Skip to main contentpaceevolve-evolution-search
Improves LLM-based evolutionary search by addressing context pollution, mode collapse, and weak collaboration through hierarchical context management, momentum-based backtracking, and adaptive sampling policies.
المهن ذات الصلةSOC
استنادا إلى تصنيف SOC المهني
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
يتجاوز الأمر المباشر Prompt المخصّص للمراجعة. افحص المصدر قبل تشغيله.
npx skills add https://github.com/ADu2021/skillXiv --skill paceevolve-evolution-searchيبقى الأمر في سطر واحد. مرّر أفقيًا لمراجعته كاملًا قبل النسخ.
تفضّل نسخة محلية؟ نزّل الملفات المتاحة حاليًا لدى SkillsMP.
المزيد من هذا المستودع
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.
| name | paceevolve-evolution-search |
| title | PACEvolve: Enabling Long-Horizon Progress-Aware Consistent Evolution |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2601.10657 |
| keywords | ["evolutionary-search","LLM-optimization","long-horizon","progress-aware","context-management"] |
| description | Improves LLM-based evolutionary search by addressing context pollution, mode collapse, and weak collaboration through hierarchical context management, momentum-based backtracking, and adaptive sampling policies. |
Overview
Enhance LLM-driven evolutionary search for long-horizon optimization tasks. Address three key failure modes: context pollution from accumulated experiment data, mode collapse from imbalanced exploration-exploitation, and weak collaboration between parallel search trajectories.
When to Use
- For autonomous optimization and search tasks requiring many candidate evaluations
- When LLMs guide evolutionary algorithms over extended search spaces
- For hyperparameter tuning, architecture search, or program synthesis
- When you need sustained self-improvement over multiple generations
When NOT to Use
- For simple single-pass optimization
- When evolutionary search already works well without LLM guidance
- For tasks with very limited evaluation budget
- In low-latency applications
Key Technical Components
Hierarchical Context Management (HCM)
Prevent context pollution by pruning irrelevant historical data.
class ContextManager:
def __init__(self, max_context_tokens=2000):
self.max_tokens = max_context_tokens
self.experiment_history = []
self.context_cache = {}
def maintain_context_hierarchy(self):
"""Organize context by relevance tiers"""
tiers = {
"recent": [],
"best": [],
"diverse": []
}
tiers["recent"] = self.experiment_history[-10:]
best_experiments = (
.experiment_history,
key= x: x[],
reverse=
)[:]
tiers[] = best_experiments
tiers[] = .select_diverse_representatives(
.experiment_history,
k=
)
tiers
():
(experiments) <= k:
experiments
clusters = .cluster_by_structure(experiments)
representatives = []
cluster clusters:
best_in_cluster = (cluster, key= x: x[])
representatives.append(best_in_cluster)
representatives[:k]
():
pruned_history = []
experiment .experiment_history:
recency_score = .compute_recency(experiment)
performance_score = .compute_performance(experiment)
relevance_score = .compute_problem_relevance(
experiment,
current_problem
)
overall_score = (
* recency_score +
* performance_score +
* relevance_score
)
overall_score > RETENTION_THRESHOLD:
pruned_history.append(experiment)
.experiment_history = pruned_history
pruned_history
():
age = (.experiment_history) - .experiment_history.index(experiment)
/ ( + age)
():
all_fitness = [e[] e .experiment_history]
min_f, max_f = (all_fitness), (all_fitness)
min_f == max_f:
(experiment[] - min_f) / (max_f - min_f)
():
experiment.get() == problem.get():
sorted
self
lambda
"fitness"
True
10
"best"
"diverse"
self
self
10
return
def
select_diverse_representatives
self, experiments, k=10
"""Select diverse experiments by clustering"""
if
len
return
self
for
in
max
lambda
"fitness"
return
def
prune_context
self, current_problem
"""Remove irrelevant historical data"""
for
in
self
self
self
self
0.3
0.3
0.4
if
self
return
def
compute_recency
self, experiment
"""Score based on how recent"""
len
self
self
return
1.0
1.0
def
compute_performance
self, experiment
"""Normalize fitness score"""
"fitness"
for
in
self
min
max
if
return
0.5
return
"fitness"
def
compute_problem_relevance
self, experiment, problem
"""Semantic relevance to current problem"""
if
"problem_id"
"id"
return
1.0
return
0.1
Momentum-Based Backtracking (MBB)
Escape local optima by reverting to diverse solutions.
class MomentumBacktracker:
def __init__(self, backtrack_window=5):
self.backtrack_window = backtrack_window
self.trajectory_history = []
self.fitness_trend = []
def detect_local_optimum(self):
"""Check if stuck in local optimum"""
if len(self.fitness_trend) < self.backtrack_window:
return False
recent_trend = self.fitness_trend[-self.backtrack_window:]
improvement = max(recent_trend) - min(recent_trend)
stagnation = improvement < STAGNATION_THRESHOLD
solution_similarity = self.compute_avg_similarity()
low_diversity = solution_similarity > SIMILARITY_THRESHOLD
is_stuck = stagnation and low_diversity
return {
"is_stuck": is_stuck,
"stagnation_score": improvement,
"diversity_score": 1.0 - solution_similarity
}
def backtrack_with_momentum(self, current_fitness):
"""Revert to promising past solution with momentum"""
checkpoint = self.select_backtrack_checkpoint()
if checkpoint is None:
return None
current_solution = self.trajectory_history[-1]["solution"]
checkpoint_solution = checkpoint["solution"]
momentum = self.compute_momentum_vector(current_solution, checkpoint_solution)
backtrack_solution = self.blend_solutions(
checkpoint_solution,
momentum,
alpha=0.7
)
return {
"solution": backtrack_solution,
"checkpoint_fitness": checkpoint["fitness"],
"momentum_direction": momentum
}
def select_backtrack_checkpoint(self):
"""Select good past solution with different structure"""
candidates = []
for i in range(max(0, len(self.trajectory_history) - 20), len(self.trajectory_history)):
solution = self.trajectory_history[i]
if solution["fitness"] > BACKTRACK_FITNESS_THRESHOLD:
diversity = self.compute_diversity_from_recent(solution)
if diversity > DIVERSITY_THRESHOLD:
candidates.append(solution)
if not candidates:
return None
return max(candidates, key=lambda x: x["fitness"])
def blend_solutions(self, base_solution, momentum, alpha=0.7):
"""Blend checkpoint with momentum direction"""
blended = alpha * base_solution + (1 - alpha) * momentum
return blended
def compute_momentum_vector(self, current, checkpoint):
"""Direction of progress from checkpoint to current"""
return current - checkpoint
def compute_diversity_from_recent(self, solution):
"""How different from recent solutions"""
recent_solutions = [s["solution"] for s in self.trajectory_history[-5:]]
similarities = [
self.compute_solution_similarity(solution["solution"], recent)
for recent in recent_solutions
]
avg_similarity = np.mean(similarities)
return 1.0 - avg_similarity
Self-Adaptive Sampling Policy
Dynamically balance exploration and exploitation.
class AdaptiveSamplingPolicy:
def __init__(self):
self.exploration_rate = 0.5
self.sampling_history = []
def compute_adaptive_rate(self, recent_progress):
"""Adjust exploration based on progress"""
improvement = np.mean(recent_progress)
if improvement > HIGH_PROGRESS_THRESHOLD:
self.exploration_rate = max(0.1, self.exploration_rate - 0.1)
elif improvement < LOW_PROGRESS_THRESHOLD:
self.exploration_rate = min(0.9, self.exploration_rate + 0.1)
return self.exploration_rate
def sample_next_candidate(self, best_candidates, random_candidates, exploration_rate):
"""Choose between exploiting best or exploring random"""
if np.random.random() < exploration_rate:
return np.random.choice(random_candidates)
else:
return np.random.choice(best_candidates)
def integrate_backtracking(self, backtrack_solution, exploration_rate):
"""Incorporate backtracking into sampling"""
backtrack_exploration = min(
exploration_rate + 0.3,
0.9
)
return backtrack_exploration
def compute_sampling_efficiency(self):
"""Track whether sampling strategy is effective"""
total_samples = len(self.sampling_history)
improvements = sum(
1 for i in range(1, len(self.sampling_history))
if self.sampling_history[i]["fitness"] > self.sampling_history[i-1]["fitness"]
)
efficiency = improvements / total_samples
return efficiency
Integration: Complete Loop
Combine all components into complete search loop.
class PACEevolveSearch:
def __init__(self, llm_generator):
self.generator = llm_generator
self.context_manager = ContextManager()
self.backtracker = MomentumBacktracker()
self.sampler = AdaptiveSamplingPolicy()
def search_iteration(self, problem, budget=100):
"""Single generation of evolutionary search"""
for step in range(budget):
context = self.context_manager.maintain_context_hierarchy()
self.context_manager.prune_context(problem)
candidates = self.generator.generate_candidates(
problem,
context,
num_candidates=10
)
candidates = self.evaluate_candidates(candidates)
stuck = self.backtracker.detect_local_optimum()
if stuck["is_stuck"]:
backtrack = self.backtracker.backtrack_with_momentum(
max(e["fitness"] for e in candidates)
)
if backtrack:
candidates.append(backtrack["solution"])
best_candidates = sorted(candidates, key=lambda x: x["fitness"], reverse=True)[:5]
self.sampler.compute_adaptive_rate(
[e["fitness"] for e in candidates]
)
if best_candidates:
self.context_manager.experiment_history.append(best_candidates[0])
return self.context_manager.experiment_history
Performance Characteristics
- State-of-the-art on LLM-SR and KernelBench benchmarks
- Discovers solutions exceeding previous records
- Sustained improvement over 100+ iterations
- Effective on Modded NanoGPT and similar tasks
Recommendations
- Initialize exploration rate at 0.5; let it adapt
- Backtrack when stagnation detected; don't backtrack too frequently
- Prune context every 10-20 iterations to prevent pollution
- Monitor sampling efficiency; adjust thresholds if needed
References
- Context pollution prevents long-horizon optimization
- Mode collapse causes premature convergence
- Momentum-based backtracking enables escape from local optima
- Adaptive sampling maintains balance without manual tuning
