Beyond Hallucinations: A Composite Score for Measuring Reliability in Open-Source Large Language Models

Conference: AAAI 2026 arXiv: 2512.24058
Code: https://github.com/rohitsalla/CRS.git
Area: Hallucination Detection Keywords: LLM reliability, calibration, robustness, uncertainty quantification, composite metric

TL;DR

This paper proposes the Composite Reliability Score (CRS), which unifies calibration, robustness, and uncertainty quantification into a single interpretable metric. A systematic evaluation of 10 open-source LLMs across 5 QA datasets reveals that Mistral-8x22B achieves the highest overall reliability (CRS=0.81), and that model size does not directly determine reliability.

Background & Motivation

Background: Open-source LLMs are increasingly deployed in high-stakes domains such as healthcare, law, and finance, yet their reliability remains uncertain—models frequently exhibit overconfidence, fragility to input perturbations, and a lack of clear uncertainty estimates.

Limitations of Prior Work: Existing evaluations are fragmented—focusing either on accuracy alone or on only one of calibration or robustness—and thus fail to comprehensively characterize deployment reliability.

Key Challenge: A single metric (e.g., ECE or accuracy) can mask weaknesses along other dimensions; a model with high accuracy but poor calibration may be more dangerous in high-stakes settings than one with slightly lower accuracy but better calibration.

Goal: Design a unified reliability measurement framework that simultaneously evaluates calibration, robustness, and uncertainty quantification.

Key Insight: Each of the three dimensions is normalized to \([0,1]\) and then combined via a weighted average—a simple yet effective approach for reflecting composite reliability.

Core Idea: \(\text{CRS} = \alpha C + \beta R + \gamma U\), where \(C\) = calibration score, \(R\) = robustness score, \(U\) = uncertainty quantification score.

Method

Overall Architecture

CRS consists of three pillars, each normalized to \([0,1]\) and combined with equal weights (\(\alpha=\beta=\gamma=1/3\)):

Key Designs

1. Calibration (C)

   • Function: Measures how well model confidence aligns with actual accuracy.
   • Mechanism: Based on ECE, converted to a positive score: \(C = \max(0, 1 - \text{ECE}_{model}/\text{ECE}_{max})\).
   • Evaluation: Baseline calibration plus two post-hoc calibration methods (temperature scaling, isotonic regression).

2. Robustness (R)

   • Function: Measures the model's ability to maintain performance under input perturbations.
   • Mechanism: Computes the relative performance drop under three perturbation types (5% typos, back-translation paraphrase, TextFooler adversarial attack): \(R = 1 - \text{Avg Acc Drop}/\text{Avg Acc}_{clean}\).
   • Design Motivation: Using relative rather than absolute accuracy drop enables fair comparison across models with different baseline capabilities.

3. Uncertainty Quantification (U)

   • Function: Evaluates the model's ability to produce high uncertainty when making incorrect predictions.
   • Mechanism: Uncertainty is estimated via MC Dropout (10 forward passes) and a 3-model ensemble; AUROC is used to assess discrimination between correct and incorrect predictions, normalized as \(U = (\text{AUROC} - 0.5)/0.5\). The better of the two methods is taken as the final \(U\) score.

CRS Interpretability Tiers

• CRS ≥ 0.8: High reliability — suitable for deployment with minimal supervision.
• 0.6–0.8: Moderate reliability — requires human oversight.
• < 0.6: Low reliability — unsuitable for safety-critical scenarios.

Key Experimental Results

Main Results (CRS Rankings)

Model	Params	C	R	U	CRS	Tier
Mistral-8x22B	22B	0.91	0.78	0.73	0.81	High
Qwen3-235B	22B	0.84	0.74	0.70	0.76	Moderate
DeepSeek R1	27B	0.87	0.76	0.63	0.75	Moderate
Gemma 2	27B	0.71	0.68	0.71	0.70	Moderate
LLaMA-3-7B	7B	0.16	0.54	0.44	0.57	Low
Falcon-7B	7B	0.00	0.51	0.41	0.52	Low

Calibration Metrics (Baseline)

Model	Avg ECE↓	Avg Brier↓	Avg NLL↓
Mistral-8x22B	0.031	0.128	0.332
Falcon-7B	0.062	0.179	0.566

Key Findings

• Only Mistral-8x22B reaches the "high reliability" tier (CRS ≥ 0.8); most models fall in the moderate range.
• Model size does not determine reliability: the 27B Gemma 2 achieves a CRS of only 0.70, while the 22B Mistral-8x22B reaches 0.81.
• DeepSeek R1 demonstrates strong calibration (\(C=0.87\)) but weak uncertainty quantification (\(U=0.63\)), illustrating that a single dimension cannot represent overall reliability.
• Sensitivity analysis shows that top and bottom model rankings remain stable across different weight configurations.
• LLaMA-3-7B exhibits severely poor calibration (\(C=0.16\)), a weakness that would be obscured by accuracy-only evaluation.

Highlights & Insights

• The three-dimensional unified metric fills a gap in LLM reliability evaluation—providing substantially richer information than accuracy or ECE alone.
• The tiered deployment recommendations (High/Moderate/Low) are directly oriented toward practical applications, offering actionable guidance for practitioners.
• The paper reveals that calibration, robustness, and uncertainty quantification are not simply positively correlated—some models are well-calibrated yet fragile to perturbations, a trade-off that only a composite metric can expose.

Limitations & Future Work

• Normalization has inherent limitations: the calibration score \(C\) is anchored to the worst model's ECE in the evaluation set, making scores incomparable across different evaluation setups.
• Evaluation is limited to QA tasks and does not cover generation, summarization, or dialogue scenarios.
• MC Dropout and ensemble-based uncertainty estimation have limited applicability to architectures without dropout layers, as found in some LLMs.
• The equal weighting scheme (\(1/3, 1/3, 1/3\)) is simple, but the relative importance of the three dimensions should vary across application domains.
• Hallucination detection is not incorporated—despite the title's claim of going "Beyond Hallucinations," the framework does not directly measure hallucination rates.

Related Work & Insights

• vs. GLUE/SuperGLUE: These benchmarks evaluate accuracy but not reliability; this work is complementary.
• vs. single-dimension calibration studies (RestoreCalib, CogCalib, etc.): Prior work focuses on calibration alone, whereas this paper unifies three dimensions.

Rating

• Novelty: ⭐⭐⭐ — The framework is intuitive but not complex; the three-dimension weighted average is relatively straightforward.
• Experimental Thoroughness: ⭐⭐⭐⭐ — 10 models, 5 datasets, three perturbation types, and two uncertainty estimation methods.
• Writing Quality: ⭐⭐⭐⭐ — Formulations are clear and the framework is interpretable.
• Value: ⭐⭐⭐⭐ — Addresses a gap in unified reliability evaluation with strong practical utility.

Additional Notes

• The methodology and experimental design of this work offer useful reference points for related research.
• Future work could validate the generalizability and scalability of the approach across broader scenarios and larger model scales.
• Potential research value exists in combining this framework with recent related work (e.g., intersections with RL/MCTS or multimodal methods).

Related Papers

• [CVPR 2026] PAS: Prelim Attention Score for Detecting Object Hallucinations in Large Vision-Language Models
• [CVPR 2025] ODE: Open-Set Evaluation of Hallucinations in Multimodal Large Language Models
• [AAAI 2026] Verb Mirage: Unveiling and Assessing Verb Concept Hallucinations in Multimodal Large Language Models
• [ACL 2025] Beyond Facts: Evaluating Intent Hallucination in Large Language Models
• [ICML 2026] Instruction Lens Score: Your Instruction Contributes a Powerful Object Hallucination Detector for Multimodal Large Language Models