We just open-sourced the state of the art in ARC-AGI-2! (GitHub link in the reply.) - The first valid solution to almost saturate the benchmark (92.5% - slight improvement on our earlier announcement). - Slashed cost ($3.9 per task) compared to previous SOTA. - A learning multi-agent harness for iterative self-improvement. We have already submitted the solution to @arcprize for verification on the (semi-) private test set. Feel free to fork it to build your own agents. The approach is quite general and not limited to ARC-AGI. We are preparing a detailed blog post. In the meantime, check out the repo and its readme on github, as well as our earlier announcement quoted below for some high-level details. For the longest-horizon, open ended, non-verifiable tasks, continual learning beyond simple in-context learning and beyond backpropagation is needed. More on that soon. At @noemon_ai we are changing how AI scales by changing how it learns. @GregKamradt @fchollet @mikeknoop

Announcing: ARC-AGI-2 top score and at a fraction of the cost through an agentic learning harness for iterative self-improvement. Public Eval — 91% @ $3.3/task. SOTA compared to even new models trained for extreme reasoning like GPT-5.4 Pro and Gemini 3 Deep Think, while also reducing their cost by more than 4x. We used Gemini 3.1 Pro (Preview) and we increased its score by 12 percentage points. @GregKamradt @fchollet @mikeknoop

Our method's effectiveness and efficiency relies on learning, i.e. internalizing lessons from experience into the model, not only iterating on model output artifacts such as code. In fact, differently from many recent approaches, we do not use any code execution at all. Learning in this case where we don't have access to the model's internals, and where the horizon is relatively short, is achieved by simply conditioning the model on the history of its past attempts and their outcomes. Learning and reasoning is grounded on the real task environment: hypotheses are tested on the grid, and feedback comes from the world, not just from the model's own reasoning trace. The core of our approach is a simple ReAct loop: We separate reasoning/planning from validating/acting: a dedicated Reasoner learns from full history to improve its natural-language instructions which it provides to a dedicated Validator that executes, validates, and returns feedback to the Reasoner. The image shows an example problem along with the natural language solution that the Reasoner hypothesized and provided to the Validator. The Reasoner describes it as "Lasers" that "shoot inwards" The meta-cognitive capabilities of the new Gemini model are also critical as they help decide when the (learning) process can be stopped.