StALT: A New Metric to Detect Genuine Reasoning in LLMs

A novel metric named Spatiotemporal Amplitude of Latent Transition (StALT) has been developed by researchers to differentiate authentic internal reasoning from excessive verbosity in large language models (LLMs). This study, which can be found as a preprint on arXiv (2605.01853v1), examines hidden-state transitions throughout various decoding steps and layers. The findings indicate that effective reasoning paths in large reasoning models (LRMs) demonstrate extensive temporal dynamics with focused layer-wise concentration, a trend that is less pronounced in non-reasoning models and knowledge-intensive areas. StALT serves as a training-free statistic that captures temporal variations between adjacent tokens, factoring in within-token layer saliency. This research explores whether lengthy solution traces from LRMs signify meaningful computation or simply overthinking.