🤖 AI Summary
Researchers propose Tiny Recursive Model (TRM), a radically parameter‑efficient recursive reasoning architecture that outperforms much larger LLMs on hard puzzle benchmarks. Using a single tiny 2‑layer network with just 7M parameters, TRM achieves 45% test accuracy on ARC‑AGI‑1 and 8% on ARC‑AGI‑2—exceeding most commercial LLMs (e.g., Gemini 2.5 Pro ≈4.9%) while using under 0.01% of their parameters. The paper also reports large gains on other tasks (Sudoku‑Extreme from 55%→87%, Maze‑Hard 75%→85%) compared with prior approaches.
Technically, TRM simplifies and improves on the earlier Hierarchical Reasoning Model (HRM), which used two small networks recursing at different temporal frequencies plus deep supervision, adaptive compute (ACT), and a 1‑step gradient/fixed‑point approximation (HRM ≈27M params). TRM instead applies recursive latent refinement in a single tiny network across multiple supervised “improvement” steps (up to Nsup ≈16), retaining the benefits of deep supervision and iterative correction without HRM’s complexity. Significance: this challenges the idea that raw scale is required for algorithmic reasoning, showing recursion + multi‑step supervision yields major generalization gains with tiny models—opening paths for energy‑efficient, deployable reasoning systems and prompting research into why recursion converges and when fixed‑point/1‑step gradient approximations hold.
Loading comments...
login to comment
loading comments...
no comments yet