366 articles tagged with #interpretability. AI-curated summaries with sentiment analysis and key takeaways from 50+ sources.
Researchers introduce DeALOG, a decentralized multi-agent framework that uses specialized AI agents coordinating through a shared natural-language log to answer complex questions spanning text, tables, and images. The system demonstrates competitive performance on multiple benchmarks while improving robustness through collaborative verification without central control.
Researchers introduce INFORM, an interpretability framework for analyzing multi-expert LLM orchestration systems, revealing that frequently routed experts often serve as structural hubs with minimal functional impact while sparsely selected experts can be critically important. The study challenges conventional assumptions about expert importance in collaborative AI systems and provides tools for understanding opaque decision-making in complex model architectures.
Researchers tested whether massive activations in transformer neural networks are architectural artifacts or functionally necessary by creating a specialized architecture (Ledger Residuals) that separates the residual stream into scratch and protected channels. The model rebuilt the massive activation pattern in the protected channel regardless, suggesting these outliers serve a functional purpose rather than being removable byproducts of design constraints.
Researchers have developed a framework using Sparse Autoencoders to extract and interpret visual, textual, and multimodal concepts from Vision Language Models, achieving 45% improvement in visual concept quality compared to existing methods. This advancement provides structured insights into how VLMs process joint image-text information, addressing a critical gap in AI interpretability research.
Researchers discovered that Dutch language models exhibit coherence illusions similar to humans, where incoherent text appears coherent when a matching distractor precedes it. Using surprisal, attention entropy, and energy metrics, they identified shared mechanisms underlying these illusions across different model architectures.
Researchers propose Graph-of-Differences (GoD), a novel approach to medical image re-identification that grounds patient matching in explicit anatomical structures rather than arbitrary spatial features. The method demonstrates significant accuracy improvements on fundus and chest X-ray images while providing clinically auditable explanations tied to named anatomical regions.
ARCO introduces an adaptive rubric framework that enables large language model agents to receive step-level interpretable rewards during multi-step reasoning tasks. By jointly evolving the reward rubric and policy through co-training, the method achieves stronger performance on question-answering benchmarks while providing explainable feedback that clarifies why each step in a trajectory succeeds or fails.
Researchers introduce w²VLA, a modular Vision-Language-Action model that separates declarative knowledge (concepts and semantics) from procedural knowledge (task execution) to enable zero-shot skill transfer across novel objects. The approach addresses brittleness in current VLA systems by restructuring information flow through compositional modulation rather than opaque transformer processing, achieving superior generalization beyond object-specific training.
Researchers demonstrate that protein contact prediction can be extracted from language model attention heads in a single forward pass, outperforming the computationally expensive Categorical Jacobian method on clean test data. The findings reveal that contact information is concentrated in a small subset of attention heads, requiring only 10 labeled proteins for head selection.
Researchers propose Dys-XAI, an influence-based explainability framework that makes deep learning predictions for dysarthria severity assessment interpretable by linking decisions to similar training examples. The method uses gradient-based influence approximations to identify supportive and competing samples, with validation experiments confirming that removing influential samples systematically alters predictions, addressing a critical gap between model performance and clinical adoptability.
Researchers introduce GLARE, an LLM-based interactive system that translates natural language questions into SQL queries to make global explanations from AI vision models more accessible and usable. The system bridges the gap between complex, static explanation artifacts and human-centered interpretability by enabling users to ask targeted questions about model behavior without needing technical expertise.
Researchers introduce Evolving Programmatic Bottlenecks (EPB), a novel framework for interpreting Neural Combinatorial Optimization models by distilling them into human-readable program portfolios. The method uses large language models to autonomously evolve interpretable programs while maintaining performance comparable to the original black-box models, addressing a critical gap in AI explainability for complex sequential decision-making systems.
Researchers developed a three-stage pipeline to automatically extract skill libraries from computer-using agent interaction data, achieving high readability (95% purity on labeled benchmarks) but failing to improve downstream policy performance across domains. The study reveals that while trajectory mining can expose interpretable skill structure, current technical limitations prevent reliable cross-domain transfer improvements.
Researchers developed an interpretable deep learning framework using EfficientNet-B0 and attention mechanisms to classify sperm morphology for male infertility diagnosis. The model achieves 90-94% accuracy on public datasets while providing visual explanations through Grad-CAM++ visualizations, addressing the clinical adoption barrier of traditional black-box AI models.
Researchers propose a cross-attention attribution method for style-captioned text-to-speech systems, adapting the DAAM framework to speech diffusion models for the first time. Analysis of 3,600 style-caption and text combinations reveals how individual words influence acoustic output, showing that style tokens condition voice characteristics globally while peaking in early generation steps and deep network layers.
Researchers introduce TreeTracer, a visual analytics tool that detects hidden biases in large language models by aggregating hundreds of stochastic generations into comparable hierarchical structures. The tool successfully exposes representational harms in LLMs like GPT-2 XL and demonstrates that standard single-output auditing methods fail to capture biases buried in lower-probability generation branches.
Researchers demonstrate that Information Lattice Learning (ILL), a technique for discovering interpretable rules in signals, naturally aligns with probabilistic graphical model structure learning when applied to probability distributions. The work reveals that ILL rules correspond to marginal constraints over abstracted variables, with maximum-entropy reconstruction creating constraint-based factor graphs rather than traditional Bayesian networks.
Researchers propose enhanced neural additive and basis models (NAM/NBM) that incorporate feature selection mechanisms to improve computational efficiency and interpretability of deep neural networks. The advancement enables these models to handle high-dimensional datasets and capture feature interactions while reducing training costs and model sizes compared to traditional approaches.
Researchers demonstrate that DiffusionGemma, a diffusion-based language model, maintains reasonable interpretability despite performing computations in latent space by mapping information through interpretable token bottlenecks. While algorithmic transparency remains more challenging than autoregressive models, the approach achieves comparable monitorability performance, suggesting diffusion models can be adequately transparent for safety and debugging purposes.
Researchers introduce Chain of Operators (CHOP), a framework that enables frozen neural operator models to handle out-of-distribution tasks without fine-tuning by constructing chains of explicit mathematical transformations. The approach demonstrates improved generalization across different PDE families while maintaining interpretability.
This arXiv survey examines explainable AI (XAI) methods applied to Answer Set Programming (ASP), a symbolic AI approach used for declarative reasoning. The paper catalogs existing explanation approaches and tools while identifying gaps in coverage across different user scenarios, establishing a foundation for future XAI research in logic-based systems.
Researchers introduce MARIC, a multi-agent framework that improves image classification by decomposing the task into collaborative reasoning steps rather than relying on single-pass vision language models. The approach uses specialized agents to analyze different visual dimensions and synthesize findings, demonstrating superior performance across multiple benchmark datasets.
Researchers introduce Neural FOXP2, a technique that identifies and steers language-specific neurons in large language models to shift their default behavior from English to other languages like Hindi or Spanish. The method uses sparse autoencoders and spectral analysis to isolate a compact set of control circuits governing language preference, enabling safer, more targeted manipulation of multilingual model behavior.
SemantiClean is a modular framework that extracts semantic signals from e-commerce session data to predict purchase intent and customer behavior while prioritizing auditability and reproducibility over raw predictive accuracy. The system uses a predefined library of 24 behavioral elements organized across four layers and implements safeguards against signal inflation, representing a shift toward transparent, governance-focused AI systems over conventional black-box optimizers.
Researchers demonstrate that sparsified Kolmogorov-Arnold Networks (KANs) can perform quantum state tomography while remaining interpretable, recovering physical structure without superior performance. The method identifies relevant Pauli measurements from 63 total measurements and reveals internal pathways consistent with known quantum mechanics, validating that neural models can be audited against established physics.
