跳转至

文章背景与核心概要

本文探讨了一种先进的AI幻觉检测与缓解框架，该框架将受HOPE启发的嵌套学习架构与连续记忆系统（CMS）及语义相似度缓存相结合。作者在一个包含310个提示词（其中包含217个认知不确定性查询和93个虚构诱导压力测试）的混合基准上进行了评估，实现了一个通过开放发言权协议（OFP）进行编排的三阶段流水线。

核心研究结果揭示了多维可靠性指标的微妙变化：性能方面，聚合的总幻觉分数在端到端上提升了其可达成范围的6.1%，其中高达97.7%的提升是在第一轮审查阶段实现的；在维度差异上，分数的提升实际上是一维的——83.5%的总提升仅由“显式语境化”驱动，而对于追踪无根据内容最为关键的“事实声明密度”指标则保持平稳；在可持续性方面，语义缓存成功处理了47.7%的模型调用，优化了计算开销。

Hallucination Mitigation with Agentic AI, Nested Learning, and AI Sustainability via Semantic Caching

Authors: Diego Gosmar, Deborah A. Dahl
Identifiers: arXiv:2605.29055 [cs.AI] | DOI: 10.48550/arXiv.2605.29055
Submission History: Submitted May 27, 2026; Revised August 12, 2026 (v2)

Hallucination Mitigation with Agentic AI, Nested Learning, and AI Sustainability via Semantic Caching

Authors: Diego Gosmar, Deborah A. Dahl
Identifiers: arXiv:2605.29055 [cs.AI] | DOI: 10.48550/arXiv.2605.29055
Submission History: Submitted May 27, 2026; Revised August 12, 2026 (v2)

Summary

本文研究了一个先进的框架，该框架通过将受HOPE启发的嵌套学习架构与连续记忆系统（CMS）和语义相似度缓存相结合，来检测和缓解AI幻觉。通过包含310个提示词（包含217个认知不确定性查询和93个虚构诱导压力测试）的混合基准进行评估，作者实现了通过开放发言权协议（OFP）编排的三阶段流水线。

关键发现在多维可靠性指标上展现出了微妙的特性： * 性能增益： 聚合的总幻觉分数在端到端上提升了其可达成范围的6.1%，其中惊人的97.7%的增益是在最初的审查阶段实现的。 * 维度差异： 分数的提升实际上是一维的——83.5%的总增益完全由显式语境化驱动，而事实声明密度（追踪无根据内容最关键的指标）则保持不变。 * 可观测性： 可观测性指标充当了一个注释通道，而非直接的响应属性，它在审查阶段跃升了147%，随后在通道停止传播时回落。 * 人工与跨模型验证： 三位人类标注员对93个压力提示词进行的独立标注显示，最终答案中有10.8%（95%置信区间为5.9–18.7）仍将虚构项目呈现为真实内容（\(\alpha = 0.586\)）。跨系列评判模型（Llama 3.1、Gemma 4和Qwen 3）重新对930个输出进行了评分，并与人类评估高度一致（\(\rho = -0.772\) vs 原评估者的 \(-0.477\)）。 * 可持续性： 语义缓存成功服务了47.7%的所有模型调用，优化了计算开销。

Summary

This paper investigates an advanced framework for detecting and mitigating AI hallucinations by coupling a HOPE-inspired Nested Learning architecture with Continuum Memory Systems (CMS) and semantic similarity caching. Evaluated on a hybrid benchmark of 310 prompts (including 217 epistemic-uncertainty queries and 93 fabrication-induction stress tests), the authors implement a three-stage pipeline orchestrated via the Open Floor Protocol (OFP).

Key findings reveal nuances in multi-dimensional reliability metrics: * Performance Gains: The aggregated Total Hallucination Score improves end-to-end by 6.1% of its attainable range, with an overwhelming 97.7% of this gain achieved during the very first review stage. * Dimensional Discrepancy: The score improvements are effectively one-dimensional—83.5% of the total gain is driven solely by Explicit Contextualization, while Factual Claim Density (the metric most critical for tracking unsupported content) remains flat. * Observability: The observability indicator acts as an annotation channel rather than a direct response property, jumping 147% at the review stage before falling back when the channel ceases propagation. * Human and Cross-Model Verification: Independent labeling by three human annotators on the 93 stress prompts revealed that 10.8% (95% CI 5.9–18.7) of final answers still present invented items as real (\(\alpha = 0.586\)). Cross-family judges (Llama 3.1, Gemma 4, and Qwen 3) re-scored 930 outputs and aligned strongly with human evaluations (\(\rho = -0.772\) vs. \(-0.477\) for the original evaluator). * Sustainability: Semantic caching successfully served 47.7% of all model calls, optimizing computational overhead.

Metadata & Reference Information

元数据与参考信息

Metadata & Reference Information

Associated Tools & Resources

关联工具与资源

  • 文献计量工具： Explorer, Connected Papers, Litmaps, scite Smart Citations
  • 代码与数据： alphaXiv, CatalyzeX Code Finder, DagsHub, Gotit.pub, Hugging Face, ScienceCast
  • 演示与推荐系统： Replicate, Hugging Face Spaces, TXYZ.AI, Influence Flower, CORE Recommender

Associated Tools & Resources

  • Bibliographic Tools: Explorer, Connected Papers, Litmaps, scite Smart Citations
  • Code & Data: alphaXiv, CatalyzeX Code Finder, DagsHub, Gotit.pub, Hugging Face, ScienceCast
  • Demos & Recommenders: Replicate, Hugging Face Spaces, TXYZ.AI, Influence Flower, CORE Recommender