Eight papers on April 22 probe limits in AI training, inference, and deployment

Wednesday's batch of eight papers divides roughly into three themes: fundamental limits on what training methods can achieve, advances in model compression for constrained hardware, and specialized architectures for industrial or embedded tasks.

Limits and ceilings. Two papers document cases where no amount of architectural refinement can overcome a structural problem. A rough-set analysis of the Derm7pt dermoscopy dataset found that 16 percent of concept profiles are internally inconsistent, placing a hard 92-percent accuracy ceiling on Concept Bottleneck Models regardless of how they are trained. Separately, researchers showed that dataset distillation methods lose their advantage when soft labels are used during training, because soft labels obscure differences in data quality and make curated subsets perform no better than random ones. On a more optimistic note, a switching-system analysis of Q-value iteration found that the algorithm identifies optimal actions in finite time, with convergence rates that may exceed the classical discount-factor bound. A related theoretical finding concerns world model fidelity: long-horizon failures stem from misaligned latent geometry rather than weak dynamics prediction, and geometric regularization addresses the gap.

Compression and edge deployment. Two papers focus on shrinking models for resource-constrained settings. The QSLM framework automates quantization of spike-driven language models, reducing memory footprint by 86.5 percent while preserving task accuracy on embedded hardware. On the hardware side, compact 1D convolutional networks running on low-power FPGAs enable vibration-based gesture recognition on furniture surfaces, consuming under 1.2 mJ per inference and requiring no complex preprocessing.

Specialized architectures and routing. A vision-language model called AD-Copilot was trained specifically for factory inspection, comparing paired industrial images to detect subtle defects and reaching 82.3 percent accuracy on industrial anomaly benchmarks, outperforming both general-purpose models and human inspectors on the evaluated tasks. In the engineering category, a complexity study of routing in orbital federated learning maps which scheduling scenarios admit polynomial-time solutions and which cross into NP-hard territory, providing clearer guidance for system designers.