Most AI progress today comes from making models bigger. A new research paper takes the opposite route. Its model has just 7 million parameters, tiny next to chatbots with hundreds of billions, and it gets better at hard puzzles simply by “thinking” in loops for longer, even past 20,000 steps.
The method is called InfiLoop. It’s described in a preprint posted to arXiv on October 8, 2026, by Pengxiang Li, Dilxat Muhtar, Di He, Guinan Su, Lu Yin and Shiwei Liu. The paper has not yet been peer-reviewed.
What is a looped Transformer?
A Transformer is the basic building block behind most modern AI models. A looped Transformer reuses the same small block over and over. Each loop adds more “thinking” without adding more parameters, so in theory the model can tackle harder problems just by running more loops.
The catch, according to the researchers, is that more loops can actually make things worse. In the usual setup, each loop simply passes its latest output to the next one. If one loop makes a mistake, it can overwrite a correct answer found earlier. The authors call this the “carry-last” rule.
They tested this with TRM, an earlier tiny reasoning model, on very hard Sudoku puzzles. Its accuracy peaked at 62.2% and then dropped to 58.1% as it kept looping. By the end, it had lost almost 20% of puzzles it had already solved.
How InfiLoop fixes it
Instead of blindly passing on the newest output, InfiLoop keeps a running summary of earlier outputs. It scores each new output and decides how much of it to accept, while older outputs slowly fade. Importantly, the memory this needs stays the same no matter how many loops the model runs.
The researchers found something interesting when they looked inside. Without being told which answers were correct, the model learned to give low weight to updates while it was still searching, accept the update that completed a puzzle, and then hold that solution steady. The authors describe this as an emergent self-verification behaviour.
The results
According to the paper, the 7M-parameter InfiLoop model scored:
- 97.9% on Sudoku-Extreme, compared with 94.2% for FPRM and 74.7% reported for TRM
- 87.4% on Maze-Hard, a 30×30 shortest-path maze test
- 49.1% on ARC-AGI-1, a visual pattern-reasoning test (pass@2, meaning two attempts per task)
- 13.6% on the harder ARC-AGI-2 (pass@2), nearly double TRM’s 7.8%
In a separate test, it kept improving on Sudoku as the loops grew to 24,960 executed layers. It also matched another model’s accuracy using half the computation, according to the authors. And every puzzle InfiLoop solved stayed solved, even when the researchers deliberately added noise to its internal state.
What to keep in mind
These are impressive numbers for such a small model, but they need context. The paper is a preprint and hasn’t been independently checked. The authors compare against the best results reported by other teams, and retrained only some of them. Each model is also trained on a specific task. This isn’t a general chatbot that can do Sudoku, write emails and answer questions.
ARC-AGI-2 is still very hard. 13.6% is a big jump for a tiny model, but most of the tasks remain unsolved. And extra loops aren’t free: each one costs computing time, so “think longer” has a price.
Why it matters
As a writer, I find this idea refreshing. Much of the AI conversation is about who has the most chips and the biggest models. Research like this suggests that smarter design can sometimes beat sheer size, which matters for places like India where computing budgets are tight. I also like that the model learned to protect its correct answers. That sounds a lot like good human problem-solving: don’t throw away what already works. I’d want to see independent tests before calling it a breakthrough, though.
For more on AI research, read how IBM added reasoning to decades-old information theory, how reasoning mode changed one AI model’s behaviour and how an MIT algorithm helps AI follow strict safety rules.