跳转至

文章背景与核心概要

尽管诸如监督思维链(CoT)微调和强化学习(RL)等后训练方法已成功提升了大语言模型(LLM)的数学推理能力,但它们在信号处理领域的应用仍未得到充分探索。本报告研究了强化微调策略,旨在利用 WirelessMATHBench-XLQwen2.5-3B-Base 模型适配于研究生级别的信号数学问题。

该研究对两种主要的训练范式进行了基准测试:1. 基于可验证奖励在 WirelessMATHBench-XL 上直接进行强化学习(Direct RL);2. 在蒸馏得到的无线领域 CoT 语料库上进行监督微调(SFT),随后进入相同的领域特定强化学习阶段。在两种范式中,作者评估了三种策略优化算法:群相对策略优化(GRPO)、群序列策略优化(GSPO)以及几何平均策略优化(GMPO),并通过实验揭示了初始化影响、算法稳定性和整体性能表现。


SignalReasoner: Assessing the Upper Bound of 3B Models for Signal Mathematical Reasoning

Authors: Guozheng Sun
Submitted: August 18, 2026
Primary Subject: Computer Science > Artificial Intelligence (cs.AI)
arXiv ID: arXiv:2608.17301 | DOI: 10.48550/arXiv.2608.17301


📋 Summary

尽管后训练方法(如监督思维链(CoT)微调和强化学习(RL))已成功增强了大语言模型(LLM)的数学推理能力,但它们在信号处理领域的应用在很大程度上仍未得到充分探索。

While post-training methods like supervised chain-of-thought (CoT) fine-tuning and reinforcement learning (RL) have successfully enhanced the mathematical reasoning of large language models (LLMs), their application to signal processing remains largely under-explored.

本报告研究了强化微调策略,旨在利用 WirelessMATHBench-XLQwen2.5-3B-Base 模型适配于研究生级别的信号数学问题。该研究对两种主要的训练范式进行了基准测试: 1. 直接强化学习(Direct RL):在 WirelessMATHBench-XL 上使用可验证奖励进行训练。 2. 监督微调(SFT):在蒸馏得到的无线领域 CoT 语料库上进行 SFT,随后进行相同的领域特定强化学习阶段。

在两种范式中,作者评估了三种策略优化算法:群相对策略优化(GRPO)群序列策略优化(GSPO)几何平均策略优化(GMPO)

This report investigates reinforcement fine-tuning strategies to adapt the Qwen2.5-3B-Base model for graduate-level signal mathematical problems using WirelessMATHBench-XL. The study benchmarks two primary training paradigms: 1. Direct RL on WirelessMATHBench-XL using verifiable rewards. 2. Supervised Fine-Tuning (SFT) on a distilled wireless-domain CoT corpus, followed by the same domain-specific RL stage.

Across both paradigms, the author evaluates three policy optimization algorithms: Group Relative Policy Optimization (GRPO), Group Sequence Policy Optimization (GSPO), and Geometric-Mean Policy Optimization (GMPO).

Key Findings

  • 初始化影响:评估了领域感知的 CoT SFT 是否能有效初始化后续的强化学习。
  • 算法稳定性:评估了 GSPO 或 GMPO 相比于 GRPO 是否具备稳定性或准确性优势。
  • 性能表现:表现最好的模型取得了 39.12% 的总体准确率,较未训练的 Base 模型(12.37%)实现了超过三倍的提升

Key Findings

  • Initialization Impact: Evaluated whether domain-aware CoT SFT effectively initializes subsequent RL.
  • Algorithm Stability: Assessed whether GSPO or GMPO provide stability or accuracy advantages over GRPO.
  • Performance: The best-performing model achieved an overall accuracy of 39.12%, marking a more than threefold improvement over the untrained Base model (12.37%).