Scholarlink

This paper explores the potential application of reinforcement learning (RL) for reasoning in large language models (LLMs) within the field of explainable artificial intelligence (XAI). DeepSeek recently introduced a training method that achieves strong reasoning capabilities in LLMs through unsupervised RL on mathematical and programming problems. We discuss how a similar approach could be adapted for XAI by training a language model using the output of an existing model as ground truth. If the model converges successfully, it could replicate the outputs of the original model while also providing a natural language (NL) reasoning process leading to these outputs. While this method presents benefits such as in-depth NL explanations and being model-agnostic, several challenges must be considered. These include the computational cost of training LLMs, the appropriate formatting of input data for different problem domains, the relevance of the relationship between the LLM and the original model, and identifying the specific applications where this method would be feasible and beneficial. Initial experiments were conducted to assess the potential of this approach, and the preliminary results are mixed.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Leveraging Large Language Models Reinforcement Learning for Explainable Artificial Intelligence

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Arthur Picard,
  • Yazan Mualla,
  • Franck Gechter

摘要

This paper explores the potential application of reinforcement learning (RL) for reasoning in large language models (LLMs) within the field of explainable artificial intelligence (XAI). DeepSeek recently introduced a training method that achieves strong reasoning capabilities in LLMs through unsupervised RL on mathematical and programming problems. We discuss how a similar approach could be adapted for XAI by training a language model using the output of an existing model as ground truth. If the model converges successfully, it could replicate the outputs of the original model while also providing a natural language (NL) reasoning process leading to these outputs. While this method presents benefits such as in-depth NL explanations and being model-agnostic, several challenges must be considered. These include the computational cost of training LLMs, the appropriate formatting of input data for different problem domains, the relevance of the relationship between the LLM and the original model, and identifying the specific applications where this method would be feasible and beneficial. Initial experiments were conducted to assess the potential of this approach, and the preliminary results are mixed.