Skip to main content Accueil Créateurs adu2021 skillxiv svs-variational-problem-synthesis
svs-variational-problem-synthesis Generate problem variants from correct model solutions while preserving answer equivalence, enabling self-play training that maintains output diversity and prevents entropy collapse.
Aller à l'installation Skills Marketplace Découvrez et explorez les compétences IA créées par la communauté.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Afficher les détails du prompt Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
npx skills add https://github.com/ADu2021/skillXiv --skill svs-variational-problem-synthesisLa commande reste sur une seule ligne. Faites défiler horizontalement pour la vérifier avant de la copier.
Vous préférez une copie locale ? Téléchargez les fichiers actuellement disponibles dans SkillsMP.
Télécharger Zip Téléchargement... Plus depuis ce dépôt meaningful-kebab-case-name Convert arXiv papers into ready-to-use agent skills using category-aware extraction. First classifies the paper into one or more of 11 research categories, then applies a specialized extraction pipeline for each category — because different types of papers produce different types of usable knowledge. A single paper can yield multiple skills if it spans categories. Use this skill whenever the user wants to turn a paper into a skill, extract practical techniques from research, build a skill library from papers, convert arXiv papers into reusable agent instructions, or batch-process multiple papers into skills. Also trigger when someone asks about extracting actionable knowledge from papers, making research practical for LLM agents, or systematically converting academic contributions into structured agent capabilities.
action-quantization-behavior-cloning Establish regret bounds for behavior cloning with discretized actions combining statistical error and quantization error terms. Prove smoothness requirements for safe quantizer design, show that learning-based quantizers fail these requirements, and propose model-based augmentation to reduce error dependence from H² to H.
adaptive-lora-personalized-ranks Dynamically allocate LoRA ranks per-layer during fine-tuning instead of using fixed uniform ranks. Learn optimal rank for each layer and subject via variational framework with discretized exponential distribution, reducing memory footprint while maintaining fidelity and text-alignment.
Métiers associés SOC
Basé sur la classification professionnelle SOC
name svs-variational-problem-synthesis title Self-Play with Variational Problem Synthesis for Sustained Reasoning Diversity version 0.0.2 engine skillxiv-v0.0.2-claude-opus-4.6 license MIT url https://arxiv.org/abs/2508.14029 keywords ["reinforcement-learning","problem-synthesis","entropy-preservation","diversity","self-play"] description Generate problem variants from correct model solutions while preserving answer equivalence, enabling self-play training that maintains output diversity and prevents entropy collapse.
Self-Play with Variational Problem Synthesis (SvS)
Core Concept
SvS addresses entropy collapse in reasoning RL by synthesizing problem variants derived from correct model solutions. The key insight is that while Pass@1 improves with RL training, Pass@k degrades due to reduced output diversity. By generating new problems from correct solutions (maintaining identical answers), SvS sustains policy entropy during training. This approach achieves 18-22% improvements on AIME benchmarks while preserving the model's ability to generate diverse high-quality solutions.
Architecture Overview
Correct Solution Extraction : Identify and isolate correct reasoning stepsInvariant Answer Identification : Determine what must remain constant across variantsProblem Variant Generation : Create mathematically equivalent problemsSelf-Play Training Loop : Iteratively improve on synthetic problemsEntropy Preservation : Monitor and maintain output diversity metrics
Implementation Steps
1. Extract and Analyze Correct Solutions
Identify successful reasoning traces for variant generation:
from typing import List , Dict , Tuple , Optional
from dataclasses import dataclass
@dataclass
class Solution :
"""Representation of a model solution."""
problem: str
reasoning_steps: List [str ]
final_answer: str
intermediate_values: Dict [str , any ]
confidence: float
is_correct: bool
class SolutionAnalyzer :
def __init__ (self, verifier: "AnswerVerifier" ):
self
def
extract_correct_solutions
self,
samples: List [Tuple [str , str , str ]],
num_best: int = 5
List
"""
Extract solutions verified as correct.
"""
for
in
self
if
self
self
1.0
if
else
0.0
lambda
True
return
def
_parse_steps
self, response: str
List
str
"""Extract reasoning steps from response."""
import
r'Step|Therefore|Thus|So'
return
for
in
if
def
_extract_values
self, response: str
Dict
str
any
"""Extract intermediate numerical values."""
import
r'(\w+)\s*=\s*([\d\.\-]+)'
return
float
for
in
2. Implement Problem Variant Generation Create mathematically equivalent problems from solutions:
class ProblemVariantGenerator :
"""Generate problem variants maintaining answer equivalence."""
def __init__ (self, problem_template_library: Dict [str , List [str ]] ):
self .templates = problem_template_library
def generate_variants (
self,
solution: Solution,
num_variants: int = 3 ,
variant_type: str = "substitution"
) -> List [Tuple [str , str ]]:
"""
Generate problem variants from correct solution.
Returns: [(new_problem, expected_answer), ...]
"""
variants = []
if variant_type == "substitution" :
for i in range (num_variants):
new_problem = self ._substitute_values(
solution.problem,
solution.intermediate_values,
variant_index=i
)
variants.append((new_problem, solution.final_answer))
elif variant_type == "reformulation" :
for i in range (num_variants):
new_problem = self ._reformulate_problem(
solution.problem,
solution.reasoning_steps,
variant_index=i
)
variants.append((new_problem, solution.final_answer))
elif variant_type == "context_swap" :
for i in range (num_variants):
new_problem = self ._swap_context(
solution.problem,
solution.final_answer,
variant_index=i
)
variants.append((new_problem, solution.final_answer))
return variants
def _substitute_values (
self,
original_problem: str ,
known_values: Dict [str , float ],
variant_index: int = 0
) -> str :
"""
Create variant by substituting different numbers.
Key: keep mathematical relationships intact.
"""
import re
problem = original_problem
seed = variant_index
for var_name, original_val in known_values.items():
if isinstance (original_val, float ) and original_val != 0 :
scale = 1.0 + (seed * 0.1 ) % 0.5
new_val = original_val * scale
pattern = re.escape(str (int (original_val) if original_val == int (original_val) else original_val))
problem = re.sub(pattern, str (int (new_val) if new_val == int (new_val) else new_val), problem)
return problem
def _reformulate_problem (
self,
problem: str ,
reasoning_steps: List [str ],
variant_index: int = 0
) -> str :
"""
Reformulate problem preserving mathematical structure.
Example: "Find x such that..." -> "Solve for x where..."
"""
reformulations = {
"find" : "calculate" ,
"what is" : "determine" ,
"compute" : "evaluate" ,
"how many" : "count the number of" ,
"if" : "suppose" ,
}
problem_lower = problem.lower()
for original, replacement in reformulations.items():
if original in problem_lower:
problem = problem.replace(original, replacement)
break
return problem
def _swap_context (
self,
original_problem: str ,
answer: str ,
variant_index: int = 0
) -> str :
"""
Create variant with different real-world context.
"""
contexts = [
"In a classroom with" ,
"At a store with" ,
"During a game with" ,
"In a garden with" ,
"For a project with" ,
]
import re
numbers = re.findall(r'\d+' , original_problem)
if variant_index < len (contexts):
context = contexts[variant_index]
else :
context = contexts[0 ]
new_problem = f"{context} {' and ' .join(numbers)} items..."
return new_problem
3. Implement Self-Play Training Loop Train model on generated variants:
class SelfPlayTrainer :
"""Trains model on self-generated problem variants."""
def __init__ (
self,
model: "LLM" ,
variant_generator: ProblemVariantGenerator,
solution_analyzer: SolutionAnalyzer,
reward_model: "RewardModel"
):
self .model = model
self .variant_generator = variant_generator
self .analyzer = solution_analyzer
self .reward_model = reward_model
self .training_history = []
def self_play_iteration (
self,
base_problems: List [str ],
num_solutions_per_problem: int = 8 ,
num_variants_per_solution: int = 2 ,
num_training_steps: int = 100
) -> Dict [str , float ]:
"""
Execute single self-play iteration.
"""
iteration_metrics = {}
print ("Generating solutions on base problems..." )
base_solutions = []
for problem in base_problems:
solutions = self ._generate_n_solutions(
problem,
num_solutions_per_problem
)
base_solutions.extend(solutions)
print ("Extracting correct solutions..." )
correct_solutions = self .analyzer.extract_correct_solutions(
[(s.problem, s.reasoning_steps, s.final_answer) for s in base_solutions],
num_best=len (base_solutions) // 2
)
iteration_metrics["num_correct" ] = len (correct_solutions)
iteration_metrics["base_accuracy" ] = len (correct_solutions) / len (base_solutions)
print ("Generating problem variants..." )
variant_dataset = []
for solution in correct_solutions:
variants = self .variant_generator.generate_variants(
solution,
num_variants=num_variants_per_solution,
variant_type="substitution"
)
variant_dataset.extend(variants)
print (f"Training on {len (variant_dataset)} variants..." )
variant_loss = self ._train_on_variants(
variant_dataset,
num_steps=num_training_steps
)
iteration_metrics["variant_loss" ] = variant_loss
print ("Measuring output diversity..." )
base_entropy, variant_entropy = self ._measure_entropy(
base_problems,
num_solutions_per_problem
)
iteration_metrics["base_entropy" ] = base_entropy
iteration_metrics["variant_entropy" ] = variant_entropy
iteration_metrics["entropy_preservation" ] = variant_entropy / (base_entropy + 1e-8 )
self .training_history.append(iteration_metrics)
return iteration_metrics
def _generate_n_solutions (
self,
problem: str ,
n: int = 8
) -> List [Solution]:
"""Generate n diverse solutions for a problem."""
solutions = []
for i in range (n):
response = self .model.generate(
problem,
temperature=0.7 + (i % 3 ) * 0.1 ,
max_tokens=500
)
solution = Solution(
problem=problem,
reasoning_steps=self .analyzer._parse_steps(response),
final_answer=self ._extract_answer(response),
intermediate_values=self .analyzer._extract_values(response),
confidence=0.5 ,
is_correct=False
)
solutions.append(solution)
return solutions
def _train_on_variants (
self,
variant_dataset: List [Tuple [str , str ]],
num_steps: int = 100 ,
batch_size: int = 4
) -> float :
"""
Train model on variant dataset using RL.
"""
total_loss = 0.0
for step in range (num_steps):
batch_variants = variant_dataset[
(step * batch_size) % len (variant_dataset):
((step + 1 ) * batch_size) % len (variant_dataset)
]
batch_loss = 0.0
for problem, expected_answer in batch_variants:
response = self .model.generate(problem, max_tokens=500 )
predicted_answer = self ._extract_answer(response)
reward = self .reward_model.compute_reward(
problem,
response,
expected_answer
)
loss = -reward * self .model.log_prob(response)
batch_loss += loss.item()
avg_batch_loss = batch_loss / len (batch_variants)
total_loss += avg_batch_loss
self .model.optimizer.zero_grad()
avg_batch_loss.backward()
torch.nn.utils.clip_grad_norm_(self .model.parameters(), 1.0 )
self .model.optimizer.step()
return total_loss / num_steps
def _measure_entropy (
self,
problems: List [str ],
num_solutions: int = 8
) -> Tuple [float , float ]:
"""
Measure output diversity (Shannon entropy of solution tokens).
"""
import torch
from scipy.stats import entropy
all_solutions = []
for problem in problems:
for _ in range (num_solutions):
response = self .model.generate(problem, max_tokens=300 )
all_solutions.append(response)
token_counts = {}
for solution in all_solutions:
tokens = solution.split()
for token in set (tokens):
token_counts[token] = token_counts.get(token, 0 ) + 1
token_probs = list (token_counts.values())
token_probs = [p / sum (token_probs) for p in token_probs]
avg_entropy = entropy(token_probs)
return avg_entropy, avg_entropy
def _extract_answer (self, response: str ) -> str :
"""Extract final answer from response."""
import re
numbers = re.findall(r'\d+(?:\.\d+)?' , response)
return numbers[-1 ] if numbers else response.split()[-1 ]
4. Track and Monitor Entropy Collapse Monitor diversity metrics:
class EntropyMonitor :
"""Monitor output diversity and entropy metrics."""
def __init__ (self, check_interval: int = 10 ):
self .check_interval = check_interval
self .entropy_history = []
self .pass_k_history = []
def compute_pass_k (
self,
problems: List [str ],
model: "LLM" ,
k_values: List [int ] = [1 , 3 , 5 ],
num_generations: int = 32
) -> Dict [int , float ]:
"""
Compute Pass@k metric (k solutions needed for one correct).
"""
pass_k_scores = {k: 0.0 for k in k_values}
for problem in problems:
solutions = []
for _ in range (num_generations):
response = model.generate(problem, max_tokens=500 )
solutions.append(response)
correct_mask = [self ._verify_solution(s, problem) for s in solutions]
for k in k_values:
if any (correct_mask[:k]):
pass_k_scores[k] += 1.0 / len (problems)
return pass_k_scores
def compute_output_entropy (
self,
problems: List [str ],
model: "LLM" ,
num_generations: int = 8
) -> float :
"""
Compute Shannon entropy of outputs.
High entropy = diverse outputs
Low entropy = repetitive outputs
"""
from scipy.stats import entropy as scipy_entropy
all_solutions = []
for problem in problems:
for _ in range (num_generations):
response = model.generate(problem, max_tokens=300 )
all_solutions.append(response)
token_freq = {}
for solution in all_solutions:
for token in set (solution.split()):
token_freq[token] = token_freq.get(token, 0 ) + 1
probs = list (token_freq.values())
probs = [p / sum (probs) for p in probs]
return scipy_entropy(probs)
def check_entropy_collapse (self, current_entropy: float ) -> bool :
"""Detect if entropy is dropping (collapse pattern)."""
if len (self .entropy_history) < 2 :
return False
recent_entropy = self .entropy_history[-1 ]
entropy_drop = recent_entropy - current_entropy
return entropy_drop > 0.1 * recent_entropy
def _verify_solution (self, solution: str , problem: str ) -> bool :
"""Verify if solution is correct."""
pass
Practical Guidance
When to Use SvS
Mathematical reasoning benchmarks (AIME, high school math)
Tasks with clear correct/incorrect answers
Scenarios where output diversity matters (Pass@k evaluation)
Self-play learning scenarios
Models where entropy collapse is observed
When NOT to Use
Creative generation (poetry, stories)
Tasks without canonical correct answers
Low-diversity tasks where Pass@1 is sole metric
Real-time inference (variant generation adds latency)
Key Hyperparameters
num_variants_per_solution : 2-5 (more = better coverage)variant_generation_type : "substitution" recommended for mathentropy_threshold : 10-20% drop before interventionself_play_iterations : 3-10 per model sizetemperature_variation : 0.7-1.0 range for diversity
Performance Expectations
AIME24 Improvement: +18.3%
AIME25 Improvement: +22.8%
Entropy Preservation: 80-95% of baseline diversity
Pass@k Ceiling: Sustained rather than degrading
Model Scales: Works across 3B-32B parameter models
Reference Researchers. (2024). Beyond Pass@1: Self-Play with Variational Problem Synthesis. arXiv preprint arXiv:2508.14029.
