AMMA-UQ: Introducing Adaptive Multi-Modal Attention Mechanism for Uncertainty Quantification in Black-Box Large Language Models

AMMA-UQ is an uncertainty quantification framework for black-box large language models. Through three key innovations—adaptive sampling, multi-modal similarity fusion, and attention aggregation—it reduces sample usage by 48.7% while improving the accuracy of confidence assessment. This framework addresses the failure of traditional uncertainty estimation methods in black-box scenarios, providing a more reliable basis for confidence judgment in LLM applications.

Background: Why Do Black-Box LLMs Need Uncertainty Quantification?

Large Language Models (LLMs) face a core challenge in practical applications: there is often a discrepancy between the model's output confidence and its actual accuracy. Users cannot directly access the model's internal logits or probability distributions (black-box scenario), which renders traditional uncertainty estimation methods ineffective. When models "confidently make mistakes", the consequences can be catastrophic—from medical diagnosis errors to financial decision-making failures.

The Consistency Hypothesis provides a theoretical basis for this: if multiple sampled outputs of the model for the same question are highly consistent with each other, the answer is more likely to be correct; conversely, if the outputs fluctuate significantly, the uncertainty is higher. However, existing methods often use fixed sampling strategies and simple similarity metrics, failing to fully utilize multi-dimensional signals, leading to low efficiency or inaccurate estimation.

Overview of the AMMA-UQ Framework

AMMA-UQ (Adaptive Multi-Modal Attention for Uncertainty Quantification) is an innovative framework extended from the Consistency Hypothesis work by Xiao et al. (UAI 2025). It addresses the uncertainty quantification problem of black-box LLMs and proposes three key technical innovations, aiming to obtain more accurate confidence estimates with fewer sampling times.

The core idea of the framework is: uncertainty quantification should not be a simple "look at discrepancies after multiple samplings", but a refined process of intelligently fusing multi-dimensional signals and dynamically adjusting sampling strategies.

Key Innovation 1: Adaptive Sampling Strategy

Traditional methods usually use a fixed number of samplings (e.g., 10 or 20 times) regardless of the actual complexity of the problem. AMMA-UQ breaks this paradigm and introduces a dynamic sampling mechanism based on entropy stabilization.

Specifically, the framework continuously monitors changes in the entropy value of the output distribution during sampling. When the entropy value stabilizes—i.e., new samplings no longer significantly change the characteristics of the output distribution—sampling stops automatically. This strategy brings significant efficiency improvements: experiments show that compared to fixed sampling, AMMA-UQ reduces sample requirements by an average of 48.7% while maintaining or even improving quantification accuracy.

The significance of this adaptive mechanism is that it allows more reasonable allocation of computing resources: fewer samplings for simple problems and more for complex ones, instead of a one-size-fits-all approach.

Key Innovation 2: Multi-Modal Similarity Fusion

A single similarity metric is often insufficient to fully capture the differences in text outputs. AMMA-UQ innovatively fuses three complementary similarity signals:

Lexical Similarity: Based on traditional metrics like ROUGE-L, it measures the degree of lexical overlap between output texts. These metrics are computationally efficient and sensitive to surface-level changes.

Semantic Similarity: Using pre-trained models like SBERT, it captures the distance of outputs in the semantic vector space. This method can understand outputs that are semantically equivalent but expressed differently.

Task-Specific Similarity: Similarity metrics designed for specific tasks, such as answer correctness judgment in question-answering tasks or information coverage evaluation in summarization tasks.

By fusing these three types of signals, AMMA-UQ can construct a more robust and comprehensive representation of output similarity than a single metric.

Key Innovation 3: Attention Mechanism Aggregation

After obtaining the multi-modal similarity matrix, AMMA-UQ introduces an attention mechanism to learn discriminative weight allocation.

Unlike simple averaging or weighted summation, the attention layer can automatically learn the importance differences between different sample pairs. In the specific implementation, the framework uses an attention network with a hidden layer dimension of 64, where the input is pairwise similarity features and the output is aggregated weights. This data-driven aggregation method allows the framework to adaptively adjust the weights of different similarity signals and the contribution of different sample pairs according to specific tasks and data characteristics.

Experimental Evaluation and Performance

AMMA-UQ has been validated on multiple standard datasets, including tasks like CoQA (Conversational Question Answering). The evaluation metrics used are AUROC (Area Under the Receiver Operating Characteristic Curve) and ECE (Expected Calibration Error), which measure uncertainty ranking ability and calibration degree respectively.

Experimental results show that AMMA-UQ outperforms baseline methods in both metrics, proving the synergistic effect of adaptive sampling, multi-modal fusion, and attention aggregation. More importantly, these improvements are achieved while significantly reducing computational overhead (nearly half the sample reduction), reflecting the practical value of the framework.

Practical Significance and Application Prospects

The proposal of AMMA-UQ has multiple implications for the LLM application ecosystem:

For API users: Reliable uncertainty estimates can be obtained without accessing the model's internal state, helping to build safer LLM applications such as risk warning and human-machine collaborative decision-making scenarios.

For resource-constrained environments: Adaptive sampling significantly reduces inference costs, making uncertainty quantification feasible in edge devices or high-frequency call scenarios.

For the research field: This framework provides a new technical path for uncertainty research of black-box models, and the ideas of attention aggregation and multi-modal fusion can be transferred to other related tasks.

Conclusion

AMMA-UQ represents an important progress in the field of uncertainty quantification for black-box LLMs. Through the triple innovations of adaptive sampling, multi-modal similarity fusion, and attention mechanism aggregation, it achieves an excellent balance between efficiency and accuracy. As LLMs are increasingly applied in key fields, such technologies that can "know what they don't know" will become more and more important.