LoopCoder-v2: Only Loop Once for Efficient Test-Time Computation Scaling

LoopCoder-v2 introduces a family of 7 billion parameter Parallel Loop Transformers (PLT) designed to optimize test-time computation by utilizing cross-loop position offsets (CLP) and shared-KV gated sliding-window attention. The model was trained on 18 trillion tokens and demonstrated significant performance improvements across various benchmarks, notably enhancing the SWE-bench Verified score from 43.0 to 64.4 with a two-loop configuration, while higher loop counts resulted in diminishing returns due to positional mismatches. This research provides insights into the trade-offs of loop count selection, which is crucial for practitioners aiming to balance computational efficiency and model performance in AI applications.