TSFM-Lens
TSFM-Lens is a pipeline for comparative mechanistic interpretability of time-series foundation models: it puts models with very different architectures on equal grounds. It asks which concepts they learn, whether those concepts are shared across models, whether they causally affect the forecast, and where in depth the forecast forms. Every claim is registered on a development corpus and tested once on a fresh sealed private corpus (held-out confirmation).
Headline result. On a fresh sealed private corpus, different architectures select the same inputs for a concept (concept transfer 20/20, on four private draws), while their causal machinery is model-specific (no concept spans more than three of seven models); 10 single-feature causal claims confirm.
Example reports (interactive, self-contained HTML)
- Seven-model report, with held-out confirmation TimesFM 2.5, Chronos-2, Sundial, Chronos-Bolt, Timer, Time-MoE and Chronos-T5-Base on the 965-series v2 dev corpus, all 19 stages, then confirmed once on a 971-series sealed private split. The report is large; allow it a moment to load.
- Four-model concept atlas (earlier run) TimesFM, Chronos-2, Sundial and Chronos-Bolt on the v2 corpus (29 SAE targets), confirmed once on the v2 private split. Superseded by the seven-model run, kept unchanged.
Code and quickstart
- GitHub repository
- Install and quickstartRun the whole pipeline on mock models on CPU in minutes.