Seer Self-Consistency: Advance Budget Estimation for Adaptive Test-Time Scaling

TL;DR

SeerSC addresses the inefficiency of test-time scaling in large language models by combining rapid System 1 entropy estimation with parallel System 2 reasoning to dynamically allocate computation per sample. By computing a confidence-weighted entropy $E(X)$ from multiple direct System 1 answers and budgeting System 2 work via $B(X)$, SeerSC enables efficient, parallelized self-consistency without sacrificing accuracy. Empirical results on diverse math and STEM benchmarks show up to 47% token reduction and 43% latency reduction, while maintaining accuracy comparable to or better than strong baselines, and the method is compatible with existing optimizations like weighted voting and path pruning. This approach provides a practical, latency-aware framework for adaptive test-time scaling across challenging reasoning tasks.

Abstract

Test-time scaling improves the inference performance of Large Language Models (LLMs) but also incurs substantial computational costs. Although recent studies have reduced token consumption through dynamic self-consistency, they remain constrained by the high latency of sequential requests. In this paper, we propose SeerSC, a dynamic self-consistency framework that simultaneously improves token efficiency and latency by integrating System 1 and System 2 reasoning. Specifically, we utilize the rapid System 1 to compute the answer entropy for given queries. This score is then used to evaluate the potential of samples for scaling, enabling dynamic self-consistency under System 2. Benefiting from the advance and accurate estimation provided by System 1, the proposed method can reduce token usage while simultaneously achieving a significant decrease in latency through parallel generation. It outperforms existing methods, achieving up to a 47% reduction in token consumption and a 43% reduction in inference latency without significant performance loss.

Figures (6)

• Figure 1: Comparison of the Proposed method with Existing Methods. Unlike sequential estimation methods, SeerSC leverages fast System 1 to guide the parallel generation of System 2.
• Figure 2: Observational Experiments on the Math Dataset.
• Figure 3: Overall workflow of the proposed SeerSC framework. We first utilize System 1 to rapidly estimate the answer diversity for each sample, and subsequently allocate varying computational resources for the System 2 phase based on the answer entropy.
• Figure 4: Latency comparison of System 1 and System 2 in SeerSC on AIME-2024, using DeepSeek-R1-Distill-Qwen-7B and Qwen3-4B (non-thinking).
• Figure 5: Latency scaling results of SeerSC compared with baseline methods on AIME-2024 and GPQA-D using Qwen3-4B (non-thinking) and DeepSeek-R1-Distill-Qwen-7B. SeerSC matches or exceeds SC under low-latency conditions, reduces latency by about 50% compared to AC and ESC on AIME-2024, and achieves the same accuracy with up to 75% lower latency on GPQA-D.