Back to AI Research

AI Research

Interpretable Adaptive Sampling for LLM Test-Time S... | AI Research

Key Takeaways

  • Interpretable Adaptive Sampling for LLM Test-Time Scaling introduces a method to dynamically adjust the amount of computation spent on individual LLM prompts...
  • Test-time scaling improves LLM reasoning by generating and aggregating multiple candidate answers, yet many pipelines use fixed per-query budgets that spend the same compute on easy and difficult prompts.
  • These fixed budgets are also difficult to inspect because they do not explain why a given prompt receives a particular number of samples.
  • We propose adaptive} test-time scaling with a lightweight fuzzy controller that maps interpretable signals, including estimated prompt complexity and model confidence, to a per-query sampling budget.
  • The controller assigns fewer samples to easier or more confident prompts and more samples to harder or less certain prompts, making inference-time compute inspectable rather than fixed or opaque.
Paper AbstractExpand

Test-time scaling improves LLM reasoning by generating and aggregating multiple candidate answers, yet many pipelines use fixed per-query budgets that spend the same compute on easy and difficult prompts. These fixed budgets are also difficult to inspect because they do not explain why a given prompt receives a particular number of samples. We propose adaptive} test-time scaling with a lightweight fuzzy controller that maps interpretable signals, including estimated prompt complexity and model confidence, to a per-query sampling budget. The controller assigns fewer samples to easier or more confident prompts and more samples to harder or less certain prompts, making inference-time compute inspectable rather than fixed or opaque. We evaluate under a fair-alignment protocol with matched decoding settings and controlled answer selection, and compare against best-of-$N$, compute-aware scaling, and self-certainty-based baselines on question-answering and mathematical reasoning tasks. Across models and datasets, adaptive fuzzy control improves over several standard baselines and remains close to a selector-matched full-budget control while reducing the average number of samples. These findings suggest that interpretable adaptive sampling is a practical direction for more efficient test-time reasoning in large language models.

Interpretable Adaptive Sampling for LLM Test-Time Scaling introduces a method to dynamically adjust the amount of computation spent on individual LLM prompts. Instead of using a fixed number of samples for every query, the authors, Mobina Kashaniyan and Ali Jannesari, propose a fuzzy controller that assigns more samples to difficult or uncertain prompts and fewer to easy ones, aiming to improve efficiency while maintaining reasoning performance.

The Problem with Fixed Budgets

Many current LLM pipelines use a fixed "test-time scaling" budget, where the model generates the same number of candidate answers for every prompt. This approach is often inefficient because it wastes computational resources on simple questions while potentially underserving complex ones. Furthermore, these fixed systems are typically "black boxes," offering no explanation for why a specific amount of compute was allocated to a particular query. This lack of transparency makes it difficult for users to audit or debug the reasoning process.

How the Fuzzy Controller Works

The researchers developed a hierarchical fuzzy controller that acts as an interpretable decision-maker. The system processes a prompt through several stages:

  1. Signal Extraction: The system calculates various "human-readable" signals, including prompt length, complexity, expected answer length, model confidence (based on token probabilities), and historical performance. 2. Fuzzy Logic Mapping: These signals are mapped to fuzzy labels like "low," "medium," or "high" using membership functions. Unlike binary thresholds, these functions allow a prompt to partially belong to multiple categories, resulting in a gradual, smooth adjustment of the sampling budget. 3. Refinement and Output: The controller uses a two-stage process to refine the budget. The first stage provides a coarse decision based on complexity and confidence, while the second stage adjusts the budget based on factors like entropy and linguistic complexity. The final output is an integer representing the number of candidate answers to generate.

Performance and Efficiency

The authors evaluated their method on the GSM8K, MATH, and SciQ datasets using Phi-3-mini and Qwen2.5-1.5B models. The results indicate that the adaptive approach achieves an accuracy–compute tradeoff:

  • Efficiency: On the MATH dataset, the controller successfully reduced the average number of samples by 10.8% to 14.5% compared to a fixed-budget baseline, with only a minor impact on accuracy.

  • Accuracy: In some cases, such as with Qwen2.5-1.5B on GSM8K, the adaptive method slightly improved accuracy while simultaneously reducing the average number of samples.

  • Transparency: Because the controller relies on explicit rules and interpretable signals, the system can report exactly which factors influenced the decision to increase or decrease the sample count for any given prompt.

Considerations for Implementation

The authors note that their method is intended to provide a transparent alternative to opaque, learned policies. While a complex neural network might fit training data more precisely, the fuzzy controller is designed to be auditable and understandable without requiring retraining. The researchers emphasize that the goal is not to prove that adaptive budgeting always outperforms full-budget sampling, but to demonstrate that it is possible to maintain high accuracy while focusing computational effort where it is most needed.

Comments (0)

No comments yet

Be the first to share your thoughts!