The bottom line: LoopCoder-v2 with two loops substantially improves code reasoning benchmarks (SWE-bench Verified: 43.0 → 64.4 points), while three or more loops become counterproductive due to growing position errors.
Anthropic demonstrates that Transformers with multi-pass blocks are more efficient at understanding and generating code, but only up to two loops deliver actual improvements. Beyond that, performance deteriorates due to positional offsets that accumulate with each additional loop.
Parallel Loop Transformers (PLT) allow computational operations to be distributed across latency by running identical Transformer blocks multiple times in sequence. However, offsets occur in the position indices between loop passes (Cross-Loop Position Offsets, CLP), and KV-cache memory consumption also grows with the number of loops. Anthropic’s team trained LoopCoder-v2, a 7-billion-parameter family of various PLT coder models, from scratch on 18 trillion tokens and subsequently performed matched instruction tuning.
The empirical results demonstrate strongly non-monotonic behavior: the two-loop model delivers broad improvements across all benchmarks – SWE-bench Verified rises from 43.0 to 64.4 points, Multi-SWE from 14.0 to 31.0 points, and code reasoning and agentic software engineering also benefit. Variants with three or more loops, by contrast, show regressions.
Source: arxiv.org · Published June 15, 2026
Lumi AI News — AI-assisted curation according to Article 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.1.