What's this about?
Real code review is an iterative conversation between developer and reviewer, which is what makes it costly and slow. Existing benchmarks model it as a single static judgement. MCR-Bench makes it dynamic — multi-turn, with the back-and-forth included.
Why does the multi-turn part matter?
Because the first review comment is rarely the useful one. The value comes from the exchange: the reviewer asks why, the author explains, the reviewer sees the constraint they missed or the author sees the bug they missed. A model evaluated on producing one good comment in isolation is being tested on the easiest and least valuable part of the job, which is why single-shot code-review benchmarks flatter models.
Why does this matter to you?
- This is directly the shape of AI-assisted review in a security workflow: findings, pushback, refinement.
- It provides a way to measure whether a review tool actually helps or just produces plausible first comments.
- The static-to-dynamic critique applies to almost every benchmark in this digest.