Beyond Myopic World Models: Long-Horizon End-to-End Training for Direct Future Prediction

TL;DR - An arXiv preprint arguing that world models fail at long-horizon imagination because they're trained on few-step losses and then recursively rolled out; it proposes training directly on an end-to-end endpoint-prediction objective. This matters for model-based RL and planning, where compounding rollout error is the main bottleneck.

• Identifies an objective/deployment mismatch: few-step losses optimize local transition fidelity, treating transitions uniformly regardless of downstream influence, while recursive inference amplifies small local errors.
• Introduces DPWM (Direct Prediction World Model), a non-recursive architecture that compresses an arbitrary-length action sequence into a single embedding and predicts the endpoint observation in one forward pass — avoiding recurrent rollout in both inference and gradient propagation.
• Reports substantial gains over recursive baselines on continuous-control and pixel-based benchmarks, with the margin growing as horizon increases, at horizons where unrolled autoregressive training becomes unstable.
• Ablation supports the central claim: recurrent backbones improve similarly when retrained with the same long-horizon endpoint objective, so the training objective — not the architecture — is the primary driver.