| 摘要: |
| 针对大语言模型（Large Language Models，LLMs）在生成作战知识时易产生“幻觉”且面临多源情报冲突的难题，提出了一种融合对抗性证据推理与自适应特征融合的可信度评估方法。首先，针对单一查询模式下情报召回率低的问题，提出多视图生成增强策略，并构建融合语义对齐与信源权威性的双因子检索机制。其次，设计基于思维链（Chain-of-Thought，CoT）的对抗性推理框架，模拟“红队”视角主动挖掘反驳证据，以抑制模型的证实偏差。最后，构建基于多层感知机（Multilayer Perceptron，MLP）的监督式决策融合模型，实现从多维认知特征到置信度评分的非线性映射。在构建的作战知识幻觉评测数据集（Combat Knowledge Hallucination Evaluation Dataset，CKHE）上的实验结果表明，该方法的Macro-F1值达到85.41%，在80%的高噪声干扰下仍能保持75%以上的性能。 |
| 关键词: 作战知识评估 大语言模型(LLM) 对抗性证据推理 自适应特征融合 |
| DOI：10.20079/j.issn.1001-893x.251213001 |
|
| 基金项目: |
|
| Assessing LLM-generated Combat Knowledge via Adversarial Reasoning and Multidimensional Adaptive Fusion |
| WANG Wei |
| (Southwest China Institute of Electronic Technology，Chengdu 610036，China) |
| Abstract: |
| To address the challenges of “hallucinations” and multi-source intelligence conflicts inherent in Large Language Models(LLMs) when generating combat knowledge，a credibility assessment method that integrates adversarial evidence reasoning and adaptive feature fusion is proposed.First，to address the low intelligence recall rate under a single query model，a multi-view generation enhancement strategy is proposed，and a two-factor retrieval mechanism integrating semantic alignment and source authority is constructed.Second，an adversarial reasoning framework based on Chain-of-Thought(CoT) is designed to simulate the perspective of the “red team” and actively mine rebuttal evidence to suppress confirmation bias in the model.Finally，a supervised decision fusion model based on Multilayer Perceptron(MLP) is constructed to achieve a nonlinear mapping from multidimensional cognitive features to confidence scores.Experimental results on the constructed Combat Knowledge Hallucination Evaluation Dataset(CKHE) show that the proposed method achieves a Macro-F1 score of 85.41%，maintaining performance above 75% even under 80% high noise interference. |
| Key words: combat knowledge assessment large language model(LLM) adversarial evidence reasoning adaptive feature fusion |
《电讯技术》编辑部 