HyperLens: New Probe Quantifies Cognitive Effort in LLMs

Researchers have introduced HyperLens, a high-resolution probe that traces confidence trajectories in large language models (LLMs) to quantify cognitive effort during inference. The work, published on arXiv, identifies an intrinsic magnification mechanism in transformer architectures where deeper layers amplify small changes in layer-wise confidence, producing a fine-grained confidence trajectory. HyperLens reveals a consistent divergence in these trajectories that separates complex from simple tasks, which the authors abstract into a quantitative cognitive effort metric. Their analysis shows that complex tasks consistently require higher cognitive effort, and they provide a mechanistic diagnosis of a common failure mode. The study aims to improve understanding of LLM inference dynamics, which have been poorly understood due to limited resolution of existing analysis tools.