메뉴
Allocate LLM reasoning budget optimally via value tree search: use residual value prediction to estimate step utility, then dynamically shift exploration-exploitation balance as budget depletes. Outperform high-budget baselines at 1/4 cost.
기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.
설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.
SOC 직업 분류 기준
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
직접 명령은 검토 Prompt를 거치지 않습니다. 실행하기 전에 소스를 확인하세요.
npx skills add https://github.com/ADu2021/skillXiv --skill bavt-budget-aware-search명령은 한 줄로 유지됩니다. 복사하기 전에 가로로 스크롤해 전체 내용을 확인하세요.
로컬 사본을 원하시나요? SkillsMP에서 현재 제공할 수 있는 파일을 다운로드하세요.
SKILL.md 표시 중
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.
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.
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 | bavt-budget-aware-search |
| title | Spend Less, Reason Better: Budget-Aware Value Tree Search for LLM Agents |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2603.12634 |
| keywords | ["Tree Search","Budget-Aware","Value Estimation","Multi-Hop Reasoning","LLM"] |
| description | Allocate LLM reasoning budget optimally via value tree search: use residual value prediction to estimate step utility, then dynamically shift exploration-exploitation balance as budget depletes. Outperform high-budget baselines at 1/4 cost. |
Agentic reasoning requires exploring multiple paths, but computational budgets are finite. Budget-Aware Value Tree (BAVT) search makes principled allocation decisions: it estimates marginal utility per step using residual value prediction, then dynamically modulates exploration strength as budget depletes.
The key insight is a power-law scaling exponent inversely proportional to remaining budget, creating smooth transitions from broad exploration to aggressive greedy exploitation.
BAVT operates through three mechanisms:
This achieves superior performance under budget constraints: 4× budget-efficient compared to baseline.
Critic estimates incremental progress, not absolute returns.
import torch
import torch.nn as nn
class ResidualValuePredictor(nn.Module):
def __init__(self, hidden_dim=768):
super().__init__()
self.hidden_dim = hidden_dim
# MLP for residual value prediction
self.predictor = nn.Sequential(
nn.Linear(hidden_dim, hidden_dim * 2),
nn.ReLU(),
nn.Linear(hidden_dim * 2, )
)
():
residual_value = .predictor(context_embeddings)
residual_value
(nn.Module):
():
().__init__()
.model = model
.value_predictor = ResidualValuePredictor(hidden_dim)
():
hidden_state = .model.encode(trajectory_text)
residual = .value_predictor(hidden_state)
residual.item()
Dynamically adjust exploration-exploitation trade-off based on remaining budget.
class BudgetAwareUCB:
def __init__(self, exploration_constant=1.0):
self.exploration_constant = exploration_constant
def compute_ucb_score(
self,
node_value,
visit_count,
total_budget,
remaining_budget
):
"""
Compute UCB score with budget-dependent exploration.
Scaling exponent: α_t = 1 / r_t (r_t = remaining budget)
"""
# Exploitation term: average value
exploitation = node_value / (visit_count + 1e-8)
# Exploration term with budget-dependent scaling
exploration_scale = self.exploration_constant / (remaining_budget + 1e-8)
# UCB = exploitation + sqrt(scale * ln(parent_visits) / visit_count)
import math
ucb_score = (
exploitation +
exploration_scale * math.sqrt(
math.log(total_budget + 1) / (visit_count + 1e-8)
)
)
return ucb_score
def select_best_child(self, node, remaining_budget, total_budget):
"""
Select child with highest UCB score.
"""
best_child = None
best_score = -float('inf')
for child in node.children:
ucb_score = self.compute_ucb_score(
child.value_sum,
child.visit_count,
total_budget,
remaining_budget
)
if ucb_score > best_score:
best_score = ucb_score
best_child = child
return best_child
Use single LLM model for both generation and critic roles.
class TreeNode:
def __init__(self, text="", parent=None):
self.text = text
self.parent = parent
self.children = []
self.visit_count = 0
self.value_sum = 0.0
self.residual_value = None
class BudgetAwareValueTreeSearch:
def __init__(self, model, critic_head, budget=1000):
self.model = model
self.critic = critic_head
self.total_budget = budget
self.remaining_budget = budget
def search(self, question, max_depth=10):
"""
Perform budget-aware tree search for question.
"""
root = TreeNode(question)
ucb = BudgetAwareUCB()
for budget_step in range(self.total_budget):
# Update remaining budget
self.remaining_budget = self.total_budget - budget_step
# Selection and expansion
leaf_node = self._select_and_expand(
root,
ucb,
max_depth
)
if leaf_node is None:
break
# Evaluation: critic estimates residual value
residual_value = self.critic.estimate_residual_value(
leaf_node.text
)
# Backup: propagate value up tree
self._backup(leaf_node, residual_value)
# Select best trajectory
best_path = self._extract_best_path(root)
return best_path
def _select_and_expand(self, node, ucb, max_depth):
"""
Traverse tree using UCB, expand at leaf.
"""
current = node
depth = 0
while current.children and depth < max_depth:
current = ucb.select_best_child(
current,
self.remaining_budget,
self.total_budget
)
depth += 1
# Expand: generate next steps
if depth < max_depth:
next_steps = self.model.generate_next_steps(
current.text,
num_candidates=3
)
for step in next_steps:
child = TreeNode(current.text + "\n" + step, parent=current)
current.children.append(child)
# Return first child for evaluation
if current.children:
return current.children[0]
return None
def _backup(self, node, value):
"""
Propagate value up tree from leaf.
"""
current = node
while current is not None:
current.visit_count += 1
current.value_sum += value
current = current.parent
def _extract_best_path(self, root):
"""
Extract highest-value path from root to leaf.
"""
path = []
current = root
while current.children:
# Select child with highest average value
best_child = max(
current.children,
key=lambda c: c.value_sum / (c.visit_count + 1e-8)
)
path.append(best_child.text)
current = best_child
return '\n'.join(path)
Use model for both roles without additional training.
class DualRoleLLMAgent:
def __init__(self, base_model):
self.model = base_model
self.critic_head = CriticHead(base_model)
def reason_with_budget(
self,
question,
budget_tokens=1000,
max_depth=10
):
"""
Multi-step reasoning within token budget.
"""
search = BudgetAwareValueTreeSearch(
self.model,
self.critic_head,
budget=budget_tokens
)
best_reasoning = search.search(question, max_depth)
return best_reasoning
def generate_next_steps(self, context, num_candidates=3):
"""
Generator role: propose next reasoning steps.
"""
prompt = f"{context}\n\nNext reasoning steps:"
steps = []
for _ in range(num_candidates):
step = self.model.generate(prompt, max_tokens=50, temperature=0.7)
steps.append(step.strip())
return steps
def compare_budgets(self, question, budgets=[100, 250, 500, 1000]):
"""
Demonstrate budget-performance trade-off.
"""
results = {}
for budget in budgets:
reasoning = self.reason_with_budget(question, budget)
# Evaluate (external)
score = self.evaluate(reasoning, question)
results[budget] = score
return results
Compare budget efficiency to standard approaches.
def benchmark_budget_efficiency():
"""
Demonstrate BAVT efficiency gains.
"""
agent = DualRoleLLMAgent(base_model)
questions = [...] # Test set
results = {
'bavt': {},
'baseline_fixed': {},
'baseline_10x_budget': {}
}
# BAVT with varying budgets
for budget in [250, 500, 1000]:
scores = []
for q in questions:
answer = agent.reason_with_budget(q, budget)
score = evaluate(answer, q)
scores.append(score)
results['bavt'][budget] = sum(scores) / len(scores)
# Baseline with fixed budget
baseline_budget = 1000
scores = []
for q in questions:
answer = agent.reason_fixed_budget(q, baseline_budget)
score = evaluate(answer, q)
scores.append(score)
results['baseline_fixed'][baseline_budget] = sum(scores) / len(scores)
# Baseline with 10x budget
scores = []
for q in questions:
answer = agent.reason_fixed_budget(q, baseline_budget * 10)
score = evaluate(answer, q)
scores.append(score)
results['baseline_10x_budget'][baseline_budget * 10] = sum(scores) / len(scores)
return results
When to Use:
When NOT to Use:
Hyperparameter Tuning:
Common Pitfalls: