Paper: arXiv 2412.12148
Authors: Bhaskarjit Sarmah, Mingshu Li, Jingrao Lyu, Sebastian Frank, Nathalia Castellanos, Stefano Pasquali, Dhagash Mehta
Abstract
To ensure and monitor large language models (LLMs) reliably, various evaluation metrics have been proposed in the literature. However, there is little research on prescribing a methodology to identify a robust threshold on these metrics even though there are many serious implications of an incorrect choice of the thresholds during deployment of the LLMs. Translating the traditional model risk management (MRM) guidelines within regulated industries such as the financial industry, we propose a step-by-step recipe for picking a threshold for a given LLM evaluation metric. We emphasize that such a methodology should start with identifying the risks of the LLM application under consideration and risk tolerance of the stakeholders. We then propose concrete and statistically rigorous procedures to determine a threshold for the given LLM evaluation metric using available ground-truth data. As a concrete example to demonstrate the proposed methodology at work, we employ it on the Faithfulness metric, as implemented in various publicly available libraries, using the publicly available HaluBench dataset. We also lay a foundation for creating systematic approaches to select thresholds, not only for LLMs but for any GenAI applications.
Complexity vs Empirical Score
- Math Complexity: 4.5/10
- Empirical Rigor: 7.0/10
- Quadrant: Street Traders — empirical and applied, lower math complexity
Why this score: The paper focuses on practical methodology (MRM guidelines, concrete step-by-step recipes) with only moderate statistical/mathematical content, while demonstrating high empirical rigor through a concrete experiment using public datasets (HaluBench) and implementation of the threshold selection procedure.
Research Flowchart
flowchart TD A["<b>Research Goal</b><br>Propose a robust methodology for threshold<br>selection for LLM evaluation metrics"] --> B["<b>Methodology Framework</b><br>Translating MRM guidelines<br>into a step-by-step recipe"] B --> C["<b>Phase 1: Risk Identification</b><br>Identify application risks &<br>stakeholder risk tolerance"] C --> D["<b>Phase 2: Threshold Determination</b><br>Apply statistical procedures on<br>ground-truth data (e.g., HaluBench)"] D --> E["<b>Application Example</b><br>Apply methodology to<br>Faithfulness metric"] E --> F["<b>Key Outcomes</b><ul><li>Systematic threshold selection process</li><li>Risk-driven decision framework</li><li>Foundation for GenAI applications</li></ul>"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes a step‑by‑step recipe for selecting a robust threshold for a given LLM evaluation metric.
Source sentence
we propose a step-by-step recipe for picking a threshold for a given LLM evaluation metric.
Assumes 2
- The Z‑score method assumes the evaluation metric scores follow a normal distribution or that the sample is large enough for the Central Limit Theorem to apply. statistical
Source sentence
The Z‑score approach is valid when the data follows a normal distribution or when the sample size is sufficiently large for the Central Limit Theorem to ensure approximate normality.
- Thresholds for bimodal (or multimodal) distributed scores can be identified by locating the midpoint between two modes. statistical
Source sentence
Determining a threshold for bimodal (or multimodal) distributed scores involves identifying the midpoint that effectively separates the two modes.
Data
- HaluBench benchmark public
Source sentence
we focus on a publicly available dataset called HaluBench
- UpTrain benchmark public
Source sentence
For UpTrain, we define gpt-4o-mini as our LLM model and set the evaluation metric to FACTUAL ACCURACY for faithfulness evaluation.
- DeepEval benchmark public
Source sentence
For the DeepEval framework, we specify our LLM model as gpt-4o-mini and set include_reason = True
- RAGAS benchmark public
Source sentence
In RAGAS, we used the default LLM hyperparameters such as temperature and top-p.
Costs
Not applicable
Finds 4
- The KDE method performs reasonably well at lower confidence levels (e.g., 80%) but fails to align with desired thresholds as confidence levels increase. performance
Source sentence
however, as the confidence level increases, the thresholds derived from the KDE method often fail to align with the desired levels.
- Empirical recall aligns with desired class separation at low confidence levels but becomes overly conservative at higher levels, sometimes defaulting to zero. performance
Source sentence
at higher confidence levels, the thresholds derived using this method exhibit overly conservative behavior, particularly for the DeepEval dataset.
- All conformal prediction methods achieve coverage rates that converge to the specified confidence levels, confirming the framework’s validity. statistical
Source sentence
confidence levels, validating the conformal prediction framework.
- GAM and polynomial logistic regression consistently achieve the highest AUC‑ROC and average precision scores across datasets. performance
Source sentence
Overall, the best scores are consistently achieved either by GAM or polynomial logistic regression, highlighting their effectiveness in this context.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 27 related papers). Coverage is partial, so this is neither replication nor refutation.
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.