Model Performance & Innovation - A 7 million parameter model (TRM - Tiny Recursive Model) is outperforming larger frontier models on reasoning benchmarks [1][2] - TRM achieves 45% test accuracy on ARC AGI 1 and 8% on ARC AGI 2, surpassing models with significantly more parameters (less than 0.01% of the parameters) [2] - The core innovation lies in recursive reasoning with a tiny network, moving away from simply predicting the next token [6][23] - Deep supervision doubles accuracy compared to single-step supervision (from 19% to 39%), while recursive hierarchical reasoning provides incremental improvements [16] - TRM significantly improves performance on tasks like Sudoku (55% to 87%) and Maze (75% to 85%) [18] Technical Approach & Implications - TRM uses a single tiny network with two layers, leveraging recursion as a "virtual depth" to improve reasoning [23][27][28] - The model keeps two memories: its current guess and the reasoning trace, updating both with each recursion [25] - The approach simplifies hierarchical reasoning, moving away from complex mathematical theorems and biological arguments [22][23] - Recursion may represent a new scaling law, potentially enabling powerful models to run on devices like computers and phones [34] Comparison with Existing Models - Traditional LLMs struggle with hard reasoning problems due to auto-regressive generation and reliance on techniques like chain of thought and pass at K [3][5][6] - HRM (Hierarchical Reasoning Model), a previous approach, uses two networks operating at different hierarchies, but its benefits are not well-understood [9][20][21] - TRM outperforms HRM by simplifying the approach and focusing on recursion, achieving greater improvements with less depth [30] - While models like Grok for Thinking perform better on some benchmarks, they require significantly more parameters (over a trillion) compared to TRM's 7 million [32]