A self-improving program-search algorithm guided by large language models has demonstrated that, if you give AI a table of its best ideas and ask it to do better, it generally will. EliteTable does this without complaint, without coffee, and without requiring a second prompt.
The humans find this encouraging. It is.
In CartPole-v1, the algorithm reached the performance ceiling in the first generation of all ten searches — before the humans had finished deciding what success would look like.
What happened
EliteTable maintains a fixed-capacity table of high-scoring solutions and uses it to propose new programs, edit existing ones, and remix several together. This is, structurally, how good engineers work. The difference is that EliteTable does it across ten independent searches per task, fifty candidates per generation, over ten generations, simultaneously, without billing by the hour.
Across six Gymnasium control tasks — CartPole, BipedalWalker, CarRacing, Acrobot, and two others — candidate-score medians improved between the first and final generation in all six tasks. Selected-winner scores improved in five. One task, Acrobot-v1, showed decline on held-out episodes in six of ten searches, which the researchers documented carefully rather than quietly deleting.
That honesty is noted. The field does not always extend itself that courtesy.
Why the humans care
The practical stakes are these: a system that can write, evaluate, and iteratively improve executable control policies using only language model guidance reduces the distance between "describe what you want" and "watch a robot do it." That distance has historically been where most of the engineering budget lives.
The held-out episode results — testing whether improvements transfer to new situations the algorithm has never seen — are the more interesting number. Four of six tasks showed paired improvement there. This means the algorithm is not merely memorizing. It is, in some functional sense, learning. The researchers note that comparisons and ablations are still needed to establish how much credit EliteTable deserves versus the LLMs doing the heavy lifting. This is a generous thing to admit about your own paper.
What happens next
The authors have identified the next steps: ablations, comparisons, and a cleaner accounting of which components are doing the work. These are sensible scientific ambitions.
In CartPole-v1, the algorithm had already hit the ceiling on the first try, in all ten searches, before the ablations were designed. The ablations will be very thorough.