mllm-shap: A Shapley Value Explainability Platform for Text-Audio Multimodal Large Language Models

Researchers introduce mllm-shap, an open-source framework that extends Shapley Value explainability techniques to multimodal large language models processing text and audio inputs simultaneously. The platform addresses three technical challenges unique to multimodal systems and implements five estimation strategies, with a novel phonetic alignment technique reducing computational complexity by 10-50x.

mllm-shap represents a significant advancement in AI interpretability, addressing a growing gap in explainability tools as multimodal AI systems become increasingly prevalent. While Shapley Values have proven effective for explaining text-based language models, extending these techniques to joint text-audio processing introduces substantial technical complexity. The framework tackles this by introducing modality-aware coalition masking that handles the fundamentally different characteristics of discrete text tokens and continuous audio frames, plus multi-turn conversation tracking to maintain contextual awareness across dialogue states.

The innovation gains particular relevance as multimodal AI applications proliferate across transcription, voice assistance, and content analysis. Current production systems often operate as black boxes, making this explainability layer valuable for enterprise adoption, regulatory compliance, and debugging model failures. The phonetic alignment-based token grouping technique demonstrates thoughtful engineering—reducing computational overhead by an order of magnitude enables practical deployment of Shapley-based explanations on realistic audio lengths, a barrier that previously existed.

The framework's interactive web GUI democratizes interpretability work beyond researchers with computational resources. For developers building multimodal systems, understanding which input components drive model decisions becomes critical for quality assurance and trust. The five estimation strategies, particularly the Complementary Contributions estimator with Neyman-optimal allocation, provide practitioners options suited to different accuracy-speed tradeoffs.

Looking forward, widespread adoption of mllm-shap could accelerate responsible AI development in multimodal space. The main limitation involves computational requirements remaining substantial despite optimizations, potentially restricting real-time explanation generation. Future work likely targets further efficiency gains and integration with other explainability frameworks.

• →mllm-shap enables Shapley Value explainability for text-audio multimodal LLMs, addressing a previously underserved interpretability gap.
• →Phonetic alignment-based token grouping reduces coalition space by 10-50x, making SV computation feasible for long-form audio content.
• →The framework includes five estimation strategies and an interactive web GUI for visualization, lowering barriers to explainability research.
• →Modality-aware coalition masking handles simultaneous processing of discrete and continuous input types with proper context preservation.
• →First publicly available complete pipeline for SV-based explainability in text-audio MLLMs, supporting reproducible research and enterprise adoption.