ABLE (Attribution-Based Large-model Embedding) introduces a framework for representing large language models by leveraging interpretability space through attribution-based embeddings. It addresses challenges in systematic model comparison by aggregating gradient-based feature attributions to capture model-specific input-sensitivity patterns.
This article explores how animating AI confidence scores and probability distributions with Swift 6 can transform "black box" models into transparent systems. This approach enhances user trust, provides real-time feedback, and aids in debugging by visualizing the AI's "thought process."
Geometry-Lite is a novel prompt-level probe designed to interpret how safety evidence develops across layers in large language models. It analyzes layer-wise margin geometry using various readouts to understand boundary formation, improving safety detection over single-layer probes.
This research proposes a framework for sentence-level interpretability in rubric-based scoring, combining Shapley-value attributions with rationales from large language models (LLMs). It compares fine-tuned language models and prompted LLMs for teaching quality assessment, finding PLMs offer better prediction accuracy despite label compression.
O OpenAI Safety Fellowship é um programa de pesquisa focado na segurança da IA, abordando aspectos críticos como robustez, interpretabilidade e alinhamento de valores humanos. O texto detalha seus objetivos e componentes técnicos, como treinamento adversarial e técnicas de explicabilidade.
This position paper argues that large language model (LLM) reasoning should be studied as latent-state trajectory formation rather than faithful surface chain-of-thought (CoT). It formalizes three competing hypotheses regarding the primary object of reasoning, impacting claims about faithfulness, interpretability, and benchmarks.
OceanCBM is the first concept bottleneck model (CBM) for spatiotemporal prediction and mechanistic interrogation of ocean dynamics. It predicts mixed layer heat content, a precursor to marine heatwaves, using mixed supervision and prescribed geophysical fluid dynamics concepts to ensure fidelity to ground-truth physics.
This guide explains the critical role of a confidence bar in AI applications for building user trust and enhancing transparency about model predictions. It details how to build a reusable, animated confidence bar using Swift 6 and SwiftUI.
This content explores the fascinating area of translating the internal processes or "thoughts" of an artificial intelligence model like Claude into understandable language. It investigates how the complex operations of AI can be interpreted and expressed to better understand its reasoning.
The startup Goodfire has released Silico, a new mechanistic interpretability tool that allows researchers to debug and adjust LLM parameters during training. This provides model makers with more fine-grained control over AI development.
The paper introduces challenges in understanding annotator disagreement regarding AI safety policies, which can arise from operational failures, policy ambiguity, or value pluralism. It highlights the difficulty of discerning the root causes of these disagreements and the unreliability of self-reported reasoning from annotators.
This research introduces a physics-based interpretability framework to assess what physics data-driven MoCap-to-radar models learn. It finds that low reconstruction error doesn't guarantee physical consistency, and temporal attention is critical for transformer-based models to learn the underlying physics.
This study introduces yvsoucom-iterkit, a deterministic and log-driven automated machine learning framework for interpretable pipeline optimization in healthcare risk prediction. It enables reproducible analysis of pipeline components, revealing that performance is driven by a small subset of interacting elements like augmentation, model choice, and imbalance handling.
This paper investigates how LLMs learn in-context, using a graph random-walk task to explore whether they pattern-match or infer latent structure. It reveals that neither account alone is sufficient, presenting evidence of simultaneous encoding of graph topologies and causal interventions.
SymptomWise é um framework que aprimora a análise de sintomas por IA, separando a compreensão da linguagem do raciocínio diagnóstico para aumentar a confiabilidade e rastreabilidade. Ele utiliza conhecimento médico especializado e inferência determinística, empregando LLMs apenas para extração de sintomas e explicações, não para o diagnóstico em si.
O artigo apresenta PRISM, uma estrutura para Reinforcement Learning que fundamenta as decisões de agentes em conceitos discretos e causalmente validados, usando-os como interface de transferência zero-shot. Ele demonstra que esses conceitos impulsionam diretamente o comportamento do agente e que a importância de um conceito pode ser dissociada de sua frequência de uso.
Gemma Scope 2 foi lançado, disponibilizando ferramentas abertas de interpretabilidade para a família Gemma 3. Isso visa ajudar a comunidade de segurança de IA a aprofundar a compreensão do comportamento de modelos de linguagem complexos.