SSR: Speculative Parallel Scaling Reasoning in Test-time

TL;DR

SSR introduces a training-free speculative decoding framework that accelerates multi-step reasoning in large language models, improving efficiency without sacrificing accuracy across multiple mathematical benchmarks.

Contribution

It presents a novel, training-free approach combining selective strategy identification and step-level speculative decoding for efficient reasoning in LLMs.

Findings

SSR improves accuracy by 13.84% on LiveMathBench.

SSR reduces computation to 30% on MATH-500.

SSR achieves strong gains over baselines in multiple benchmarks.

Abstract

Large language models (LLMs) have achieved impressive results on multi-step mathematical reasoning, yet at the cost of high computational overhead. This challenge is particularly acute for test-time scaling methods such as parallel decoding, which increase answer diversity but scale poorly in efficiency. To address this efficiency-accuracy trade-off, we propose SSR (Speculative Parallel Scaling Reasoning), a training-free framework that leverages a key insight: by introducing speculative decoding at the step level, we can accelerate reasoning without sacrificing correctness. SSR integrates two components: a Selective Parallel Module (SPM) that identifies a small set of promising reasoning strategies via model-internal scoring, and Step-level Speculative Decoding (SSD), which enables efficient draft-target collaboration for fine-grained reasoning acceleration. Experiments on three mathematical benchmarks-AIME 2024, MATH-500, and LiveMathBench - demonstrate that SSR achieves strong gains over baselines. For instance, on LiveMathBench, SSR improves pass@1 accuracy by 13.84% while reducing computation to 80.5% of the baseline FLOPs. On MATH-500, SSR reduces compute to only 30% with no loss in accuracy.