Research

Paper

AI LLM March 09, 2026

Adaptive Loops and Memory in Transformers: Think Harder or Know More?

Authors

Markus Frey, Behzad Shomali, Ali Hamza Bashir, David Berghaus, Mehdi Ali

Abstract

Chain-of-thought (CoT) prompting enables reasoning in language models but requires explicit verbalization of intermediate steps. Looped transformers offer an alternative by iteratively refining representations within hidden states. This parameter efficiency comes at a cost, as looped models lack the storage capacity of deeper models which use unique weights per layer. In this work, we investigate transformer models that feature both adaptive per-layer looping, where each transformer block learns to iterate its hidden state via a learned halting mechanism, and gated memory banks, that provide additional learned storage. We find that looping primarily benefits mathematical reasoning, while memory banks help recover performance on commonsense tasks compared to parameter and FLOP matched models. Combining both mechanisms yields a model that outperforms an iso-FLOP baseline -- with three times the number of layers -- on math benchmarks. Analysis of model internals reveals layer specialization: early layers learn to loop minimally and access memory sparingly, while later layers do both more heavily.

Metadata

arXiv ID: 2603.08391
Provider: ARXIV
Primary Category: cs.CL
Published: 2026-03-09
Fetched: 2026-03-10 05:43

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2603.08391v1</id>\n    <title>Adaptive Loops and Memory in Transformers: Think Harder or Know More?</title>\n    <updated>2026-03-09T13:49:03Z</updated>\n    <link href='https://arxiv.org/abs/2603.08391v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2603.08391v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Chain-of-thought (CoT) prompting enables reasoning in language models but requires explicit verbalization of intermediate steps. Looped transformers offer an alternative by iteratively refining representations within hidden states. This parameter efficiency comes at a cost, as looped models lack the storage capacity of deeper models which use unique weights per layer. In this work, we investigate transformer models that feature both adaptive per-layer looping, where each transformer block learns to iterate its hidden state via a learned halting mechanism, and gated memory banks, that provide additional learned storage. We find that looping primarily benefits mathematical reasoning, while memory banks help recover performance on commonsense tasks compared to parameter and FLOP matched models. Combining both mechanisms yields a model that outperforms an iso-FLOP baseline -- with three times the number of layers -- on math benchmarks. Analysis of model internals reveals layer specialization: early layers learn to loop minimally and access memory sparingly, while later layers do both more heavily.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.CL'/>\n    <published>2026-03-09T13:49:03Z</published>\n    <arxiv:comment>Published at Latent &amp; Implicit Thinking Workshop @ ICLR 2026</arxiv:comment>\n    <arxiv:primary_category term='cs.CL'/>\n    <author>\n      <name>Markus Frey</name>\n    </author>\n    <author>\n      <name>Behzad Shomali</name>\n    </author>\n    <author>\n      <name>Ali Hamza Bashir</name>\n    </author>\n    <author>\n      <name>David Berghaus</name>\n    </author>\n    <author>\n      <name>Mehdi Ali</name>\n    </author>\n  </entry>"
}