AIP-Speech introduces an inference-only benchmark for evaluating how Speech/Audio LLMs degrade under controlled background interference. It mixes clean speech with real-world background recordings at different SNRs and analyzes degradation using acoustic descriptors such as ZCR, spectral centroid, and spectral rolloff. Across five open Speech LLMs, spectral descriptors show the strongest associations with performance degradation across ASR and KWS. The paper also examines accent-wise robustness and whether prompting can improve resilience to background interference.

The paper studies whether TopoLoss, a lightweight training-time spatial regularizer, improves mechanistic interpretability in ViTs by organizing neurons into spatially coherent causal circuits. On ImageNet-100, TopoLoss substantially improves causal sufficiency (2.79× vs. 1.68× baseline) and reduces SAE feature complexity, but does not improve neuron-level monosemanticity, suggesting it reorganizes information at the circuit level rather than disentangling individual neurons.

The paper proposes an end-to-end deep learning pipeline for liver cancer detection and HCC staging from CT/MRI images, combining YOLOv8 for liver detection, MedSAM for segmentation, and ResNet-18/EfficientNet-B2 for T1–T4 stage classification. On TCGA-LIHC, the pipeline achieved strong detection/segmentation performance, while EfficientNet-B2 performed best with transfer learning; however, the relatively low classification accuracy highlights limitations in reliable tumor-stage prediction.