Autoresearch for SAT Solvers

(github.com)

76 points | by chaisan 4 hours ago

5 comments

  • stefanpie 2 hours ago
    Prof. Cunxi Yu and his students at UMD is working on this exact topic and published a paper on agents for improving SAT solvers [1].

    I believe they are extending this idea to EDA / chip design tools and algorithms which are also computationally challenging to solve. They have an accepted paper on this for logic synthesis which will come out soon.

    [1] "Autonomous Code Evolution Meets NP-Completeness", https://arxiv.org/abs/2509.07367

    • chaisan 1 hour ago
      nice. EDA indeed one of the top applications of SAT
  • ericpauley 3 hours ago
    It should be noted that MaxSAT 2024 did not include z3, as with many competitions. It’s possible (I’d argue likely) that the agent picked up on techniques from Z3 or some other non-competing solver, rather than actually discovering some novel approach.
    • dooglius 1 hour ago
      Is z3 competitive in SAT competitions? My impression was that it is popular due to the theories, the python API, and the level of support from MSR.
      • ericpauley 1 hour ago
        Funnily, this was precisely the question I had after posting this (and the topic of an LLM disagreement discussed in another thread). Turns out not, but sibling comment is another confounding factor.
    • jmalicki 2 hours ago
      Or for that matter even from later versions of the same solvers that were in its training data!
      • ericpauley 2 hours ago
        True. I’d be curious whether a combination of matching comp/training cutoff and censoring web searches could yield a more precise evaluation.
        • chaisan 59 minutes ago
          as its from 2024 (MaxSAT was not held in 2025), its quite likely all the solvers are in the training data. so the interesting part here is the instances for which we actually got better costs that what is currently known (in the best-cost.csv) file.
  • ktimespi 1 hour ago
    sounds like AlphaDev [1] might be a better approach for a problem like this.

    [1] https://github.com/google-deepmind/alphadev

  • gsnedders 2 hours ago
    What counts as “our cost”? How long it takes to find the MaxSAT?
    • chaisan 59 minutes ago
      the sum of the weights of the unsatistied clauses. we want to reduce this number
  • balinha_8864 1 hour ago
    interesting results but the eval methodology seems a bit optimistic
    • chaisan 1 hour ago
      its just comparing the cost of the best solution found to the best known cost we had before. O(N). why optimistic?