tiny recursive model
In plain wordsA very small AI model that solves hard puzzles by going over its own answer again and again, improving it each time, instead of relying on sheer size.
A small neural network that solves reasoning puzzles by recursion: a single two-layer network of about 7 million parameters repeatedly updates a latent 'reasoning' state and its current answer over a fixed number of steps, and is trained to improve the answer at every round. Trained on about a thousand examples per task with heavy data augmentation, it reached 45% on ARC-AGI-1ARC-AGI, prop. n.: Puzzles made of colored grids — and, since 2026, small video-game-like worlds — that people solve easily but AI long found hard, designed to test whether a machine can learn a new skill from just a few examples. and 8% on ARC-AGI-2 — ahead of many large language modelslarge language model, n.: A computer program trained on enormous amounts of text to predict the words that come next. At vast scale, that simple skill lets it chat, answer questions, translate, summarize and write code. while having under a ten-thousandth of their parameters — and strong results on Sudoku and maze puzzles. It simplified the earlier hierarchical reasoning modelhierarchical reasoning model, n.: A small, brain-inspired AI model that solves puzzles with two linked parts — a slow planner and a fast worker — instead of writing out its reasoning in words., replacing its two networks with one and dispensing with its biological motivation and fixed-point gradient approximation, suggesting that iterating a tiny network can substitute for scale on narrow tasks.
A 7-million-parameter tiny recursive model beat the chatbot on our grid puzzles, though it can't write an email.
Origin
Introduced by Alexia Jolicoeur-Martineau of Samsung's SAIL Montréal laboratory in 'Less is More: Recursive Reasoning with Tiny Networks', posted to arXiv on 6 October 2025.
Usage note
The results concern grid puzzles with extensive task-specific training; such models are not general-purpose language models.
Related
- hierarchical reasoning modelhierarchical reasoning model, n.: A small, brain-inspired AI model that solves puzzles with two linked parts — a slow planner and a fast worker — instead of writing out its reasoning in words.
- ARC-AGIARC-AGI, prop. n.: Puzzles made of colored grids — and, since 2026, small video-game-like worlds — that people solve easily but AI long found hard, designed to test whether a machine can learn a new skill from just a few examples.
- latent reasoninglatent reasoning, n.: Getting an AI to do its step-by-step thinking inside its internal numbers rather than writing it out in words — potentially faster and more flexible, but much harder for people to inspect.
- recurrent neural networkrecurrent neural network, n.: A neural network that reads a sequence, such as a sentence or an audio clip, one piece at a time, carrying a running summary forward so that what came earlier can shape how later pieces are understood.
- test-time computetest-time compute, n.: The computing effort an AI model spends while answering a question, as opposed to while being trained. Letting a model think longer or try many answers can make it markedly smarter.
- small language modelsmall language model, n.: A language model built to be compact enough to run cheaply and quickly, even on a laptop or phone, trading some knowledge and skill for speed, privacy and lower cost.
- large language modellarge language model, n.: A computer program trained on enormous amounts of text to predict the words that come next. At vast scale, that simple skill lets it chat, answer questions, translate, summarize and write code.
Sources
- Jolicoeur-Martineau, A. (2025). Less is More: Recursive Reasoning with Tiny Networks. arXiv:2510.04871. arxiv.org/abs/2510.04871
Mentioned in
hierarchical reasoning modelhierarchical reasoning model, n.: A small, brain-inspired AI model that solves puzzles with two linked parts — a slow planner and a fast worker — instead of writing out its reasoning in words.