PTS Pivotal Token Search

Guides you through PTS (Pivotal Token Search), an open-source framework for finding the reasoning events that change whether a language model solves a task. The workflow explains the three scales PTS searches (latent workspace meta-tokens, emitted pivotal tokens, and reasoning sentences called thought anchors), installs PTS from its README after asking you, and shows how to run a search on a task with a small local model. It explains how to read the output: probability shifts, the shared category labels, and why latent readout scores are not comparable to probability changes. It covers the hosted visualizer, linking events into chains, and exporting results as datasets for preference tuning or steering. It states the limits plainly: searches are expensive because they sample many continuations, results are observational unless you run an intervention with controls, and findings depend on the task, the model and the sampling settings. This plugin contains instructions only; everything runs on your own machine.