Scale Selective Test Time

Summary

SCALE addresses inefficient resource allocation in test-time scaling by implementing selective resource allocation based on sub-problem difficulty. Rather than uniformly distributing computational resources across all steps, the approach decomposes mathematical problems into sequential sub-problems, assesses difficulty, and dynamically assigns simple problems to fast processing while directing complex ones to deliberate reasoning.

Core Technique

Problem Decomposition: Break mathematical problems into smaller steps or sub-problems. For math:

• Step 1: Parse and identify key quantities
• Step 2: Determine solution strategy
• Step 3: Execute calculations
• Step 4: Verify answer

Difficulty Assessment: For each sub-problem, estimate complexity via:

• Token count of problem statement
• Number of dependencies on previous steps
• Complexity of mathematical operations involved
• Uncertainty in model confidence

Dual-System Allocation:

• System 1 (Fast): Single-shot generation with minimal reasoning, optimal for straightforward calculations
• System 2 (Deliberate): Multiple sampling, majority voting, extended reasoning for complex steps

Implementation

Difficulty scorer: Train a lightweight classifier:

difficulty = mlp(encode(subproblem))  # 0 (easy) to 1 (hard)

Dynamic allocation:

if difficulty < 0.3:
    # System 1: single pass
    answer = model.generate(subproblem, max_tokens=50)
else:
    # System 2: deliberate reasoning
    answers = [model.generate(subproblem, max_tokens=500) for _ in range(5)]
    answer = majority_vote(answers)

Token counting: Track token usage:

total_tokens = sum(difficulty_threshold * tokens_system1 + (1 - difficulty_threshold) * tokens_system2)

Accuracy gains: Hard sub-problems get more compute, soft problems use minimal resources, overall accuracy improves while token usage decreases.

When to Use

• Mathematical reasoning tasks with variable problem difficulty
• Applications where token budget is constrained but accuracy is critical
• Scenarios where problem decomposition is natural
• Automated grading systems (AIME, AMC, etc.) with measurable accuracy

When NOT to Use

• Tasks where uniform reasoning across all steps is necessary
• Scenarios without clear problem decomposition
• Applications where difficulty assessment is unreliable
• Real-time inference where dynamic allocation overhead matters

Key References

• System 1 and System 2 thinking in cognitive science
• Adaptive computation and dynamic neural networks
• Majority voting and ensemble methods for reasoning
• Difficulty estimation and sampling strategies