Tag
This paper introduces Reasoning Exposure Prompting (REP), a method that uses shadow-model demonstrations in code-like formats to elicit hidden reasoning traces from LLMs, showing that interface-level trace hiding is insufficient to prevent extraction of useful reasoning signals.
This paper presents the first model extraction attack on graph classification under strict black-box constraints, exploiting subgraph explanations to estimate decision boundaries. The findings reveal that mandated explainability interfaces create exploitable security vulnerabilities in Graph Neural Network services.
A user reports that the 3.6 GB Gemma 4 e4b model extracted from Google AI Edge Gallery on Android outperforms larger 3.7 GB Unsloth versions and community ports, raising questions about hidden optimizations.