From Static to Dynamic: Benchmarking Real-World Code Review with MCR-Bench

TL;DR - MCR-Bench evaluates LLMs on realistic, multi-round code review with defect lifecycle tracking. Results show current models struggle increasingly as reviews lengthen, exposing weaknesses in temporal alignment and long-range memory.

• Contains 2,269 real-world review tasks across five programming languages.
• Annotates defect descriptions, types, severity levels, and state changes across review rounds.
• Mainstream LLMs perform poorly at both defect detection and lifecycle tracking, with performance degrading over additional rounds.
• Semantically complex and low-salience defects are missed more often; errors are linked to cross-round temporal misalignment and inadequate long-range memory.