In a paradigm-shifting revelation for AI research, Samsung’s Advanced Institute of Technology (SAIT) unveiled the Tiny Recursive Model (TRM) on October 7, 2025, via a groundbreaking arXiv paper titled “Less is More: Recursive Reasoning with Tiny Networks.” Crafted by Senior AI Researcher Alexia Jolicoeur-Martineau, this featherweight 7-million-parameter network eclipses behemoths like Google’s Gemini 2.5 Pro and OpenAI’s o3-mini on grueling reasoning benchmarks—proving that clever recursion trumps sheer scale in cracking complex puzzles. At under 0.01% the size of trillion-parameter titans, TRM heralds an era where affordability meets superior smarts, challenging the “bigger is better” dogma that’s dominated AI for years.
TRM’s secret sauce? A streamlined recursive loop that mimics human-like self-correction, iteratively refining answers without ballooning compute demands. Starting with an embedded question xx, initial answer yy, and latent state zz, the two-layer transformer (with rotary embeddings and SwiGLU activations) performs up to 16 supervised steps. In each, a “deep recursion” of three passes—two gradient-free for exploration, one for learning—unfolds into a “latent recursion” of six updates: tweaking zz via the network, then polishing yy. This emulates 42-layer depth per step, using Adaptive Computational Time (ACT) to halt via a simple Q-head probability. For fixed-context tasks like Sudoku, it swaps self-attention for an MLP (5M params); larger grids like ARC-AGI retain attention (7M params). Trained on scant data (~1,000 examples) with heavy augmentation—shuffles for Sudoku, rotations for mazes—TRM leverages Exponential Moving Average for stability, dodging overfitting that plagues scaled-up rivals.
The results are staggering. On Sudoku-Extreme (9×9 grids), TRM nails 87.4% accuracy, dwarfing its predecessor Hierarchical Reasoning Model (HRM) at 55%. Maze-Hard (30×30 paths) sees 85.3% success, up from HRM’s 74.5%. But the crown jewel is ARC-AGI, AI’s Everest for abstract reasoning: TRM scores 44.6% on ARC-AGI-1 and 7.8% on ARC-AGI-2, outpacing Gemini 2.5 Pro (37%/4.9%), o3-mini-high (34.5%/3%), and DeepSeek R1 (15.8%/1.3%). Even Grok-4-thinking (1.7T params) lags at 16% on ARC-AGI-2, while bespoke tweaks hit 29.4%—still shy of TRM’s efficiency. Ablations confirm recursion’s magic: sans it, accuracy plummets to 56.5% on Sudoku.
Jolicoeur-Martineau champions this minimalism: “The idea that one must rely on massive foundational models trained for millions of dollars… is a trap. With recursive reasoning, it turns out that ‘less is more’.” Community buzz echoes her: X users dub it “10,000× smaller yet smarter,” with Sebastian Raschka praising its HRM simplification as a “two-step loop that updates reasoning state.” Open-sourced on GitHub under MIT license, TRM’s code includes training scripts for a single NVIDIA L40S GPU—democratizing elite reasoning for indie devs and startups.
This isn’t just a win for Samsung; it’s a reckoning for AI’s scale obsession. As labor shortages and energy costs soar, TRM spotlights recursion as a sustainable path to AGI-like feats on structured tasks, from logistics puzzles to drug discovery grids. Yet caveats linger: it’s a solver, not a conversationalist, excelling in visuals but untested on open-ended prose. Future tweaks could hybridize it with LLMs, but for now, TRM whispers a profound truth: In the quest for intelligence, tiny thinkers may lead the charge.
Leave a Reply