layer-wise-analysis

This paper introduces a diversity-aware, layer-wise scoring method for KV cache eviction in large language models, incorporating attention dispersion and redundancy to improve performance on LongBench datasets.

LaRA is a layer-wise representation analysis framework that detects data contamination in RL post-trained LLMs by measuring geometric deviations across model layers, outperforming output-level baselines.

Introduces Geometry-Lite, a compact probe that analyzes layer-wise margin geometry to interpret how safety evidence forms across layers in LLMs, improving over single-layer probes while maintaining interpretability.