ResearchPod Summary
As Large Language Models (LLMs) and Agentic AI systems are deployed in cybersecurity operations, ensuring trust and provenance in generated outputs becomes vital. Retrieval-Augmented Generation (RAG) grounds LLMs using external information like cyber incident logs. However, retrieval alone does not guarantee that the model correctly interprets or utilizes specific evidence, especially when processing structured, highly similar security logs. Existing attribution techniques—such as LLM-as-a-judge models, similarity measures, and token-level importance scores—either suffer from self-assessment biases or dilute critical signals across overlapping, non-discriminative tokens. This paper asks: How can we reliably trace an LLM-generated security response back to its exact originating log segment at a holistic, segment level?
To bridge this evidence attribution gap, the authors propose Topological Attribution Distance (TAD), a segment-level attribution metric inspired by Topological Data Analysis (TDA). TAD evaluates the global geometric shape of a model's response in its hidden-state embedding space across all transformer layers. By performing segment-level ablation—observing how the inclusion or omission of a specific retrieved log deforms the response's embedding geometry—TAD quantifies the influence of each source. Specifically, the method computes persistence diagrams of the response's embedding geometry with and without each candidate log segment, and measures the topological cost using the Wasserstein distance.
Evaluating TAD on real-world cyberattack logs across Direct, Regular, and Indirect difficulty scenarios, the authors compare the framework against multiple baselines including similarity measures, ROUGE-L, self-evaluation citation, LLM judges, and token-level attribution. The results demonstrate that TAD significantly outperforms existing baselines, achieving over 97% average accuracy in identifying top-attributed logs and maintaining more than a 39% precision gap over the best-performing baseline across all test scenarios.
By providing an explainable, provenance-aware measure that operates directly on an LLM's internal hidden states, TAD offers a rigorous tool for evidence verification. This capability is essential for auditing autonomous security agents, reducing hallucinations, and building operational trust in critical cybersecurity and decision-making workflows.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.