CodeReviewer: large-scale pre-training for review tasks
Early foundation model for automating code review activities—quality estimation and comment generation from diffs and review data.
Source: Automating Code Review Activities by Large-Scale Pre-training
Before the agent era, CodeReviewer showed you could pre-train specifically for review: diffs in, quality estimates and review comments out, using large-scale code-review corpora. It became a standard baseline for later LLM fine-tunes (LLaMA-Reviewer and friends).
It is still useful context: many “new” LLM review demos rediscover tasks this line of work already framed—comment generation, refinement, quality scoring—without the agent scaffolding.
Why it matters for code review
When you evaluate a flashy multi-agent reviewer, keep CodeReviewer-era tasks in the mix. If a system cannot beat a specialized review pretrain on comment quality, the agents may be expensive theater around a weaker core.