The video discusses new 2025 AI models, Hierarchical Reasoning Models (HRM) and Tiny Recursive Models (TRM), which use recursion and iterative refinement at inference time to enhance complex reasoning without increasing model size. These models represent a shift from traditional large language models by enabling deeper, task-specific reasoning through recursive architectures, potentially leading to more efficient and powerful AI systems that combine recursion with large-scale knowledge.
In this episode of Decoded, the discussion centers on the emerging trend of using recursion in AI models to enhance reasoning capabilities without merely increasing model size. The conversation highlights two pivotal 2025 papers introducing Hierarchical Reasoning Models (HRM) and Tiny Recursive Models (TRM), which leverage recursion at inference time to improve performance on complex tasks like Sudoku and mazes. Unlike traditional large language models (LLMs) that process inputs in a single feed-forward pass, these recursive models repeatedly apply the same weights through multiple loops, enabling deeper reasoning through iterative refinement.
The hosts delve into the limitations of standard LLMs, particularly their inability to efficiently perform tasks requiring multiple sequential reasoning steps, such as sorting or solving incompressible problems. LLMs lack an external memory mechanism akin to a Turing machine’s tape, which restricts their computational efficiency and reasoning depth. While chain-of-thought prompting introduces a form of recursion in output token space, it remains bounded by the training data and does not equate to inherent recursive reasoning within the model architecture itself.
HRM models draw inspiration from biological neural processes, incorporating multiple hierarchical levels that operate at different frequencies, mimicking brain activity. The key innovation lies in their training method, which uses truncated backpropagation through time combined with a fixed-point iteration approach called deep equilibrium learning. This technique avoids the vanishing gradient problems typical of traditional recurrent neural networks by not backpropagating through all recursion steps, instead iteratively refining latent states across multiple passes. This approach allows a relatively small model to achieve state-of-the-art results on complex reasoning tasks without extensive pretraining.
TRM builds on HRM by simplifying the architecture, collapsing hierarchical networks into a single shared network and refining the backpropagation strategy to include one full recursion step. This results in a smaller, more efficient model that outperforms HRM on benchmark tasks despite having fewer parameters. Both models employ an expectation-maximization-like iterative optimization process, updating latent variables and memory states to progressively improve task performance. This recursive reasoning mechanism enables the models to discover problem-solving strategies without explicit chain-of-thought supervision, marking a significant advancement over traditional feed-forward LLMs.
The broader implication of this research is a potential paradigm shift in AI development, where recursion and iterative refinement become central to building models capable of complex reasoning. While current recursive models are task-specific, integrating these techniques with large-scale, general-purpose LLMs could yield powerful hybrid systems that combine vast knowledge representation with efficient, scalable reasoning. This approach challenges the prevailing notion that bigger models are inherently better, suggesting instead that smarter architectures leveraging recursion may unlock new frontiers in artificial intelligence.