PTS Visualizer — Pivotal Token Search
15
26 commits
2 linked in READMEs
updated Jul 15, 2026
Interactive visualization for Pivotal Token Search: the reasoning events that shift a language model's probability of solving a task, at three representational scales — latent workspace meta-tokens, emitted pivotal tokens, and sentence-level thought anchors — as a single kind of object.
Per-scale event counts, causal-link count, and separate distributions for emitted probability deltas and latent readout scores (they are different quantities and are never binned together).
Every event in context, filterable by scale, category, valence, score, and (for latent events) layer. Latent readouts are shown as readout scores, never as probability deltas.
The latent → token → sentence → outcome graph for a query, with nodes shaped by scale and edges from the recorded causal links.
t-SNE of event embeddings, colored by category or impact.
All scales on one shared generation axis: latent meta-tokens, pivotal tokens, thought-anchor sentences, and the resulting success-probability curve — plus a workspace heatmap of meta-token readout scores by position.
Built for the unified PTS reasoning-event datasets, which drive every view:
codelion/Qwen3-0.6B-ptscodelion/DeepSeek-R1-Distill-Qwen-1.5B-ptsAny PTS file can also be loaded by pasting its dataset id or uploading a JSONL (legacy pivotal-token and thought-anchor files are upgraded on the fly). Steering-vector datasets are a separate downstream artifact for OptiLLM and are not the focus here.
# Clone the repository
git clone https://github.com/codelion/pts
cd pts/visualizer
# Install dependencies
pip install -r requirements.txt
# Run the app
python app.py
If you use this tool in your research, please cite:
@software{pts,
title = {PTS: Pivotal Token Search},
author = {Asankhaya Sharma},
year = {2025},
publisher = {GitHub},
url = {https://github.com/codelion/pts}
}
PTS Visualizer — Pivotal Token Search
15
26 commits
2 linked in READMEs
updated Jul 15, 2026
Interactive visualization for Pivotal Token Search: the reasoning events that shift a language model's probability of solving a task, at three representational scales — latent workspace meta-tokens, emitted pivotal tokens, and sentence-level thought anchors — as a single kind of object.
Per-scale event counts, causal-link count, and separate distributions for emitted probability deltas and latent readout scores (they are different quantities and are never binned together).
Every event in context, filterable by scale, category, valence, score, and (for latent events) layer. Latent readouts are shown as readout scores, never as probability deltas.
The latent → token → sentence → outcome graph for a query, with nodes shaped by scale and edges from the recorded causal links.
t-SNE of event embeddings, colored by category or impact.
All scales on one shared generation axis: latent meta-tokens, pivotal tokens, thought-anchor sentences, and the resulting success-probability curve — plus a workspace heatmap of meta-token readout scores by position.
Built for the unified PTS reasoning-event datasets, which drive every view:
codelion/Qwen3-0.6B-ptscodelion/DeepSeek-R1-Distill-Qwen-1.5B-ptsAny PTS file can also be loaded by pasting its dataset id or uploading a JSONL (legacy pivotal-token and thought-anchor files are upgraded on the fly). Steering-vector datasets are a separate downstream artifact for OptiLLM and are not the focus here.
# Clone the repository
git clone https://github.com/codelion/pts
cd pts/visualizer
# Install dependencies
pip install -r requirements.txt
# Run the app
python app.py
If you use this tool in your research, please cite:
@software{pts,
title = {PTS: Pivotal Token Search},
author = {Asankhaya Sharma},
year = {2025},
publisher = {GitHub},
url = {https://github.com/codelion/pts}
}