NeuroGolf 2026

Tiny ONNX networks that solve ARC-AGI puzzles · 用极小的 ONNX 网络求解 ARC-AGI 谜题

---

📖 Overview · 项目概览

EN — Solution framework for the 2026 NeuroGolf Championship on Kaggle. The competition asks for ONNX networks that solve ARC-AGI tasks while minimizing memory_bytes + parameter_count. A task that passes every train / test / arc-gen example earns max(1, 25 − ln(memory + params)) points; one that fails any example scores zero.

中文 —— Kaggle 2026 NeuroGolf 锦标赛 的求解框架。比赛要求为每个 ARC-AGI 任务提交一个 ONNX 网络，在通过全部 train / test / arc-gen 样例的前提下，最小化 内存字节数 + 参数数量。单任务得分为 max(1, 25 − ln(memory + params))，任一样例失败即得零分。

---

📊 Results · 当前成绩

Submission · 提交	Solvers added · 新增求解器	Tasks pass · 通过任务	Score · 公开分数
v1	identity, zero, single-color, remap	4	81.57
v2	+ transpose	6	131.57
v3	+ marker-crop (opset 11)	7	149.93
v4	+ static-crop, kron-scale, resize-scale, shape-aware flip/rot180	15	276.86
v5	+ rot90-ccw (transpose ∘ flip-v)	16	290.62
v6	+ bbox-strip (Slice + Sub + ReduceMax/ArgMax + Gather + Less + Mul)	17	303.98
v7	+ shift (Slice+Concat+Pad) + tile-h (shape-aware Mod-Gather)	19	~337 (pending)
v8	+ palindrome H / V / 2D (Where-based mirror-concat)	26	~435 (pending)
v9	+ majority-fill (ReduceSum→TopK→Where→OneHot)	27	~453 (pending)
v10	+ variable-kron (Div+Min+Mul masked Resize-by-N)	29	480.56
v11	+ conv 1×1 / 3×3 / 5×5 masked (least-squares + bias)	31	~509 (pending)
v12	+ bbox-color-extract, split-and, scale-detector, variable-shift, gravity-right	35	567.92
v13	+ flood-fill-enclosure (border BFS unrolled with Pad+Conv+Max)	36	578.92
v14	+ column-label (ArgMax pairwise ranking for per-column top-row labeling)	37	592.92
v15	+ repeat-top-rows (runtime period detection P=2/3/4 with weighted candidate mix)	39	~605 (pending)
v16	+ block-mask (N×N→N²×N² masked tiling + channel-0 recovery with Tile×inv_mask)	40	~616 (pending)
v17	+ gravity-right-diag (per-channel center-of-mass diagonal shift with stay/shift mask)	41	~628 (pending)
v18	+ gravity-down (size-aware per-channel column sink via content-mask height + ramp threshold)	45	697.88
v19	+ self-fractal (Tile × Kron-upscaled selector mask; fixed / most-frequent colour)	47	726.38
v20	+ rot-tile (four-quadrant rotational tiling I / rot90 / rot180 / rot270)	48	742.85
v21	+ gravity-up (stable column rise via TopK sort-key + GatherElements, multi-colour columns)	49	755.96
v22	+ periodic-fill (period in-painting; static period enumeration + log-doubling shift-Max propagation)	53	786.58
v23	+ split-logic (two-half boolean combine: and/or/xor/nor over L-R or T-B split, ± divider)	62	~924 (pending)
v24	+ connect-dots (fill the gap between the first and last same-colour dot per row/column)	64	~949 (pending)
v25	+ mirror-complete (restore an erased half as the shape-aware mirror of the present half)	66	975.27
v26	+ denoise (remove isolated single cells via depthwise 3×3 same-colour neighbour count)	67	~988 (pending)
v27	+ connect-fill (fill the gap between aligned same-colour dots with a fixed colour)	70	~1024 (pending)
v28	+ drop-into-wall (colour-1 cells fall into a full colour-5 wall row in their column)	71	~1037 (pending)
v29	+ color-bbox-fill (replace each colour's scattered cells with its solid bounding box)	72	~1051 (pending)

---

🚀 Quick Start · 快速开始

# 1. Install deps · 安装依赖
pip install -r requirements.txt

# 2. Download competition data · 下载竞赛数据
python scripts/download.py

# 3. Build networks for all 400 tasks · 为全部 400 个任务构建网络
python scripts/build_all.py

# 4. Package the .onnx files · 打包 .onnx 文件
python scripts/package_submission.py

# 5. Submit to Kaggle · 提交到 Kaggle
python scripts/submit.py submissions/submission.zip "describe your run"

Kaggle requires the submission file be named exactly submission.zip. Kaggle 要求提交文件必须命名为 submission.zip。

---

🧩 Solvers · 求解器一览

Each solver is a callable (task: dict) → Optional[onnx.ModelProto]. The pipeline runs every solver, verifies each candidate against the official scorer, and keeps the highest-scoring network.

每个求解器是一个 (task: dict) → Optional[onnx.ModelProto] 调用，pipeline 会运行所有求解器，用官方评分器验证每个候选网络，保留得分最高的网络。

Solver · 求解器	Pattern · 识别模式	ONNX ops · 算子	Params · 参数
solve_identity	output ≡ input · 输出等于输入	Identity	0
solve_remap	per-pixel color lookup · 像素级颜色映射	1×1 Conv	100
solve_single_color	uniform fill, same shape · 同尺寸纯色填充	1×1 Conv	100
solve_transpose	output is input transposed · 输出为输入的转置	Transpose	0
solve_static_crop	fixed-offset, fixed-size subrect of input · 定长定位裁剪	Slice + Pad	~14
solve_kron_scale	N×N pixel expansion, constant input shape · 同尺寸像素块复制	2× Gather + Pad	~40
solve_resize_scale	N× nearest upscale, variable input shape · 变尺寸 N 倍最近邻放大	Slice + Resize	~24
solve_marker_crop	crop a fixed window around a unique marker pixel · 围绕唯一标记像素的定长裁剪	Slice → ReduceSum → ArgMax → Slice → Pad	~27
solve_flip_h_aware / solve_flip_v_aware	content-aware horizontal / vertical flip · 形状感知水平 / 垂直翻转	ReduceSum + ReduceMax + Mul(arange) + Sub + Clip + Cast + Gather + Mul	~37
solve_rot180_aware	two stacked shape-aware flips · 双轴形状感知翻转	flip-h ∘ flip-v	~74
solve_rot90_ccw_aware / solve_rot90_cw_aware	shape-aware 90° rotation · 形状感知 90° 旋转	Transpose ∘ flip-v	~37
solve_bbox_strip	crop input to the bbox of non-background cells · 非背景外接矩形提取	ReduceSum + Sub + row/col ArgMax + Gather + Less + Mul	~76
solve_shift	constant-shape translation with color-0 fill on the gap · 定形平移并以色 0 填充	Slice + Concat(with init fill) + Pad	~50
solve_tile_h	horizontal tile-N with variable input width · 变宽 N 倍水平复刻	ReduceSum + ReduceMax + Mod + Gather + Less + Mul	~67
solve_palindrome_h / solve_palindrome_v	input mirror-concatenated to its right / bottom edge · 镜像拼接	shape-aware Where + Gather + Less mask	~68
solve_palindrome_2d	four-quadrant 2D mirror · 四象限二维镜像	palindrome-h ∘ palindrome-v	~133
solve_majority_fill	constant-shape rectangle filled with the majority non-bg color · 常尺寸主色填充	ReduceSum + TopK + Greater + And + Where + OneHot + Mul + Pad	~32
solve_variable_kron	scale by N where N is count(non-zero) or count(distinct colors) · 变 N 倍 Kronecker（N 来自输入特征）	ReduceSum + Cast + Div + Min (float) + Gather ×2 + Less + Mul	~138
solve_bbox_color_extract	crop input to bbox of majority/rarest color · 按主色/稀有色外接矩形裁剪	ReduceSum + ArgMax/ArgMin + OneHot + Mul + Cast + Gather + Mod + Less	~66
solve_split_and	split input along a color-5 vertical separator, AND the two halves · 沿颜色5分隔列分割并对两半做AND	Slice + Pad + Sub + Mul + Less + Cast + And	~8129
solve_split_logic	split into two halves (L-R or T-B, ± divider) and combine with and/or/xor/nor/nand · 切两半并做布尔运算	Slice + ReduceSum + Mul/Max/Sub (boolean) + Pad + Concat	~20
solve_connect_dots	fill the span between the first and last same-colour dot in each row/column · 连接每行/列首尾同色点之间的间隔	Slice + CumSum + Gather (reverse) + Greater + And + Concat	~37
solve_mirror_complete	restore an erased half as the shape-aware mirror of the present half · 用形状感知镜像补全被擦除的另一半	shape-aware flip (ReduceMax + Sub + Clip + Gather) + Mul + Add merge	~42
solve_denoise	remove isolated single cells (no same-colour 8-neighbour) · 删除无同色邻居的孤立点	depthwise 3×3 Conv (hollow kernel) + Greater + Mul + Concat	~81
solve_connect_fill	fill the gap between aligned same-colour dots with one fixed colour · 用固定色连接同行/列的成对同色点	CumSum ×4 (h/v prefix+suffix) + Greater + And + Or + ReduceMax + Concat	~64
solve_drop_into_wall	colour-1 cells fall into a full colour-5 wall row in their column · 颜色1沿列落入整行的颜色5墙	Gather + ReduceSum/ReduceMax (wall + column detect) + Mul (outer product) + Relu + Concat	~12
solve_color_bbox_fill	replace each colour's scattered cells with its solid bounding box · 用实心外接矩形填充每个颜色的散点	ReduceMax (row/col span) + CumSum ×4 + Greater + And + Mul (outer) + Concat	~80
solve_scale_detector	N× nearest-neighbor upscale or downscale · N 倍最近邻放大或缩小	Slice + Resize	~24
solve_variable_shift	shift by fixed offset with zero-fill, variable input shape · 定偏移量平移零填充，变输入尺寸	Slice + Pad + Concat	~50
solve_gravity_right	each row's cells slide right until blocked · 重力向右：每行格子右移直到被挡住	ReduceSum + CumSum + Where + Mul	~94
solve_gravity_down	each column's cells fall to the bottom edge · 重力向下：每列格子下落堆积到底部	ReduceSum + ReduceMax (content-mask height) + Less + Mul + Slice + Concat	~66
solve_gravity_up	each column's cells rise to the top, preserving order (multi-colour) · 重力向上：每列保序上浮，支持多色列	ReduceSum + sort-key + TopK + Tile + GatherElements	~42
solve_periodic_fill	restore a periodic tiling from a partially-erased grid · 从被部分擦除的网格恢复周期平铺	per-period Pad+Slice+Max (log-doubling) + Where period select + bbox clip	~9k
solve_self_fractal	N×N input tiles into an N²×N² self-similar fractal, keyed by a selector colour · N×N 输入自相似放大为 N²×N²，按选择色决定哪些块保留	Slice + Tile + Gather ×2 (Kron mask) + ArgMax + Mul + Pad + Concat	~57
solve_rot_tile	N×N square tiles into 2N×2N as four rotations (I / rot90 / rot180 / rot270) · N×N 方阵拼为 2N×2N 的四象限旋转	Slice + Transpose + Gather ×4 (index-reverse) + Concat ×3 + Pad	~23
solve_gravity_right_diag	per-channel diagonal slide: compute center-of-mass, shift non-rightmost cells toward it · 按通道对角线滑移：计算质心，将非最右像素移向质心方向	ReduceSum + ReduceMax + ArgMax + Mul + Slice + Concat + Min + Sub	150
solve_conv3x3	least-squares fit of a 3×3 conv (no bias) · 无偏置 3×3 卷积拟合	3×3 Conv	900
solve_conv1x1_masked / solve_conv3x3_masked / solve_conv5x5_masked	K×K conv + bias, masked to non-padding cells · K×K 卷积带偏置和非填充掩码	Conv + ReduceSum + Mul	100 / 910 / 2510
solve_zero	output is empty grid · 输出为空网格	Sub	0
solve_flood_fill_enclosure	fill color-0 cells fully enclosed by a single source color · 填充被单一颜色完全包围的色 0 区域	ArgMax + Equal + 58×(Pad+Conv+Greater+Max) border flood-fill + Gather + Sub + Slice + Concat	~1.8k
solve_block_mask	N×N→N²×N²: tile input row/col-wise, mask[0-block]→zero, channel-0 recovery via Tile×inv_mask · 块掩码平铺：按块复制输入，零块掩蔽，通道 0 恢复	Slice + ReduceSum + Less + Tile + Mul + Add	78

---

🗂️ Project Layout · 项目结构

NeuroGolf/
├── src/neurogolf/
│   ├── grids.py          # one-hot conversions      · 独热编码转换
│   ├── onnx_ops.py       # ONNX graph helpers       · ONNX 图构建辅助
│   ├── verify.py         # clean-room scorer        · 评分器独立实现
│   ├── pipeline.py       # solver orchestration     · 求解器调度
│   └── solvers/          # per-family solvers       · 各类求解器
├── scripts/
│   ├── download.py             # pull data via Kaggle CLI · 下载比赛数据
│   ├── build_all.py            # run pipeline over tasks  · 跑全量任务
│   ├── package_submission.py   # zip the .onnx files      · 打包提交
│   └── submit.py               # post to Kaggle           · 上传 Kaggle
├── networks/             # generated .onnx files    · 生成的网络
├── submissions/          # packaged submission.zip  · 打包的提交文件
└── tests/                # pytest cases             · 单元测试

---

📏 Scoring & Constraints · 评分与约束

Per-task score · 单任务得分

points = max(1, 25 − ln(memory_bytes + params))   # if all examples pass · 全样例通过时
       = 0                                         # otherwise · 否则

Constraints · 约束条件

Aspect · 项目	Requirement · 要求
Input / output tensor · 输入输出张量	(1, 10, 30, 30) float32, one-hot · 独热编码
Tensor names · 张量名	"input" / "output"
Output decoding · 输出解码	thresholded at > 0.0 · 以 > 0.0 阈值化
File size · 文件大小	≤ 1.44 MB per .onnx · 单文件不超过 1.44 MB
Banned ops · 禁用算子	LOOP, SCAN, NONZERO, UNIQUE, SCRIPT, FUNCTION, COMPRESS, any Sequence*, graph-typed attributes · 任何 Sequence*、子图属性
Shapes · 形状	statically inferable; declare value_info when shape inference can't see through dynamic ops · 必须静态可推断；动态算子需显式声明 value_info
ARC-gen grids · arc-gen 样例	grids larger than 30×30 are skipped by the scorer · 超过 30×30 的样例由评分器自动跳过

---

🔬 Tests · 单元测试

python -m pytest tests/ -q

20 cases cover the one-hot round-trip contract and per-family solver shape contracts (Identity / Conv1×1 / Transpose / Slice+Pad / Gather+Pad / Resize / shape-aware flip / rot / bbox / shift / tile-h).

20 个测试覆盖独热编码的往返一致性，以及各求解器输出图的形状契约 （Identity / Conv1×1 / Transpose / Slice+Pad / Gather+Pad / Resize / 形状感知翻转 / 旋转 / 外接矩形 / 平移 / 横向复刻）。

---

🛣️ Roadmap · 后续计划

• Bounding-box extraction (ReduceSum + ArgMax row/col projections) · 外接矩形提取（行列投影 + ArgMax）
• Shape-aware flip / rotate that handles top-left alignment · 处理左上角对齐的形状感知翻转 / 旋转
• Scale-by-N tasks via static Resize or Tile · 用 Resize / Tile 实现 N 倍缩放任务
• Constant-shape translation (shift with color-0 gap) · 定形平移并填补色 0 间隙
• Shape-aware horizontal tile-N (Mod-Gather-mask pipeline) · 形状感知横向 N 倍复刻
• Generalize marker-crop to handle multi-pixel markers and per-color crop rules · 通用化标记裁剪以支持多像素标记
• Detect "input contains a filled rectangle of color C" tasks (57 candidates) · 识别"输入含纯色矩形"类任务
• Reduce double (1,10,30,30) intermediates in rot180 and flip via sparse / fused Gather · 借助稀疏或融合 Gather 降低 rot180 / flip 的双 (1,10,30,30) 中间张量内存
• Cache ONNX Runtime traces to speed up iteration · 缓存 ONNX Runtime trace 加速迭代

---

📜 License · 许可证

MIT © 2026 zikuanqi — see LICENSE · 详见 LICENSE