A newly posted arXiv paper argues that current AI systems risk outpacing human understanding and control, and that fresh mathematical work is needed to make them legible, steerable, and cooperative. The author structures the call by mathematical subfield so researchers can identify where their expertise applies, and frames it as an open invitation to the mathematics community.
A new arXiv paper examines the internal representations of Whisper, an automatic speech recognition model, by applying sparse autoencoders to its encodings. The authors note that interpretability research has focused mostly on text-based transformers, leaving speech systems comparatively unstudied. Their work aims to make the features learned by Whisper's encoder more understandable.
A new arXiv preprint examines whether supportive empathy in large language models can be controlled through activation steering, as has been done for traits like honesty and refusal. Using the EPITOME dataset, the authors analyze the geometry of the underlying mechanisms and find that empathy does not map cleanly onto a single controllable direction, indicating a multi-axial structure. The work also looks at how persona settings influence these empathy-related representations.
A new arXiv paper introduces IMPACT-VLA, a method for tracing how much each input modality — camera images, proprioceptive state, and language instructions — contributes to a vision-language-action policy's decisions at different points during task execution. The approach relies on counterfactual trajectories to isolate the effect of individual inputs, addressing the difficulty of interpreting these multimodal robot policies. The abstract excerpt does not detail experimental results or benchmarks.
A new arXiv paper introduces MACCHIATO, a specialized training algorithm for ReLU-based multilayer perceptrons that solve Boolean tasks. The method is designed to produce networks that are certifiably interpretable while guaranteeing correct generalization across the full truth table. The authors frame the work as a response to the growing gap between AI capability and the ability to explain model behavior.
Anthropic researchers published a paper outlining a mathematical approach to reverse-engineering how transformer models compute internally, treating attention heads and MLP layers as composable circuits. The framework aims to make the internal mechanisms of these models more tractable to study and explain. It is intended as a foundation for interpretability work rather than a description of any specific deployed system.
A new arXiv paper argues that an AI system's output should not be read as a description of a fact or world state, but as an engineered representation. The authors propose a semantic framework for describing such systems so their representations can be checked for correctness. The work is a conceptual contribution rather than a model release or benchmark.
Researchers present CT-SAFR, a multi-layered verification framework intended to make chain-of-thought reasoning by large language models safe and interpretable in autonomous robots. The framework responds to findings that reasoning models often do not verbalize their true decision processes, adding verification layers to support trustworthy AI-driven robotic decisions.
A newly released arXiv paper addresses how to measure the confidence behind agentic AI actions, arguing this is essential as such systems enter safety-critical applications. The authors note that agentic workflows fail in more complex ways than traditional machine learning systems and propose deriving calibrated confidence estimates directly from the model's internal representations.
A position paper examines how large language models generate natural-language accounts of their own decisions, a practice known as self-explanation. The authors argue that current explanations are often merely plausible-sounding rather than genuinely useful, and they outline what would be needed to make them actionable for real-world use.
A new analysis examines the ideas behind GPT-6 Astra, focusing on transformer architectures that reuse the same blocks across multiple passes rather than adding more layers. It reviews recent work on recurrent depth, where looping blocks can increase effective model depth and enable internal computation that is not exposed in the visible output. The piece also discusses hidden chains of thought and what this implies for interpreting model reasoning.