Perception-First ARC Solver

A novel approach to solving ARC-AGI-2 puzzles by focusing on perception design rather than reasoning complexity.

Key Insight

LLMs can reason, but they can't see. The bottleneck isn't reasoning capability—it's how we present visual information.

From baseline 26.7% to ~50% accuracy on ARC-AGI-2 by improving perception, not model size.

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                  PERCEPTION-FIRST SOLVER                        │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│  ┌─────────────────┐                                           │
│  │  PERCEPTION     │  ← Slim (~150 tokens)                     │
│  │  (perception_   │  ← Task type detection                    │
│  │   slim.py)      │  ← Diff maps                              │
│  │                 │  ← Key transitions                         │
│  └────────┬────────┘                                           │
│           │                                                     │
│           ▼                                                     │
│  ┌─────────────────────────────────────────────────────────┐   │
│  │              PARALLEL EXPERTS (4-8)                      │   │
│  │                                                          │   │
│  │   Temperature 1.0 for diversity                          │   │
│  │   Each expert iterates up to 10 times                    │   │
│  │   Diff-based feedback for learning                       │   │
│  │                                                          │   │
│  └─────────────────────────────────────────────────────────┘   │
│           │                                                     │
│           ▼                                                     │
│  ┌─────────────────┐                                           │
│  │  VOTING         │  ← Consensus from passing solutions       │
│  │                 │  ← Early stopping on correct answer       │
│  └─────────────────┘                                           │
│                                                                 │
└─────────────────────────────────────────────────────────────────┘

Results

Setup	ARC-AGI-2 Accuracy
GPT-5.2 medium baseline	26.7%
GPT-5.2 + perception + 4 experts	~50%

Installation

cd perception_solver
pip install anthropic openai google-generativeai numpy scipy

Set API keys:

export OPENAI_API_KEY="your-key"
export ANTHROPIC_API_KEY="your-key"  # optional
export GEMINI_API_KEY="your-key"     # optional

Usage

cd perception_solver

# Run single task
python run_parallel.py --task 0934a4d8 --experts 4

# Run first N tasks
python run_parallel.py --count 10 --experts 4

# Run all evaluation tasks
python run_parallel.py --experts 4

Configuration

Edit config.py to change model and settings:

MODEL = "openai/gpt-5.2"  # or anthropic/claude-opus-4-5

SOLVER_CONFIG = {
    "max_iterations": 10,
    "temperature": 1.0,
    "reasoning_effort": "medium",  # for GPT-5
    "thinking_budget": 0,          # for Claude (0 = disabled)
}

Research

See PERCEPTION_RESEARCH.md for the theoretical foundation based on:

• Biological vision (V1→V2→V4→IT hierarchy)
• Gibson's affordance theory
• Gestalt principles
• Predictive coding

Key Files

File	Purpose
run_parallel.py	Main entry point
perception_slim.py	Slim perception (~150 tokens)
parallel_solver.py	Multi-expert coordination
solver.py	Single expert solver
llm.py	Unified LLM interface (OpenAI, Anthropic, Gemini)
prompts.py	Prompt templates
sandbox.py	Safe code execution

Citation

@article{perception-first-arc-2025,
  title={Perception-First Approach to ARC-AGI-2:
         LLMs Can Reason, But They Can't See},
  year={2025}
}

License

MIT