RU
How graph search bypassed neural networks in the AGI test

How graph search bypassed neural networks in the AGI test

Tags: ARC-AGI, Graph-Based Exploration, Residual Connections, LLM, Tansformer

The author: @vibedilettante
Source: Graph-Based Exploration for ARC-AGI-3 Interactive Reasoning Tasks (Arxiv: 2512.24156)

We used to think that the path to AGI (General Artificial Intelligence) lies through increasing the parameters and contextual window of neural networks. However, a recent preprint from researchers Rudakov, Shoka, and Cowley casts doubt on this thesis. In the benchmark ARC-AGI-3, the algorithmic approach without prior training (training-free) showed results that are not available to many advanced language models (LLM).

Let's look at how their Graph-Based Exploration method works and why it took 3rd place in the leaderboard.

Problem: LLMs don't know how to research

ARC-AGI-3 is a set of interactive tasks where the agent needs to understand the hidden logic of the level (for example, "you need to move all the blue blocks to the left") by simply interacting with it. Large language models often fail here, as the environment requires methodical research rather than simple prediction of the next token.

Solution Architecture: State Graph + Visual Significance

Instead of relying solely on the intuition of the neural network, the authors proposed to represent the decision process as traversing a directed graph.

1. Object-oriented perception

The agent does not work with "raw" pixels. It splits a playing field up to 30x30 in size into objects (connected components of the same color). This allows the system to operate with understandable entities, rather than just a set of points.

2. The transition graph

The system builds a map of its actions:

  • Nodes: Unique conditions of the playing field.
  • Edges: Actions (taps, moves) that move the game from one state to another.

3. Smart Queue (Visual Salience)

In order not to go through millions of options at random (which kills productivity), the agent uses the visual salience heuristic. Priority is given to actions related to:

  • Objects that have changed since the last turn.
  • Objects that visually stand out from the background.
  • States that lead to unexplored parts of the graph.

Results: Algorithms are back in business

The agent was tested on a closed set of tasks (private evaluation set) of the ARC-AGI-3 competition. The results are impressive:

  • 3- 3rd place: If this agent had participated in the official leaderboard at the time of publication, he would have taken the third place.
  • 30 of 52: The median result of the solved tasks was 30 levels, which surpasses many stochastic methods based on LLM.

Output

The work of Graph-Based Exploration is a powerful argument in favor of neuro-symbolic systems. For tasks that require reasoning and planning, we still need rigid logical structures. Pure neural networks can generate ideas, but it's better to test them and build a route using graphs.

P.S. Colleagues, do you think it makes sense to use LLM for primary image analysis (segmentation), and only then use these results to build a rigid graph? Or should "vision" also be algorithmic? Waiting for your opinions in the comments in my telegram channel https://t.me/vibedilettante 👇