File size: 18,745 Bytes
a7e2535 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 | """
SOAR-style Program Synthesis for ARC-AGI
Evolutionary search: Sample programs → Execute → Refine → Vote
Based on:
- SOAR (Pourcel et al., 2507.14172)
- Product of Experts (Franzen et al., 2505.07859)
- CodeIt (Butt et al., 2402.04858)
"""
import copy
import json
import time
import random
import traceback
import signal
from typing import List, Dict, Tuple, Optional, Any
from collections import Counter, defaultdict
from concurrent.futures import ProcessPoolExecutor, TimeoutError as FuturesTimeout
import numpy as np
from arc_data import (
grids_equal, grid_to_string, grid_to_numpy, numpy_to_grid,
D8_TRANSFORMS, augment_task, reverse_d8, reverse_color_permutation,
create_color_permutation, apply_color_permutation
)
# ============================================================
# Safe code execution with timeout
# ============================================================
def _execute_transform(code: str, input_grid: List[List[int]], timeout_sec: float = 5.0) -> Optional[List[List[int]]]:
"""Execute a transform function safely with timeout."""
try:
namespace = {}
exec(code, namespace)
if "transform" not in namespace:
return None
transform_fn = namespace["transform"]
result = transform_fn(copy.deepcopy(input_grid))
# Validate output
if not isinstance(result, list):
return None
if len(result) == 0:
return None
for row in result:
if not isinstance(row, list):
return None
for cell in row:
if not isinstance(cell, (int, float)):
return None
if int(cell) < 0 or int(cell) > 9:
return None
# Normalize to int
result = [[int(c) for c in row] for row in result]
return result
except Exception:
return None
def evaluate_program_on_task(code: str, task: Dict) -> Tuple[float, Optional[List[List[int]]]]:
"""
Evaluate a program on all training pairs.
Returns (accuracy, test_output).
accuracy = fraction of training pairs correctly predicted.
"""
train_pairs = task["train"]
correct = 0
total = len(train_pairs)
for pair in train_pairs:
pred = _execute_transform(code, pair["input"])
if pred is not None and grids_equal(pred, pair["output"]):
correct += 1
accuracy = correct / total if total > 0 else 0.0
# Get test output
test_output = None
if task.get("test") and len(task["test"]) > 0:
test_output = _execute_transform(code, task["test"][0]["input"])
return accuracy, test_output
# ============================================================
# Program generation prompts
# ============================================================
SAMPLING_PROMPT = """You are an expert at solving abstract reasoning puzzles. Given input-output grid examples, write a Python function that transforms the input grid to the output grid.
Each grid is a 2D list of integers 0-9, where each integer represents a color:
0=black, 1=blue, 2=red, 3=green, 4=yellow, 5=grey, 6=magenta, 7=orange, 8=cyan, 9=maroon
{examples}
Write a Python function `transform(input_grid: list[list[int]]) -> list[list[int]]` that correctly transforms any input to its output. You may use numpy. Think step by step about the pattern, then write the code.
```python
{code_prefix}
```"""
REFINEMENT_PROMPT = """The following Python function was supposed to transform input grids to output grids, but it has errors.
{examples}
Previous attempt:
```python
{previous_code}
```
{error_feedback}
Please fix the function to correctly handle all examples. Write the corrected `transform` function:
```python
{code_prefix}
```"""
def format_examples_for_prompt(task: Dict) -> str:
"""Format task examples for the prompt."""
parts = []
for i, pair in enumerate(task["train"]):
parts.append(f"Example {i+1}:")
parts.append(f" Input ({len(pair['input'])}x{len(pair['input'][0])}):")
for row in pair["input"]:
parts.append(f" {row}")
parts.append(f" Output ({len(pair['output'])}x{len(pair['output'][0])}):")
for row in pair["output"]:
parts.append(f" {row}")
if task.get("test") and len(task["test"]) > 0:
test_inp = task["test"][0]["input"]
parts.append(f"\nTest Input ({len(test_inp)}x{len(test_inp[0])}):")
for row in test_inp:
parts.append(f" {row}")
return "\n".join(parts)
def get_error_feedback(code: str, task: Dict) -> str:
"""Generate error feedback for a failed program."""
feedback_parts = []
for i, pair in enumerate(task["train"]):
pred = _execute_transform(code, pair["input"])
if pred is None:
feedback_parts.append(f"Example {i+1}: EXECUTION ERROR")
elif not grids_equal(pred, pair["output"]):
feedback_parts.append(f"Example {i+1}: WRONG OUTPUT")
feedback_parts.append(f" Expected: {pair['output']}")
feedback_parts.append(f" Got: {pred}")
else:
feedback_parts.append(f"Example {i+1}: CORRECT")
return "\n".join(feedback_parts)
# ============================================================
# Program extraction from LLM output
# ============================================================
def extract_python_code(text: str) -> Optional[str]:
"""Extract Python code from LLM response."""
# Try to find code block
if "```python" in text:
start = text.index("```python") + len("```python")
end = text.find("```", start)
if end != -1:
code = text[start:end].strip()
if "def transform" in code:
return code
if "```" in text:
parts = text.split("```")
for i, part in enumerate(parts):
if i % 2 == 1: # Odd indices are code blocks
code = part.strip()
if code.startswith("python\n"):
code = code[7:]
if "def transform" in code:
return code
# Try to find the function definition directly
if "def transform" in text:
start = text.index("def transform")
# Find the end of the function (next def, or end of text)
lines = text[start:].split("\n")
func_lines = [lines[0]]
for line in lines[1:]:
if line.strip() and not line.startswith(" ") and not line.startswith("\t"):
if line.startswith("def ") or line.startswith("class "):
break
func_lines.append(line)
code = "\n".join(func_lines).strip()
return code
return None
# ============================================================
# Weighted Majority Voting (SOAR-style ensembling)
# ============================================================
def grid_to_hashable(grid: List[List[int]]) -> tuple:
"""Convert grid to hashable tuple for voting."""
return tuple(tuple(row) for row in grid)
def hashable_to_grid(h: tuple) -> List[List[int]]:
"""Convert hashable tuple back to grid."""
return [list(row) for row in h]
def weighted_majority_vote(programs: List[Tuple[str, float, Optional[List[List[int]]]]],
top_k: int = 2) -> List[List[List[int]]]:
"""
Weighted majority voting over program outputs.
Each program has (code, accuracy, test_output).
Score = sum of accuracies for programs producing the same output.
Returns top_k outputs.
"""
vote_scores = defaultdict(float)
vote_programs = defaultdict(list)
for code, accuracy, test_output in programs:
if test_output is None:
continue
key = grid_to_hashable(test_output)
vote_scores[key] += accuracy
vote_programs[key].append(code)
if not vote_scores:
return []
# Sort by score
sorted_votes = sorted(vote_scores.items(), key=lambda x: -x[1])
results = []
for key, score in sorted_votes[:top_k]:
results.append(hashable_to_grid(key))
return results
# ============================================================
# Program Synthesis Search (simplified SOAR)
# ============================================================
class ProgramSynthesisEngine:
"""
Evolutionary program synthesis engine for ARC tasks.
Uses an LLM to sample and refine Python programs.
"""
def __init__(self, model=None, tokenizer=None, max_samples: int = 50,
max_refinements: int = 50, temperature: float = 0.8):
self.model = model
self.tokenizer = tokenizer
self.max_samples = max_samples
self.max_refinements = max_refinements
self.temperature = temperature
self.programs_cache = {}
def generate_program(self, prompt: str, max_new_tokens: int = 1024) -> Optional[str]:
"""Generate a program from the LLM."""
if self.model is None or self.tokenizer is None:
return None
inputs = self.tokenizer(prompt, return_tensors="pt", truncation=True, max_length=4096)
inputs = {k: v.to(self.model.device) for k, v in inputs.items()}
with __import__('torch').no_grad():
outputs = self.model.generate(
**inputs,
max_new_tokens=max_new_tokens,
temperature=self.temperature,
top_p=0.95,
do_sample=True,
pad_token_id=self.tokenizer.eos_token_id,
)
text = self.tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
return extract_python_code(text)
def sample_programs(self, task: Dict, n_samples: int = None) -> List[Tuple[str, float, Optional[List[List[int]]]]]:
"""Sample programs for a task."""
if n_samples is None:
n_samples = self.max_samples
examples_text = format_examples_for_prompt(task)
results = []
for i in range(n_samples):
prompt = SAMPLING_PROMPT.format(
examples=examples_text,
code_prefix="import numpy as np\n\ndef transform(input_grid: list[list[int]]) -> list[list[int]]:"
)
code = self.generate_program(prompt)
if code is None:
continue
accuracy, test_output = evaluate_program_on_task(code, task)
results.append((code, accuracy, test_output))
# Early termination if we find a perfect program
if accuracy == 1.0:
break
return results
def refine_programs(self, task: Dict, programs: List[Tuple[str, float, Optional[List[List[int]]]]],
n_refinements: int = None) -> List[Tuple[str, float, Optional[List[List[int]]]]]:
"""Refine programs using execution feedback (REX bandit-style)."""
if n_refinements is None:
n_refinements = self.max_refinements
examples_text = format_examples_for_prompt(task)
all_programs = list(programs)
# Sort by accuracy (exploitation), but also explore low-accuracy ones
for i in range(n_refinements):
if not all_programs:
break
# REX-style selection: Thompson sampling with exploration
# Simplified: alternate between best (exploit) and random (explore)
if i % 2 == 0 and all_programs:
# Exploit: refine best program
parent = max(all_programs, key=lambda x: x[1])
else:
# Explore: refine random program
parent = random.choice(all_programs)
parent_code, parent_acc, _ = parent
if parent_acc == 1.0:
break # Already perfect
error_feedback = get_error_feedback(parent_code, task)
prompt = REFINEMENT_PROMPT.format(
examples=examples_text,
previous_code=parent_code,
error_feedback=error_feedback,
code_prefix="import numpy as np\n\ndef transform(input_grid: list[list[int]]) -> list[list[int]]:"
)
new_code = self.generate_program(prompt)
if new_code is None:
continue
accuracy, test_output = evaluate_program_on_task(new_code, task)
all_programs.append((new_code, accuracy, test_output))
if accuracy == 1.0:
break
return all_programs
def solve_task(self, task: Dict) -> List[List[List[int]]]:
"""
Full solve pipeline for a single ARC task:
1. Sample programs
2. Refine promising ones
3. Weighted majority vote → top-2 answers
"""
# Phase 1: Sample
programs = self.sample_programs(task)
# Phase 2: Refine
programs = self.refine_programs(task, programs)
# Filter to programs that get at least some training examples right
good_programs = [(c, a, o) for c, a, o in programs if a > 0 and o is not None]
if not good_programs:
# Fallback: use any program that produces output
good_programs = [(c, a, o) for c, a, o in programs if o is not None]
# Phase 3: Vote
predictions = weighted_majority_vote(good_programs, top_k=2)
return predictions
# ============================================================
# Handcrafted DSL primitives for common ARC patterns
# ============================================================
class ARCPrimitives:
"""Common ARC transformation primitives that can be composed."""
@staticmethod
def get_unique_colors(grid: List[List[int]]) -> set:
return set(c for row in grid for c in row)
@staticmethod
def get_background_color(grid: List[List[int]]) -> int:
"""Most common color is usually background."""
counter = Counter(c for row in grid for c in row)
return counter.most_common(1)[0][0]
@staticmethod
def find_objects(grid: List[List[int]], bg_color: int = 0) -> List[Dict]:
"""Find connected components (objects) in the grid."""
arr = grid_to_numpy(grid)
h, w = arr.shape
visited = np.zeros_like(arr, dtype=bool)
objects = []
def bfs(r, c):
color = arr[r, c]
cells = []
stack = [(r, c)]
while stack:
cr, cc = stack.pop()
if cr < 0 or cr >= h or cc < 0 or cc >= w:
continue
if visited[cr, cc] or arr[cr, cc] != color:
continue
visited[cr, cc] = True
cells.append((cr, cc))
for dr, dc in [(-1,0),(1,0),(0,-1),(0,1)]:
stack.append((cr+dr, cc+dc))
return cells
for r in range(h):
for c in range(w):
if not visited[r, c] and arr[r, c] != bg_color:
cells = bfs(r, c)
if cells:
min_r = min(cr for cr, _ in cells)
max_r = max(cr for cr, _ in cells)
min_c = min(cc for _, cc in cells)
max_c = max(cc for _, cc in cells)
objects.append({
"color": arr[r, c],
"cells": cells,
"bbox": (min_r, min_c, max_r, max_c),
"size": len(cells),
})
return objects
@staticmethod
def crop_grid(grid: List[List[int]], r1: int, c1: int, r2: int, c2: int) -> List[List[int]]:
"""Crop grid to bounding box."""
return [row[c1:c2+1] for row in grid[r1:r2+1]]
@staticmethod
def scale_grid(grid: List[List[int]], factor: int) -> List[List[int]]:
"""Scale grid by integer factor."""
result = []
for row in grid:
new_row = []
for c in row:
new_row.extend([c] * factor)
for _ in range(factor):
result.append(list(new_row))
return result
@staticmethod
def tile_grid(grid: List[List[int]], rows: int, cols: int) -> List[List[int]]:
"""Tile grid in a rows×cols pattern."""
h = len(grid)
w = len(grid[0]) if grid else 0
result = []
for tr in range(rows):
for r in range(h):
new_row = []
for tc in range(cols):
new_row.extend(grid[r])
result.append(new_row)
return result
# ============================================================
# Quick test
# ============================================================
if __name__ == "__main__":
from arc_data import load_arc_dataset_from_hf
print("Loading ARC tasks...")
tasks = load_arc_dataset_from_hf("arc-agi-community/arc-agi-2", "train")
# Test program evaluation
task = tasks[0]
print(f"\nTask 0: {len(task['train'])} demos")
for i, pair in enumerate(task["train"]):
print(f" Demo {i}: {len(pair['input'])}x{len(pair['input'][0])} -> {len(pair['output'])}x{len(pair['output'][0])}")
# Test a simple program
test_code = """
def transform(input_grid):
# Simple test: return the input as-is
return input_grid
"""
acc, out = evaluate_program_on_task(test_code, task)
print(f"\nIdentity program accuracy: {acc:.2f}")
# Test voting
programs = [
(test_code, 0.5, [[1,2],[3,4]]),
(test_code, 0.8, [[1,2],[3,4]]),
(test_code, 0.3, [[5,6],[7,8]]),
]
votes = weighted_majority_vote(programs, top_k=2)
print(f"Voting result: {len(votes)} candidates")
print(f" Top: {votes[0]}")
# Test primitives
prims = ARCPrimitives()
bg = prims.get_background_color(task["train"][0]["input"])
objects = prims.find_objects(task["train"][0]["input"], bg)
print(f"\nBackground color: {bg}")
print(f"Objects found: {len(objects)}")
for obj in objects[:3]:
print(f" Color {obj['color']}, size {obj['size']}, bbox {obj['bbox']}")
print("\n✅ Program synthesis module tests passed!")
|