Standard feed-forward transformers (LLMs) have inherent architectural limitations for complex, multi-step reasoning, as they cannot perform algorithms requiring more sequential steps than the model has layers.
Two recent papers on Hierarchical Reasoning Models (HRM) and Tiny Recursive Models (TRM) reintroduce recursion at inference time, enabling small models (7-27M parameters) to outperform massive LLMs on specific reasoning benchmarks like ARC-Prize.
The key innovation is a stable training method, inspired by Deep Equilibrium Models, that backpropagates through a single, truncated recursive step, overcoming the historical instability of Recurrent Neural Networks (RNNs).
The future of AI may involve hybrid systems that combine the vast knowledge and powerful latent representations of large models with the efficient, recursive reasoning capabilities of smaller models like TRM.
12 quotes
Concerns Raised
Recursive models like HRM and TRM are currently task-specific and not general-purpose.
Training LLMs to reason via Chain of Thought requires human-labeled traces, which are not always available for novel problems.
Backpropagation through time remains a fundamental challenge for deep recurrent models, despite recent workarounds.
Opportunities Identified
Combining the recursive reasoning of TRM-like models with the broad knowledge of large-scale LLMs.
Using large models to create powerful latent representations and then applying smaller recursive models to reason within that space.
Further exploration of the "outer refinement loop" and truncated backpropagation as powerful new training techniques.