LLM merge solvers reach 59% on Java conflicts

LLM solvers matched developers on 55% to 59% of real Java merge conflicts, versus 36.7% for AutoMerge, in a ConflictBench evaluation; the 18-to-22-point edge came mostly from coverage because traditional tools abstained on 20% to 90% of conflicts. The calibrated LLM judge still accepted four of five outputs that failed a deterministic structural check. As teams adopt stacked pull requests for agent-sized review units, an AI resolver can propose the patch, but parsers, compilers, and tests should retain veto power.