Generate adversarial test suites that catch LLM code errors by analyzing constraint differences between correct and incorrect solutions. SAGA framework improves test detection rate 9.55% and verifier accuracy 12.14% by using human bug patterns and multi-input differential analysis.
Generate adversarial test suites that catch LLM code errors by analyzing constraint differences between correct and incorrect solutions. SAGA framework improves test detection rate 9.55% and verifier accuracy 12.14% by using human bug patterns and multi-input differential analysis.
Adversarial Code Verification: Generating Discriminative Test Cases from Solution Patterns
Current benchmarks for LLM code generation use limited, homogeneous test suites that inflate performance metrics and fail to catch real errors. A model scoring 95% on standard benchmarks may fail 50% of real submissions because test suites are too narrow, allowing solutions with subtle bugs to pass. SAGA solves this through strategic analysis: first, extract constraint insights from correct solutions to identify edge cases; second, analyze patterns in human bug submissions to create targeted test cases; third, generate diverse inputs exploring different error classes. The result is test suites that reliably distinguish correct from incorrect code, giving RL systems honest reward signals.
When training code generation models with RL, weak test suites corrupt the reward signal—models learn to pass bad evaluations rather than produce correct code. SAGA test suites catch bugs that simple unit tests miss: off-by-one errors in loops, incorrect boundary handling, type confusion, and logic inversions. The human-LLM collaborative approach combines human insight about what makes problems hard with LLM efficiency at generating diverse test cases.
Core Concept
SAGA generates adversarial tests through three coordinated phases. First, multidimensional analysis examines correct solutions to extract constraint patterns: what boundary conditions do they handle, what invariants do they maintain? Second, differential analysis compares failed solutions with corrected versions to identify common error patterns—these patterns become targets for test generation. Third, dual-input strategy combines both ground-truth and buggy solutions to produce tests that distinguish right from wrong. The framework introduces new evaluation metrics beyond simple accuracy: detection rate (what proportion of known bugs does the test suite find), verifier accuracy (does the suite correctly identify all incorrect solutions), and distinct error pattern coverage (breadth of error types detected).
Architecture Overview
Multidimensional Analyzer: Extracts constraint patterns, boundary conditions, and invariants from correct solutions
Differential Analyzer: Compares correct vs. buggy submissions to identify error patterns
Constraint-Differential Test Generator: Creates test cases targeting specific constraint violations and error patterns
Test Discriminator: Validates that generated tests actually distinguish correct from incorrect code
Human-LLM Collaborative Pipeline: Combines human problem insight with LLM test generation at scale
Implementation
This example demonstrates extracting constraint patterns from correct solutions to guide test generation.
# Constraint-based adversarial test generationimport ast
import torch
classConstraintAnalyzer:
def__init__(self):
self.constraints = {}
self.boundaries = {}
defanalyze_correct_solution(self, solution_code, problem_description):
"""Extract constraints and boundary conditions from correct code."""# Parse solution AST
tree = ast.parse(solution_code)
# Extract constraint patterns
constraints = {
'loop_bounds': [],
'conditionals': [],
'type_checks': [],
'boundary_values': []
}
# Analyze loops: what are iteration bounds?for node in ast.walk(tree):
ifisinstance(node, ast.For):
# Extract loop rangeifisinstance(node.iter, ast.Call):
ifisinstance(node.iter.func, ast.Name) and node.iter.func.id == 'range':
# range(N), range(0, N), range(0, N, 1)
bounds = self._extract_range_bounds(node.iter)
constraints['loop_bounds'].append(bounds)
elifisinstance(node, ast.If):
# Extract condition patterns
condition = ast.unparse(node.test)
constraints['conditionals'].append(condition)
elifisinstance(node, ast.Compare):
# Extract comparison patterns (boundary checks)iflen(node.ops) > 0:
op_type = type(node.ops[0]).__name__
constraints['boundary_values'].append({
'operator': op_type,
'condition': ast.unparse(node)
})
return constraints
defextract_boundary_test_cases(self, constraints, variable_bounds):
"""Generate boundary test cases from extracted constraints."""
test_cases = []
# For each boundary condition, create test cases at boundariesfor boundary in constraints['boundary_values']:
# Extract the compared valueif'N'in boundary['condition']:
# Boundary around N: test N-1, N, N+1for n in variable_bounds:
test_cases.append(n - 1if n > 0else0)
test_cases.append(n)
test_cases.append(n + 1)
# Test zero and large values
test_cases.extend([0, 1, variable_bounds[-1], variable_bounds[-1] + 1])
returnlist(set(test_cases)) # Remove duplicates
This example shows differential analysis: comparing correct and incorrect solutions to identify bug patterns.
classDifferentialBugAnalyzer:
def__init__(self):
self.error_patterns = {}
defanalyze_bug_difference(self, correct_code, buggy_code, test_input, expected_output):
"""Identify what makes the buggy code fail on this input."""# Execute both versionstry:
correct_output = execute_solution(correct_code, test_input)
except:
returnNonetry:
buggy_output = execute_solution(buggy_code, test_input)
except Exception as e:
return {
'error_type': 'runtime_error',
'error_msg': str(e),
'test_input': test_input
}
# Compare outputsif correct_output != buggy_output:
# Analyze AST differences
correct_ast = ast.parse(correct_code)
buggy_ast = ast.parse(buggy_code)
diffs = self._compute_ast_differences(correct_ast, buggy_ast)
return {
'error_type': 'wrong_output',
'test_input': test_input,
'expected': expected_output,
'actual': buggy_output,
'code_differences': diffs,
'error_pattern': self._classify_error_pattern(diffs)
}
returnNonedef_classify_error_pattern(self, code_diffs):
"""Classify bug type from code differences."""ifany('off-by-one'in diff for diff in code_diffs):
return'off_by_one'elifany('<='in diff or'>='in diff for diff in code_diffs):
return'boundary_condition'elifany('not'in diff or'!'in diff for diff in code_diffs):
return'logic_inversion'elifany('int'in diff or'float'in diff for diff in code_diffs):
return'type_error'else:
return'other'defcollect_error_patterns(self, correct_solutions, buggy_submissions, test_suite):
"""Analyze all buggy submissions to build error pattern library."""
error_patterns = {}
for buggy_code in buggy_submissions:
for correct_code in correct_solutions:
for test_input, expected_output in test_suite:
bug_info = self.analyze_bug_difference(
correct_code, buggy_code, test_input, expected_output
)
if bug_info:
pattern = bug_info['error_pattern']
if pattern notin error_patterns:
error_patterns[pattern] = []
error_patterns[pattern].append(bug_info)
return error_patterns
This example demonstrates SAGA test generation combining constraints and error patterns.
classSAGATestGenerator:
def__init__(self, constraint_analyzer, bug_analyzer):
self.constraint_analyzer = constraint_analyzer
self.bug_analyzer = bug_analyzer
defgenerate_adversarial_tests(self, correct_solutions, buggy_submissions, problem_spec):
"""Generate adversarial test suite using SAGA framework."""
test_suite = []
# Phase 1: Extract constraints from correct solutions
all_constraints = []
all_boundaries = []
for correct_code in correct_solutions:
constraints = self.constraint_analyzer.analyze_correct_solution(
correct_code, problem_spec
)
all_constraints.append(constraints)
boundaries = self.constraint_analyzer.extract_boundary_test_cases(
constraints, [1, 10, 100, 1000]
)
all_boundaries.extend(boundaries)
# Phase 2: Analyze error patterns in buggy submissions
error_patterns = self.bug_analyzer.collect_error_patterns(
correct_solutions, buggy_submissions, test_suite
)
# Phase 3: Generate tests targeting each error patternfor pattern_type, bug_examples in error_patterns.items():
# What inputs trigger this pattern?
triggering_inputs = [example['test_input'] for example in bug_examples]
# Generate variations on triggering inputs
generated_tests = self._mutate_inputs(triggering_inputs, pattern_type)
test_suite.extend(generated_tests)
# Phase 4: Add boundary tests
boundary_tests = self._generate_boundary_tests(all_boundaries)
test_suite.extend(boundary_tests)
return test_suite
def_mutate_inputs(self, base_inputs, error_pattern):
"""Create input variations targeting specific error patterns."""
mutations = []
for base_input in base_inputs:
# For off-by-one errors: test N-1, N, N+1if error_pattern == 'off_by_one':
ifisinstance(base_input, int):
mutations.extend([base_input - 1, base_input, base_input + 1])
elifisinstance(base_input, list):
mutations.append(base_input[:-1]) # Shorter list
mutations.append(base_input + [0]) # Longer list# For boundary errors: test at limitselif error_pattern == 'boundary_condition':
mutations.append(0)
mutations.append(1)
mutations.append(len(base_input) - 1ifisinstance(base_input, (list, str)) else base_input)
# For logic inversion: complementary caseselif error_pattern == 'logic_inversion':
mutations.append(not base_input ifisinstance(base_input, bool) else -base_input)
return mutations
defcompute_test_quality_metrics(self, test_suite, correct_solutions, buggy_submissions):
"""Compute detection rate and verifier accuracy."""
detection_rate = 0.0
verifier_accuracy = 0.0
bugs_detected = 0
total_bugs = len(buggy_submissions)
for buggy_code in buggy_submissions:
for test_input, expected_output in test_suite:
try:
actual_output = execute_solution(buggy_code, test_input)
if actual_output != expected_output:
bugs_detected += 1break# This test catches this bugexcept:
bugs_detected += 1break
detection_rate = bugs_detected / total_bugs if total_bugs > 0else1.0# Verifier accuracy: correct solutions pass all tests
correct_pass_all = 0for correct_code in correct_solutions:
ifall(
execute_solution(correct_code, test_input) == expected_output
for test_input, expected_output in test_suite
):
correct_pass_all += 1
verifier_accuracy = correct_pass_all / len(correct_solutions)
return {
'detection_rate': detection_rate,
'verifier_accuracy': verifier_accuracy,
'bugs_detected': bugs_detected,
'total_bugs': total_bugs
}
Practical Guidance
Hyperparameter
Recommended Value
Purpose
Boundary offset range
[-2, -1, 0, +1, +2]
Comprehensive boundary testing
Max test cases per error pattern
50
Balance coverage vs. test suite bloat
Constraint extraction depth
3 iterations
Find nested constraints
Error pattern categories
8+ types
Broad error coverage
Detection rate target
> 95%
Catch most buggy submissions
Verifier accuracy target
> 98%
Minimize false positives
Buggy submission sample size
100+
Representative error patterns
When to use: Apply SAGA when evaluating code generation models with RL, where test quality directly affects learned behavior. Use when you have access to buggy human submissions (competitive programming sites) or generated solutions to analyze. Ideal for creating reliable evaluation benchmarks and improving RL reward signals.
When NOT to use: Skip if you only have correct reference solutions without buggy examples—differential analysis requires both. Avoid if test generation latency matters (SAGA is compute-intensive). Don't use for simple, well-defined problems where boundary testing suffices. Skip if your problem domain lacks clear error patterns (e.g., creative writing evaluation).
Common pitfalls: Using only boundary cases misses logic inversion and type errors. Not analyzing enough buggy submissions leads to sparse error pattern coverage. Generating tests without validation allows degenerate cases that don't actually distinguish correct from incorrect code. Over-weighting boundary tests at expense of pattern-based tests reduces real-world error detection. Forgetting to verify that correct solutions pass the test suite before using it for evaluation. Not computing detection rate separately for different error pattern types hides coverage gaps.
Reference
Code Verification Team. (2025). Rethinking Verification for LLM Code Generation: From Generation to Testing. arXiv preprint arXiv:2507.06920. https://arxiv.org/abs/2507.06920
