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.
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.
Z3 is capable (it’s an SMT solver, not just SAT), but it’s not very fast at boolean satifiability and not at all competitive with modern SOTA SAT solvers. Try comparing it to Chaff or Glucose e.g.
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.
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.
One problem here is it's very easy to overtune to a past problem set -- even accidentally. You can often significantly improve performance just by changing your random number generator seed until you happen to pick the right assignment for the first few variables of some of the harder problems.
It would be interesting to take the resulting solver and apply it to an unknown data set.
yess. loads of space for further exploration here. there is an attempt to keep things as general as possible in the expert.md file, but hard to mitigate overfitting fully. however, changing the seed will not get you much further with all else in the solver constant. unless you try a number of seed that exponentially scales with the size of the problem
Very interesting. For me the key question is whether this kind of agent can generalize to real SAT application domains, not only benchmark instances. In problems like timetabling, encoding choices, auxiliary variables, and branching strategy can matter a lot. If it can help there too, this is a very meaningful direction.
sure. in the limit, everything is parameter tuning. with large enough NP-hard problems, the complexity of the search space is big enough that its infeasible to get to a better state by just tuning params in any reasonable amount of time.
I beg to disagree. Integer programming solvers have improved orders of magnitude in the past 20 years. The basic algorithm (branch and bound) is the same.
The big commercial solvers basically are very good at picking up structures and selecting the tuning parameters that work better for specific problem types.
Not as many changes to the files under library as I expected to see. Most changes seemed to be under a single ‘add stuff’ commit. If some of the solvers are randomised, then repeatedly running and recording best solution found will continually improve over time and give the illusion of the agent making algorithmic advancements, won’t it?
I guess my point was that I don't see many algo changes in the commit history, which is a shame if this has been lost; library/* files are largely unchanged from the initial commits. But each time the agent runs, it has access to the best solutions found so far and can start from there, often using randomisation, which the agent claims helps it escape local minima e.g. 'simulated annealing as a universal improver'. It would be nice to see how its learnt knowledge performs when applied to unseen problems in a restricted timeframe.
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
It would be interesting to take the resulting solver and apply it to an unknown data set.
It is parameter tuning. We have been doing it for centuries.
The big commercial solvers basically are very good at picking up structures and selecting the tuning parameters that work better for specific problem types.
[1] https://github.com/google-deepmind/alphadev