Memory-Efficient Looped Transformer: Decoupling Compute from Memory in Looped Language Models

2026-05-08

AI Summary

This work addresses the challenge of linearly growing memory consumption in recurrent large language models during multi-step reasoning, which stems from maintaining separate key-value (KV) caches at each inference step and hinders scalability. To overcome this limitation, the authors propose MELT, an architecture that decouples reasoning depth from memory usage by sharing a single KV cache across layers and dynamically updating it through a learnable gating mechanism. A two-stage lightweight post-training strategy—comprising chunked training, interpolation-based transition, and attention-alignment distillation—is introduced to stabilize optimization. Fine-tuned from the Ouro pretrained model, MELT achieves constant memory footprint while matching the performance of standard LLMs, significantly outperforming same-scale baselines and substantially reducing memory overhead compared to the original Ouro.

Abstract

Recurrent LLM architectures have emerged as a promising approach for improving reasoning, as they enable multi-step computation in the embedding space without generating intermediate tokens. Models such as Ouro perform reasoning by iteratively updating internal representations while retaining a standard Key-Value (KV) cache across iterations, causing memory consumption to grow linearly with reasoning depth. Consequently, increasing the number of reasoning iterations can lead to prohibitive memory usage, limiting the practical scalability of such architectures. In this work, we propose Memory-Efficient Looped Transformer (MELT), a novel architecture that decouples reasoning depth from memory consumption. Instead of using a standard KV cache per layer and loop, MELT maintains a single KV cache per layer that is shared across reasoning loops. This cache is updated over time via a learnable gating mechanism. To enable stable and efficient training under this architecture, we propose to train MELT using chunk-wise training in a two phase procedure: interpolated transition, followed by attention-aligned distillation, both from the LoopLM starting model to MELT. Empirically, we show that MELT models fine-tuned from pretrained Ouro parameters outperform standard LLMs of comparable size, while maintaining a memory footprint comparable to those models and dramatically smaller than Ouro's. Overall, MELT achieves constant-memory iterative reasoning without sacrificing LoopLM performance, using only a lightweight post-training procedure.

Problem

• recurrent LLM
• memory consumption
• reasoning depth
• KV cache
• iterative reasoning

Innovation

• Looped Transformer
• Memory Efficiency
• KV Cache Sharing
• Iterative Reasoning
• Attention Distillation