Latent Dynamics · 2026-11-02
SPT-Flow: One Integration, Many Futures in Continuous-Time World Models
Mingyang Lyu, Yinqian Sun, Gongce Cao, Yuhao Zhang, Lu Jia, Qian Zhang, Bo Wang, Pin Lyu
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
SPT-Flow is a continuous-time world model that samples one trajectory-level latent per rollout and trains via rollout matching to demonstration readout states (without intermediate ground-truth inputs). The authors argue this prevents “decorative latent conditioning” from teacher forcing. On LIBERO (30k steps), they report higher best-of-10 PSNR and lower mean LPIPS with comparable per-sample inference latency.
Method
- Uses an ODE-based physical-time flow (PT-Flow) with a trajectory-level latent ε sampled once per rollout; inference integrates once and decodes at readout times.
- Trains with rollout matching: sample ε, integrate from z0 without intermediate ground-truth states, and match decoded predictions to demonstration states at multiple readout times (G=4).
- Evaluates both accuracy and latent utilization using posterior specificity and a latent swap diagnostic comparing rollouts conditioned on matched vs swapped inferred latents.
Limitation
The semantic mode coverage under matched conditions in the robotic setting remains unverified.
Abstract (from arXiv)
Continuous-time world models generate a latent trajectory through a single ODE integration, but adding stochastic conditioning does not ensure that the model uses it. In our teacher-forced baseline, posterior means vary by only about 1% of the prior scale, and matched and swapped latents yield indistinguishable reconstruction PSNR at the reported precision. We study a supervision-shortcut hypothesis: ground-truth intermediate states can reduce the need for trajectory-associated latent information. We introduce SPT-Flow, which samples one latent per trajectory and matches a free-running ODE rollout to demonstration states at multiple readout times. Re-anchoring experiments support the proposed mechanism. On LIBERO, with both models trained for 30k optimizer steps, SPT-Flow improves best-of-10 PSNR over our retrained ODEWorld baseline by 2.0 and 2.4 dB at the mid-trajectory and final frames and reduces mean LPIPS by 23%, with comparable per-sample inference cost. The rollout control recovers most of the accuracy gain; learned latent conditioning adds diversity and best-of-K gains while reducing mean sample accuracy relative to this control. The diagnostics support trajectory-associated latent utilization; semantic mode coverage under matched conditions in the robotic setting remains unverified.
Related papers
- Roz: A Taste Test for World Model Latents
- Structured Latent Dynamics for Long-Horizon Embodied World Models
- T2-z: Test-Time Optimization of Environment Latents for Robotic World Models
- Identifiable World Models from Pretrained Diffusion Representations
- How Many Candidates Does a One-Step World Model Need? A Random-Codebook Account of Nearest-of-m Training on Frozen DINO Tokens
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First announced by arXiv: 2026-11-02 · Indexed by Pub-AI: 2026-11-02 UTC
SPT-Flow: One Integration, Many Futures in Continuous-Time World Models
First announced by arXiv: 2026-11-02 · Original record ↗
- arxiv
- 2610.41273
Explicit source versions (1)
- v1 · Date not supplied
- Preprint available
Source date: 2026-11-02 · First detected: 2026-11-02 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · RSS guid=oai:arXiv.org:2610.41273v1; arxiv:announce_type=new; item pubDate (official metadata)
Abstract · arxiv
Continuous-time world models generate a latent trajectory through a single ODE integration, but adding stochastic conditioning does not ensure that the model uses it. In our teacher-forced baseline, posterior means vary by only about 1% of the prior scale, and matched and swapped latents yield indistinguishable reconstruction PSNR at the reported precision. We study a supervision-shortcut hypothesis: ground-truth intermediate states can reduce the need for trajectory-associated latent information. We introduce SPT-Flow, which samples one latent per trajectory and matches a free-running ODE rollout to demonstration states at multiple readout times. Re-anchoring experiments support the proposed mechanism. On LIBERO, with both models trained for 30k optimizer steps, SPT-Flow improves best-of-10 PSNR over our retrained ODEWorld baseline by 2.0 and 2.4 dB at the mid-trajectory and final frames and reduces mean LPIPS by 23%, with comparable per-sample inference cost. The rollout control recovers most of the accuracy gain; learned latent conditioning adds diversity and best-of-K gains while reducing mean sample accuracy relative to this control. The diagnostics support trajectory-associated latent utilization; semantic mode coverage under matched conditions in the robotic setting remains unverified.
