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

2026-06-16

AI Summary

This work addresses the latency and linearly growing key-value (KV) cache memory overhead in recurrent Transformers during inference, both stemming from sequential recurrence. To overcome these limitations, the authors propose a parallel recurrent Transformer architecture featuring cross-loop positional offsets (CLP) and a gated sliding-window attention mechanism with shared KV caches, enabling efficient exploration of recurrence depth during both training and inference. The study systematically reveals, for the first time, a non-monotonic relationship between recurrence depth and the trade-off between performance gains and computational cost, identifying two recurrence steps as the optimal balance and providing principled selection criteria. Experimental results demonstrate that the two-step model significantly outperforms non-recurrent baselines across multiple code generation and agent-based software engineering benchmarks, improving the SWE-bench Verified score from 43.0 to 64.4, while deeper recurrences lead to performance degradation.

Abstract

Looped Transformers scale latent computation by repeatedly applying shared blocks, but sequential looping increases latency and KV-cache memory with the loop count. Parallel loop Transformers (PLT) alleviate this cost through cross-loop position offsets (CLP) and shared-KV gated sliding-window attention, making loop count a practical design choice. We therefore study PLT loop-count selection through a gain--cost view: an extra loop may refine representations, but CLP also introduces a positional mismatch at each loop boundary. We instantiate this study by training LoopCoder-v2, a family of 7B PLT coders with different loop counts, from scratch on 18T tokens, followed by matched instruction tuning and evaluation. Empirically, the two-loop variant delivers broad gains over the non-looped baseline across code generation, code reasoning, agentic software engineering, and tool-use benchmarks, improving SWE-bench Verified from 43.0 to 64.4 points and Multi-SWE from 14.0 to 31.0 points. In contrast, variants with three or more loops regress, revealing a strongly non-monotonic loop-count effect. Our diagnostics show that loop 2 provides the main productive refinement, while later loops yield diminishing, oscillatory updates and reduced representational diversity. Because the CLP-induced mismatch remains roughly fixed as refinement gains shrink, the offset cost increasingly dominates. This gain--cost trade-off explains PLT's saturation at two loops and provides diagnostics for loop-count selection.

Problem

• loop count
• Parallel Loop Transformers
• cross-loop position offsets
• computation scaling
• gain-cost trade-off

Innovation

• Parallel Loop Transformers
• cross-loop position offsets
• shared-KV attention
• loop-count selection
• gain-cost trade-off