This research identifies a bias in semi-structured clinical interviews where models trained on interviewer prompts achieve high classification scores without using participant language. This has implications for AI models that rely on such data, highlighting the need to focus on participant language to ensure accurate and interpretable results.

Researchers proposed a method to improve AI model interpretability in farm monitoring, specifically for sheep facial pain recognition. They found that current language-grounded explanations are not effective and introduced a concept bottleneck to address this issue, resulting in improved accuracy and demonstrable learned concepts.

This research paper introduces the concept of xeno-interpretability, which examines the internal representations of large language models (LLMs) that cannot be adequately expressed in human terms. The study highlights the limitations of current interpretability methods and proposes a new approach to understanding model-native representations.

A new method for model interpretability, ADORE, is released as an open-source Python package. It uses first- and second-order derivatives to capture nonlinear feature interactions and provides a unified analytical framework for global feature importance and local sample contributions. This improves upon existing methods like LIME and SHAP, making it a useful tool for model explainability and decision-making in AI development.

Researchers analyzed the behavior of Transformers and state-space models through the lens of layer importance, revealing fundamental differences between the two families. The study decomposed layer importance into necessity and plasticity, showing that these concepts align in state-space models but anti-align in transformers. This has implications for model compression, fine-tuning, and interpretability.

Researchers explore the possibility of using graph-based representations to improve the interpretability of Large Language Model (LLM) Natural Language Inference (NLI) systems. They introduce a pipeline that decomposes input text into atomic propositions and represents them as graphs, achieving 89.7% accuracy on the SNLI dataset. The study highlights the trade-off between accuracy and interpretability, with the graph-based approach trailing behind its text-based counterpart by 1.9 points. The authors also demonstrate that combining graph and text modalities can achieve higher accuracy.

This paper proposes a method to integrate Large Language Models (LLMs) with the Analytic Hierarchy Process (AHP) for transparent multi-criteria decision-making. The authors create a benchmark and an end-to-end approach to enable LLMs to perform the complete AHP workflow, improving alignment with expert judgments in real-world decision problems.