TRUE: A Trustworthy Unified Explanation Framework for Large Language Model Reasoning

Researchers introduce TRUE (Trustworthy Unified Explanation Framework), a new methodology for interpreting and verifying the reasoning processes of large language models across multiple analytical levels. The framework combines executable verification, structural analysis, and causal failure mode detection to provide transparent insights into LLM decision-making, addressing critical gaps in current interpretability methods.

The interpretability of large language models has emerged as a critical research challenge as these systems assume increasingly consequential roles in enterprise and consumer applications. The TRUE framework tackles a fundamental limitation in existing explanation methods: they typically operate at single-instance level without revealing broader patterns about reasoning stability or systematic failure modes. This research represents a meaningful advancement in trustworthy AI by proposing a multi-tiered approach that functions simultaneously at instance, local structural, and class levels.

The framework's innovation lies in its three-component architecture. First, it redefines reasoning traces as executable specifications rather than abstract representations, introducing blind execution verification to validate operational integrity. Second, it constructs feasible-region DAGs through structured perturbations, enabling researchers to map the input space where reasoning remains stable and valid. Third, it employs causal failure mode analysis with Shapley value quantification to identify recurring failure patterns and measure their systematic impact across model classes.

For the AI development community, this work addresses pressing concerns about LLM reliability and auditability. As organizations deploy LLMs in regulated sectors—finance, healthcare, legal—stakeholders increasingly demand transparent reasoning verification rather than opaque decision outputs. The framework's ability to characterize failure modes with quantified importance could facilitate better model selection and refinement strategies.

Looking forward, TRUE's multi-level verification approach may influence how AI systems undergo compliance testing and safety audits. The methodology's emphasis on executable verification could become a standard requirement in high-stakes applications, potentially shaping how language model architectures are evaluated and improved. The research demonstrates a path toward more defensible and interpretable AI systems, though widespread adoption depends on integration into existing model development workflows.

• →TRUE framework enables multi-level verification of LLM reasoning through executable specifications, structural DAG modeling, and causal failure analysis.
• →The approach identifies recurring failure patterns at the class level and quantifies their impact using Shapley values for systematic understanding.
• →Feasible-region DAG construction reveals which input neighborhoods maintain reasoning stability, advancing local interpretability beyond single-instance analysis.
• →The framework addresses critical gaps in current explanation methods by providing verifiable structural insights rather than superficial reasoning descriptions.
• →Integration of executable verification could establish new standards for LLM evaluation in regulated industries requiring transparent decision-making audits.