For the past ten years, artificial intelligence has been dominated by LLMs. GPT-5, Gemini, Claude: models with hundreds of billions of parameters, trained on oceans of data, capable of writing a novel or coding an application in seconds.
However, a Canadian researcher has just demonstrated that comparable, or even superior, performance can be achieved on reasoning tasks with a model 200 times smaller and 200 times more energy-efficient. At Yiaho, we closely follow these advancements and are always ready to innovate. Today, we’re going to explain the world of Tiny Recursive Models, or TRMs!
What is a Tiny Recursive Model?
A TRM is an extremely compact artificial neural network, with only 1 to 10 million parameters, designed specifically for structured reasoning. Unlike LLMs (Large Language Models) which learn to predict the next word in a sentence, TRMs learn to solve problems step-by-step, like a human solving an equation or a logic puzzle.
The secret? Four technical pillars:
- No massive pre-training: gone are the terabytes of text scraped from the internet. The model is trained on a targeted, synthetic dataset, often generated by the model itself.
- Deep supervision: each layer of the network receives a direct learning signal, not just the final output. This forces the model to learn useful representations at all levels.
- Deep recursion: the network calls itself iteratively, like a function that recalls itself until convergence.
- Recursive data augmentation: with each iteration, the model generates new examples from its own errors, enriching its own training corpus.
Result: a model that doesn’t memorize, but truly reasons. For example, the TinyReasoner-7M, with only 7 million parameters, excels at complex puzzles like Sudoku, mazes, or ARC-AGI evaluation, benchmarks where even giants like Gemini 2.5 Pro struggle.
Backtracking: The Big Difference from LLMs
Imagine you’re playing chess. An LLM plays move by move, never going back. If it makes a mistake on the 5th move, the whole game is lost. A TRM, however, can say: “This path leads to failure. I’ll go back three moves and try something else.”
This is recursive backtracking. The model:
- Generates a partial reasoning sequence.
- Evaluates it with an internal scoring function (e.g., “Does this step follow logical rules?”).
- If the score is too low, it goes back, modifies a previous step, and restarts.
- It repeats until a valid solution is obtained.
Major consequence: the TRM does not generate its response token by token. It works internally, refines, corrects, optimizes… then delivers a complete, coherent, and final answer in one block. It’s like a mathematician filling pages of rough work before showing you the final, unblemished proof.
This approach drastically reduces “hallucinations,” those absurd errors that LLMs often make, and allows for much higher precision on tasks requiring pure logic.
The Paper That Shook the Markets
It all began on October 6, 2025, with the publication of “Less is More: Recursive Reasoning with Tiny Networks” by Alexia Jolicoeur-Martineau, principal researcher at Samsung SAIT AI Lab in Montreal. (Paper available in the sources at the bottom of the page).
In this document, she presents the Tiny Recursive Model (TRM), inspired by the Hierarchical Reasoning Model (HRM), but simplified to the extreme. The 7-million-parameter TRM:
- Achieves 45% on ARC-AGI-1 and 8% on ARC-AGI-2, outperforming models like DeepSeek-R1 or Gemini 2.5 Pro.
- Beats LLMs on tasks like Sudoku, mazes, and abstract reasoning, with only 1000 training examples.
- Runs on a single GPU RTX 4090 for inference, and trains in a few hours.
- Consumes 0.5 watt-hours per inference, compared to 100+ for an equivalent LLM.
Two days after publication, Samsung’s stock jumped 10.3% on the Seoul Stock Exchange! Analysts are already talking about revolutionary edge AI: a model capable of reasoning like a human, but fitting into a smartwatch or an industrial sensor. The buzz spread on Reddit and Medium, where open-source AI communities are enthusiastic about this proof that “less is more.” In plain English, we can say that simplicity can often be more effective than complexity!
Why TRMs Can Change Everything?
1. They work without the cloud
No need to send your data to a remote server. The model runs locally, on your phone, your car, your fridge. Privacy guaranteed, and near-zero latency.
2. They are energy efficient
A TRM consumes the equivalent of an LED bulb. An LLM? The equivalent of a small building. In a world where AI already accounts for 3% of global electricity consumption, this is crucial for mass deployment.
3. They are reliable by design
Thanks to backtracking, the model almost never hallucinates on structured tasks. It checks, corrects, validates. Ideal for medicine (logical diagnosis), finance (calculation verification), security (route planning), or even video games.
4. They cost almost nothing to train
A few thousand dollars and a weekend are enough. Any lab, startup, or student can create one. The code is already open-source on GitHub, under the SamsungSAILMontreal/TinyRecursiveModels repo, with a ready-to-use PyTorch implementation!
In comparison, training an LLM like GPT-3 costs millions and requires GPU farms. TRMs democratize reasoning AI, making what was once reserved for tech giants accessible.
TRMs: The Future of AI?
TRMs cannot write poems, hold fluid conversations, or generate creative images. Their domain? Pure reasoning: math, logic, planning, diagnosis, games, code verification. They shine where LLMs stumble, such as on ARC-AGI’s abstract puzzles, which test general intelligence without massive data bias.
But the future is promising:
- TRM + LLM Hybrids: the TRM handles the heavy reasoning, the LLM drafts the response in natural language for a more human interaction.
- Integration into Galaxy S26: Samsung is already preparing embedded assistants based on TRMs for local and energy-efficient AI.
- Expanding Open-Source: the paper’s code is on arXiv, and GitHub forks are proliferating. Tutorials for training your own TRM in 3 hours are already emerging on Medium.
With the recent excitement, from heated discussions on X to analyses on Hugging Face, TRMs could well mark the turning point towards more efficient and accessible AI.
In summary: less really is more!
Tiny Recursive Models are not a passing fad. They are proof that in AI, size is not the solution to everything. With a few million parameters, good architecture, and a lot of ingenuity, complex problems can be solved reliably, quickly, and economically.
Alexia Jolicoeur-Martineau has demonstrated it: recursion and backtracking transform “small” models into reasoning champions. TRMs are AI that thinks before it speaks. Which reminds us of a forgotten lesson: sometimes, intelligence is born from constraint.
Source: Arxiv / Less is More: Recursive Reasoning with Tiny Networks
