---
license: mit
tags:
  - arc-agi
  - arc-agi-2
  - program-synthesis
  - symbolic-reasoning
  - puzzle-solver
  - cegis
  - no-llm
  - no-neural-network
datasets:
  - arc-agi-2
metrics:
  - accuracy
model-index:
  - name: Verantyx-arc-agi2-7.4
    results:
      - task:
          type: arc-agi-2
          name: ARC-AGI-2
        dataset:
          type: arc-agi-2
          name: ARC-AGI-2 Training
          split: training
        metrics:
          - type: accuracy
            value: 7.4
            name: Accuracy (%)
---

# Verantyx ARC-AGI-2 — 7.4%

**Pure program synthesis solver for ARC-AGI-2** — no neural networks, no LLMs, no hardcoded patterns.

## Score

| Split | Score | Method |
|---|---|---|
| Training (1000 tasks) | **74/1000 = 7.4%** | DSL synthesis + CEGIS verification |

## Approach

Zero-shot program synthesis:

1. **Decompose** — Break input→output relationships into composable grid transformations
2. **Synthesize** — Generate candidate programs from 50+ DSL operations
3. **Verify** — CEGIS against ALL training pairs (must match exactly)
4. **Compose** — 2-step pipelines when single rules don't suffice

**No cheating:** Test outputs are never accessed. Only training I/O pairs are used for rule discovery.

## Key Innovation: Neighborhood Rule Learning

The most powerful operation: learns a deterministic mapping from each cell's local neighborhood (3×3 or 5×5 window) to its output value. **Solves 240 of 1000 training tasks alone.**

This is not a neural network — it's an exact lookup table built from training examples and verified to be consistent across all pairs.

## DSL Operations (50+)

| Category | Operations |
|---|---|
| **Geometric** | rotate_90/180/270, flip_h/v, mirror_h/v/hv, transpose, reverse_rows/cols, roll_rows/cols |
| **Structural** | crop_bbox, crop_to_color, extract_largest/smallest_region, extract_unique_subgrid |
| **Morphological** | erode, dilate, hollow_regions, fill_interior, extract_border, fill_enclosed |
| **Color** | colormap, replace_color, recolor_by_size, keep_one_color, remove_color |
| **Sorting** | row_sort, col_sort (by color_count, sum, first_nonbg) |
| **Gravity** | gravity_all/up/down/left/right |
| **Connection** | connect_h/v/hv, spread_color (4 directions) |
| **Tiling** | tile_to_output, corners_mirror, stack_h/v/h_flip/v_flip, self_tile, diagonal_tile |
| **Subgrid** | subgrid_select/overlay/diff (automatic separator detection) |
| **Dedup** | dedup_rows, dedup_cols |
| **Learned** | neighborhood_rule (radius 1-2 neighborhood mapping) |

## Score Progression

| Version | Score | Key Change |
|---|---|---|
| v1 | 1.6% | Initial DSL: colormap, mirror, scale |
| v5 | 2.5% | WholeGridProgram class, rotations |
| v10 | 2.9% | Subgrid ops, CompositeProgram |
| v12 | 4.1% | extract_region, stack ops |
| v14 | 5.3% | corners_mirror, connect ops |
| v17 | 6.1% | neighborhood_rule learning |
| **v19** | **7.4%** | **+18 DSL ops, priority reorder** |

## Usage

```bash
git clone https://github.com/Ag3497120/verantyx-arc-agi2
cd verantyx-arc-agi2

# Download ARC-AGI-2 data
git clone https://github.com/arcprize/arc-agi-2.git /tmp/arc-agi-2

# Run evaluation
python -m arc.eval_cross --split training
```

## Properties

- ✅ **No neural networks** — pure symbolic reasoning
- ✅ **No LLMs** — no language model of any kind
- ✅ **No hardcoded patterns** — all rules are synthesized from training data
- ✅ **No test data leakage** — only training I/O pairs are used
- ✅ **Deterministic** — same input always produces same output
- ✅ **Zero dependencies** — pure Python, no pip install needed
- ✅ **Fast** — ~0.4s per task average

## Links

- **GitHub**: [Ag3497120/verantyx-arc-agi2](https://github.com/Ag3497120/verantyx-arc-agi2)
- **HLE Solver**: [kofdai/verantyx-hle-2.6](https://huggingface.co/kofdai/verantyx-hle-2.6) (same Verantyx philosophy)
- **ARC-AGI-2**: [arcprize.org](https://arcprize.org/)