Interview, Webinar
Recursion Is The Next Scaling Law In AI
- Research Trend: The episode focuses on a 2025 trend in AI research utilizing recursion at inference time to improve model reasoning, rather than relying solely on scaling model parameters.
- Key Papers: The discussion centers on two specific architectures: Hierarchical Reasoning Models (HRM) and Tiny Recursive Models (TRM).
- HRM Performance: The HRM model achieved state-of-the-art results on the ARC-Prize 1 and 2 benchmarks using only 27 million parameters trained on just 1,000 tasks without pre-training.
- TRM Performance: The TRM model, a further simplification, reduced parameters to 7 million while increasing ARC-Prize 1 accuracy to 87%, outperforming the HRM and larger standard models.
- LLM Limitation (Sorting): Standard LLMs face a theoretical lower bound on sorting tasks (n log n) because they lack an external memory tape and cannot perform more than their fixed number of layers allow in a single feed-forward pass.
- RNN Historical Context: Recurrent Neural Networks (RNNs) naturally support recursion but historically failed at scale due to "backpropagation through time" (BPTT), causing vanishing or exploding gradients and excessive memory requirements for hidden states.
- HRM Architecture: HRMs employ a three-level hierarchy: a low-level network (TL steps), a high-level network (TH steps), and an outer refinement loop (NSUP steps), all sharing weights to enable recursive processing.
- Training Technique (DEQ): HRMs utilize a Deep Equilibrium Model (DEQ) approach with truncated backpropagation through time (specifically backpropagating only once per loop) to avoid BPTT gradient issues while maintaining performance.
- Memory as Computation: The recurrence in these models acts as a dynamic, latent memory cache (similar to a Turing machine tape), allowing the model to iteratively refine hidden states rather than processing all context in a single pass.
- TRM Architectural Simplification: TRMs collapse the HRM's separate low-level and high-level networks into a single shared network (weight-sharing), reducing the architecture to one transformer layer while maintaining distinct latent variables (ZL and ZH).
- Optimization Method: TRMs introduce a "latent recursion" step during training where gradients are backpropagated through a full recursive loop, effectively turning the training process into an Expectation-Maximization (EM) algorithm within the latent space.
- Test-Time Compute: Research indicates that while training on higher recursion counts (e.g., 16) is beneficial, testing with only a single recursion step yields nearly equivalent performance to testing with high counts.
- Bioplausibility Debate: The speakers note that while bioplausibility often inspires initial architectures, the most effective ML systems often deviate from biological reality to optimize for GPU efficiency and computational depth.
- Chain-of-Thought vs. Inherent Recursion: Current LLM reasoning relies on Chain-of-Thought (CoT), which is bounded by human-labeled traces and discrete token spaces, whereas recursive models can discover reasoning strategies internally within continuous latent spaces.
- Future Direction: The hosts suggest a future convergence where giant LLMs are combined with the recursive inference architectures of TRMs/HRMs to achieve efficient, deep reasoning without the parameter bloat of current models.
- Generalization Gap: Unlike LLMs which are general-purpose, current recursive models are task-specific; the next research goal is adapting these efficient recursive architectures to general-purpose agents capable of zero-shot reasoning.