Yes—at least for a narrow kind of visual puzzle, one small model has produced answers without writing out verbalized intermediate reasoning. BDH-CQ uses recurrent memory and iterative computation in a latent space to solve ARC-style tasks. Its reported result is 29.5% pass@2 on the public ARC-AGI-1 evaluation set, but that finding does not show that AI reasoning generally works without language or that the model is broadly intelligent.
What “reasoning without words” means in this study
BDH-CQ is a model described in the preprint “BDH-CQ: In-Context Learning with Recurrent Latent Reasoning,” submitted to arXiv on August 10, 2026. It is designed for problems in which a system must infer a rule from examples and apply it to a new puzzle.
As an Amazon Associate I earn from qualifying purchases.
At inference time, the examples update a recurrent memory. When the query arrives, the model computes an answer through iterative operations in a high-dimensional latent space rather than generating a verbal explanation of its intermediate steps. Zuzanna Stamirowska, a complexity scientist and Pathway CEO, described the process to Science News this way: “Once the query arrives, the model doesn’t write out its thinking in words at all.” She also said, “Nothing in between ever converts into language.”
That is a specific claim about the model’s output and computation in this setup. It does not mean that the model has no internal representations, nor does it establish that every part of its development or operation is language-free. The important distinction is that its intermediate inference is not expressed as a written chain of reasoning.
#1 Best Overall
What result did BDH-CQ report?
The preprint authors report that a 150-million-parameter configuration achieved 29.5% pass@2 on the public ARC-AGI-1 evaluation set. Pass@2 means the reported score allows two attempts; it should not be read as a one-attempt success rate or generalized to every version of ARC-AGI.
The authors also report a computed inference cost of $0.0007 per task. This is an author-reported estimate, not a universal price for AI reasoning. Science News noted that a cost comparison with another model used a different calculation method, so the figures are not a like-for-like measurement. The per-task estimate is best understood as one reported operating point under the authors’ method.
Rank #2
- brand: Pearson
- ARTIFICIAL INTELLIGENCE: A MODERN APPROACH, 4TH EDITION
What ARC-AGI-1 tests—and what it does not
ARC-AGI was introduced by François Chollet in a 2019 paper as a benchmark framed around fluid intelligence. ARC Prize describes its emphasis as measuring how efficiently a system acquires skills on unfamiliar tasks, taking prior knowledge, experience and generalization difficulty into account. In ARC-style puzzles, a solver must infer a transformation from a small set of visual examples and apply it to a new input.
This makes the benchmark relevant to the question of whether a model can learn a new rule from examples. But a score on one evaluation set is not a complete measure of intelligence, and the BDH-CQ result is specifically for ARC-AGI-1. It does not establish equivalent performance on later ARC-AGI versions, ordinary language tasks, or general-purpose reasoning problems.
What kinds of puzzles did it handle?
Science News reported that BDH-CQ did better on some tasks involving turning or moving shapes than on tasks involving color changes or combinations of rules. The report also said its performance on harder ordering and nesting puzzles improved after it saw an example of similar difficulty.
These observations suggest that demonstrations can matter, including demonstrations that resemble a task’s difficulty. They do not establish a universal ranking of puzzle types or guarantee that adding a similar example will improve performance on every problem.
Does avoiding written reasoning make the model more efficient?
Not necessarily as a general rule. A model that does not generate a long written chain can avoid the work of producing those words, and BDH-CQ’s authors report a low computed cost for their setup. But this result alone does not isolate whether the cost comes from the latent-reasoning design, the training method, the benchmark-specific choices, or some combination.
Recommended Free Tools
Yuntian Deng, a computer scientist at the University of Waterloo, called the finding “an interesting efficiency result,” while cautioning that it does not show BDH-CQ’s architecture is better than other approaches. He said more testing is needed to distinguish the effects of model design from training. The cost comparison reported by Science News is also not directly comparable because the calculations differed.
Best Value
BDH-CQ was built specifically for ARC-style problems. Jonas Geiping, a machine-learning researcher at the ELLIS Institute Tübingen and the Max Planck Institute for Intelligent Systems, noted that this makes direct comparison with general-purpose AI systems difficult. A benchmark-specific result should not be presented as proof that latent reasoning is cheaper or more capable across AI applications.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What is lost when a model does not show its steps?
Written reasoning can make an answer easier for a person to follow, question or debug. Latent intermediate computation is less directly inspectable: a correct output does not reveal which internal operations led to it, and an incorrect one may be harder to diagnose.
There is a caveat on the other side, too. Experts quoted by Science News noted that a model’s written chain of thought may not faithfully reveal how it reached an answer. So generated explanations are not automatically transparent evidence of internal reasoning. The trade-off is between having a readable account to examine and relying more heavily on the answer and external tests, not between perfect transparency and none.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minutePC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11How strong is the evidence?
The result comes from an arXiv preprint, not a peer-reviewed paper. Science News reported on September 22, 2026 that the study had not been peer-reviewed at that time. The available evidence cited here does not establish whether peer review or independent replication has happened since then.
That makes the result an interesting demonstration, not a settled conclusion about how AI systems reason. It shows that one model can achieve a measurable score on a defined visual-reasoning evaluation without verbalizing its intermediate inference. It does not show that all reasoning can happen without language, that latent reasoning is inherently superior, or that the benchmark score transfers to broader tasks.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
