Improve multimodal mathematical reasoning through iterative reflection cycles where an outcome reward model provides feedback on reasoning quality, and correct solutions are incorporated back into training—enabling continuous model adaptation beyond static datasets.
Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Improve multimodal mathematical reasoning through iterative reflection cycles where an outcome reward model provides feedback on reasoning quality, and correct solutions are incorporated back into training—enabling continuous model adaptation beyond static datasets.
Iteratively Improve Mathematical Reasoning Through Self-Reflection
Traditional mathematical training relies on static datasets of teacher-generated solutions, which capture only fixed reasoning patterns. MathSE enables continuous improvement through iterative refinement: models generate solutions, an outcome reward model (ORM) evaluates reasoning quality, and successful reasoning paths are fed back into training. This creates a virtuous cycle where the model adapts to progressively more difficult problems.
By treating mathematical reasoning as self-evolving rather than static, models develop robust problem-solving that generalizes beyond training distribution.
Core Concept
MathSE implements a closed-loop learning cycle:
Inference - Model generates solutions to problems (including harder ones beyond training)
Reflection - Outcome Reward Model rates solution quality and reasoning soundness
Refinement - Correct reasoning paths from this iteration feed back into training
Iteration - Process repeats, exposing model to progressively harder problems
This approach contrasts sharply with static datasets which distill only teacher reasoning patterns. Through self-evolution, the model continuously discovers new solution strategies and generalizes to novel problem structures.
Architecture Overview
Base Model: Multimodal model (vision + text) for mathematical reasoning
Outcome Reward Model (ORM): Evaluates correctness and reasoning quality
Solution Generator: Creates candidate solutions for problems
"""
Evaluate solution for given problem.
Args:
problem_embedding: Encoded problem [batch, embedding_dim]
solution_embedding: Encoded solution [batch, embedding_dim]
Returns:
correctness_score: Expected correctness [batch, 1]
quality_score: Reasoning quality [batch, 1]
"""
# Encode
self
self
# Combine features
1
# Predict scores
self
self
return
def
train_orm
orm: OutcomeRewardModel, train_pairs, num_epochs: int = 10
"""
Train outcome reward model on labeled solution pairs.
Args:
orm: OutcomeRewardModel instance
train_pairs: List of (problem, solution, label) tuples
where label is {correct: bool, quality: float}
num_epochs: Training epochs
"""
1e-4
for
in
range
0
for
in
# Forward pass
0
0
# Loss on correctness
1.0
if
'correct'
else
0.0
# Loss on reasoning quality
'quality'
0.5
0.5
print
f"ORM Epoch {epoch}: Loss {total_loss / len(train_pairs):.4f}"
Step 2: Solution Generation and Reflection
Generate solutions and reflect on their correctness.
classSolutionGenerator:
"""
Generates mathematical solutions and evaluates them.
"""def__init__(self, model, orm: OutcomeRewardModel, max_tokens: int = 1024):
"""
Args:
model: Multimodal model for solution generation
orm: Outcome Reward Model for evaluation
max_tokens: Maximum tokens in generated solution
"""self.model = model
self.orm = orm
self.max_tokens = max_tokens
defgenerate_solution(self, problem: str, num_attempts: int = 1) -> List[Dict]:
"""
Generate multiple solution attempts for a problem.
Args:
problem: Problem description
num_attempts: Number of solution attempts
Returns:
solutions: List of {solution_text, reasoning, score}
"""
solutions = []
for attempt inrange(num_attempts):
# Generate solution with sampling for diversity
prompt = f"""Solve this mathematical problem step-by-step:
{problem}
Provide a detailed solution with all intermediate steps:"""
solution_text = self.model.generate(
prompt, max_tokens=self.max_tokens,
temperature=0.7if attempt > 0else0.5# Deterministic first attempt
)
solutions.append({
'solution_text': solution_text,
'attempt': attempt
})
return solutions
defevaluate_solutions(self, problem: str, solutions: List[Dict],
ground_truth: str = None) -> List[Dict]:
"""
Evaluate solutions using ORM and ground truth comparison.
Args:
problem: Original problem
solutions: Generated solutions
ground_truth: Correct answer (if available)
Returns:
evaluated: Solutions with scores and labels
"""# Encode problem once
problem_embedding = self.model.encode(problem)
evaluated = []
for solution in solutions:
# Encode solution
solution_embedding = self.model.encode(solution['solution_text'])
# Get ORM scoreswith torch.no_grad():
correctness, quality = self.orm(
problem_embedding.unsqueeze(0),
solution_embedding.unsqueeze(0)
)
correctness_score = correctness.item()
quality_score = quality.item()
# Verify against ground truth if available
is_correct = Falseif ground_truth:
is_correct = self._check_correctness(
solution['solution_text'], ground_truth
)
evaluated.append({
**solution,
'orm_correctness': correctness_score,
'orm_quality': quality_score,
'verified_correct': is_correct,
'combined_score': 0.7 * correctness_score + 0.3 * quality_score
})
return evaluated
def_check_correctness(self, solution_text: str, ground_truth: str) -> bool:
"""Check if solution matches ground truth."""# Extract final answer from solution
lines = solution_text.split('\n')
final_answer = lines[-1].strip() if lines else''# Simple matchingif final_answer == ground_truth:
returnTrue# Numeric matching (if both are numbers)try:
sol_num = float(final_answer)
truth_num = float(ground_truth)
returnabs(sol_num - truth_num) < 1e-6except:
returnFalse
Step 3: Trajectory Replay Buffer
Maintain collection of successful solutions for replay during training.
from collections import deque
import json
classTrajectoryBuffer:
"""
Stores successful solution trajectories for iterative training.
"""def__init__(self, capacity: int = 10000, quality_threshold: float = 0.7):
"""
Args:
capacity: Maximum stored trajectories
quality_threshold: Minimum quality to store
"""self.buffer = deque(maxlen=capacity)
self.quality_threshold = quality_threshold
defadd_trajectory(self, problem: str, solution: str, score: float,
is_correct: bool, difficulty: str = 'medium'):
"""
Add successful trajectory to buffer.
Args:
problem: Problem description
solution: Complete solution text
score: ORM-computed quality score
is_correct: Whether answer is verified correct
difficulty: Problem difficulty estimate
"""if score >= self.quality_threshold or is_correct:
trajectory = {
'problem': problem,
'solution': solution,
'score': score,
'correct': is_correct,
'difficulty': difficulty,
'timestamp': time.time()
}
self.buffer.append(trajectory)
defsample_batch(self, batch_size: int) -> List[Dict]:
"""
Sample trajectory batch, emphasizing recent and high-quality solutions.
Args:
batch_size: Batch size
Returns:
batch: Sampled trajectories
"""iflen(self.buffer) == 0:
return []
# Weight by recency and quality
trajectories = list(self.buffer)
weights = []
now = time.time()
for traj in trajectories:
# Recency factor: newer solutions weighted higher
age = now - traj['timestamp']
recency = 1.0 / (1.0 + age / 3600) # Half-life: 1 hour# Quality factor
quality = traj['score']
if traj['correct']:
quality = 1.0
weight = recency * quality
weights.append(weight)
# Sample according to weightsimport random
total_weight = sum(weights)
normalized_weights = [w / total_weight for w in weights]
batch = random.choices(trajectories, weights=normalized_weights,
k=min(batch_size, len(trajectories)))
return batch
defget_statistics(self) -> Dict:
"""Compute buffer statistics."""iflen(self.buffer) == 0:
return {'size': 0}
trajectories = list(self.buffer)
correct_count = sum(1for t in trajectories if t['correct'])
return {
'size': len(trajectories),
'correct_count': correct_count,
'correct_ratio': correct_count / len(trajectories),
'avg_score': sum(t['score'] for t in trajectories) / len(trajectories)
}
Step 4: Iterative Training Loop
Implement the main self-evolving training cycle.
defself_evolving_training(base_model, orm: OutcomeRewardModel,
problem_pool: List[str], num_iterations: int = 5):
"""
Main training loop for mathematical reasoning self-evolution.
Args:
base_model: Multimodal model to train
orm: Trained outcome reward model
problem_pool: Pool of problems to solve
num_iterations: Number of self-evolution iterations
"""
generator = SolutionGenerator(base_model, orm)
buffer = TrajectoryBuffer(capacity=5000)
for iteration inrange(num_iterations):
print(f"\n=== Self-Evolution Iteration {iteration + 1} ===")
# Sample problems (mix of training and harder problems)
num_sample = int(len(problem_pool) * 0.7)
sampled_problems = random.sample(problem_pool, num_sample)
# Add harder problems beyond training distributionif iteration > 0:
harder_problems = problem_pool[num_sample:]
sampled_problems.extend(harder_problems[:len(sampled_problems) // 4])
# Generate and evaluate solutionsfor problem in sampled_problems:
# Generate solution attempts
solutions = generator.generate_solution(problem, num_attempts=2)
# Evaluate with ORM
evaluated = generator.evaluate_solutions(problem, solutions)
# Store best solutions
best = max(evaluated, key=lambda x: x['combined_score'])
buffer.add_trajectory(
problem, best['solution_text'],
best['combined_score'], best['verified_correct']
)
# Print iteration statistics
stats = buffer.get_statistics()
print(f"Buffer size: {stats['size']}")
print(f"Correct ratio: {stats['correct_ratio']:.2%}")
print(f"Avg score: {stats['avg_score']:.3f}")
# Fine-tune on collected trajectoriesprint("Fine-tuning on successful trajectories...")
batch = buffer.sample_batch(batch_size=32)
optimizer = torch.optim.Adam(base_model.parameters(), lr=1e-5)
for epoch inrange(2):
total_loss = 0for trajectory in batch:
prompt = f"Problem: {trajectory['problem']}\n\nSolution:"
logits = base_model.forward(prompt)
# Supervise on ground truth solution
target_ids = base_model.tokenize(trajectory['solution'])
loss = compute_language_loss(logits, target_ids)
optimizer.zero_grad()
loss.backward()
optimizer.step()
total_loss += loss.item()
print(f" Epoch {epoch}: Loss {total_loss / len(batch):.4f}")
return base_model
