Claim-Level Evidence Admissibility: A New Framework to Enhance the Reliability of Structured Outputs from Large Language Models

This study proposes the EGCR (Claim-Level Evidence Admissibility) framework, which filters unreliable outputs from large language models (LLMs) through claim-level evidence admissibility assessment. It is suitable for high-risk scenarios such as cybersecurity risk assessment and AI deployment decision-making. The original authors are the research team from Nanchang University, and it was released on GitHub in 2024.

LLMs have great potential in structured outputs (data extraction, risk assessment, etc.), but they tend to generate incorrect or unreliable content, which can have serious consequences in high-risk scenarios. Existing overall output verification methods have flaws: either they discard a large amount of correct information due to a few errors, or they miss detecting errors. A fine-grained claim-level assessment mechanism is needed.

1. Fine-grained claim-level assessment: Break down outputs into independent claim units and accurately judge the evidence support for each assertion; 2. Evidence admissibility criteria: Evidence existence, relevance, sufficiency, and consistency; 3. Selective registration mechanism: Only claims that pass the assessment are included in the final output.

Two controlled datasets were constructed: 1. Cybersecurity risk suite: Covers vulnerability cases such as SQL injection and XSS; 2. AI deployment risk suite: Focuses on compliance and ethical risks (data privacy, algorithm fairness, etc.).

Open-source models tested: Granite4.1(8B), Llama3.1(8B), Qwen2.5(7B), Qwen3(30B), Gemma4(31B). The results show that the EGCR framework can effectively improve the output quality of various models, and the improvement effect is consistent.

1. Strategy trade-off: Adjusting the evidence threshold can balance reliability and information integrity; 2. Multi-dimensional assessment: Covers dimensions such as consistency, selectivity, and contradiction detection; 3. Baseline comparison: EGCR outperforms traditional methods like overall verification and confidence filtering in balancing precision and recall.

Application scenarios: Cybersecurity (reducing false positives and false negatives in vulnerability assessment), AI governance (compliance risk assessment); Can be extended to fields such as healthcare, law, and finance; The open-source implementation provides a foundation for subsequent research and promotes the development of LLM reliability assessment.