories and experiments judged by quality, novelty and validity, and argues that modifying the evaluation criteria alongside exploration improves automated discovery.

Details in the text explain that LLMs are used to generate candidate objective functions and that those candidates are aggregated rather than relying on a single fixed metric. The aggregation is correlation-weighted, which the authors present as producing a stable, self-correcting evaluation criterion that changes as understanding deepens.

How was it tested on 3-SAT and what were the results?

The authors applied their meta-optimization framework to algorithm discovery for 3-SAT problems using digital MemComputing machines and compared measured scaling with problem size N. They report a reduction in the algorithmic scaling exponent from approximately N^{2.51} (the baseline) to approximately N^{1.33} after applying consensus objective aggregation. On the largest instances the paper tested, that change in scaling translated into a reported ~67× speedup.

Those numerical claims are central to the paper: the two concrete data points the authors highlight are the change in scaling exponents (~N^{2.51} to ~N^{1.33}) and the ~67× improvement on the largest tested problems. The manuscript includes six figures across 35 pages that document the experiments and results.

Why it matters

Framing discovery as meta-optimization changes the role of evaluation criteria from fixed yardsticks to tunable components of search. If the results hold beyond the specific 3-SAT and MemComputing tests in this paper, dynamically aggregating objective functions could reduce the burden of hand-designing metrics and allow automated systems to self-correct evaluation biases. That would matter for any domain where objective design limits search efficiency, because the paper ties improved evaluation directly to measurable algorithmic scaling gains.

What to watch

Check for code, data or replication material linked to the arXiv entry and for follow-up tests on other problem classes beyond 3-SAT; the paper lists connections to code and data discovery tools on the arXiv page but the main reported benchmarks are limited to the MemComputing 3-SAT experiments. Also watch for independent reproduction of the reported scaling change from ~N^{2.51} to ~N^{1.33} and the ~67× speedup.

Reference: Yuan-Hang Zhang, Chesson Sipling, Massimiliano Di Ventra, "Scientific discovery as meta-optimization: a combinatorial optimization case study," arXiv:2606.26728 (submitted 25 Jun 2026), https://doi.org/10.48550/arXiv.2606.26728.