PathFM
A flow-matching-based time series foundation model that addresses trajectory dependency by generating complete future trajectories rather than predicting only marginal quantiles. It uses step distillation and a shared history attention cache to optimize sampling efficiency.
Timeline (last 90 days)
PathFM uses flow matching to generate complete future paths.
PathFM uses step distillation to compress the number of sampling steps to one or two, greatly reducing the computational overhead of generating multiple paths.
PathFM uses a shared history attention cache to reuse the history encoding across multiple samplings, avoiding repeated computation.
The time for PathFM to generate multiple paths for a single sequence is roughly equivalent to the time of one quantile model query.
PathFM's one- to two-step sampler allows the model to optimize directly for calibration quality during training.
PathFM achieves competitive marginal accuracy on the existing fev-bench and TIME benchmarks.
On fps-bench, PathFM's overall path score is higher than all evaluated external baselines.
In chaotic system experiments, PathFM's flow model is able to maintain long-term variability, whereas strong baseline models quickly collapse and the trajectories they generate tend to degenerate.
All Facts (8)
PathFM uses flow matching to generate complete future paths.
50%UnverifiedPathFM uses step distillation to compress the number of sampling steps to one or two, greatly reducing the computational overhead of generating multiple paths.
50%UnverifiedPathFM uses a shared history attention cache to reuse the history encoding across multiple samplings, avoiding repeated computation.
50%UnverifiedThe time for PathFM to generate multiple paths for a single sequence is roughly equivalent to the time of one quantile model query.
50%UnverifiedPathFM's one- to two-step sampler allows the model to optimize directly for calibration quality during training.
50%UnverifiedPathFM achieves competitive marginal accuracy on the existing fev-bench and TIME benchmarks.
50%UnverifiedOn fps-bench, PathFM's overall path score is higher than all evaluated external baselines.
50%UnverifiedIn chaotic system experiments, PathFM's flow model is able to maintain long-term variability, whereas strong baseline models quickly collapse and the trajectories they generate tend to degenerate.
50%