← Research overview

Long-Horizon Consistency for Causal Video & Interactive World Models

Research project

Study how representation drift compounds through causal rollout, and how models preserve motion, state, and control.

Research Intern, Samsung AI Center–Cambridge · Mar–Sep 2026
CONTROLLED PATHACCUMULATING DRIFTSYNTHETIC ROLLOUT
Schematic illustration · Not generated video or measured research results.

When generation becomes a rollout

A convincing short clip does not guarantee a coherent long sequence. Reusing generated history can accumulate errors in motion, identity, and response to controls.

From visible errors to internal state

This research studies causal text-to-video and action-conditioned generation across increasing rollout lengths. Probes examine VAE latents, transformer hidden states, and key-value histories. Comparisons between online, refreshed, and recomputed histories help separate visible drift from memory effects.

Partition-consistency post-training explores how to reduce errors across rollout boundaries while preserving short-horizon quality.

Reading the illustration

The upper row shows an idealized controlled trajectory; the lower row shows a deliberately constructed accumulating offset. Both are synthetic drawings. They illustrate the evaluation question, not outputs or measured improvements from a video model.