LAGO framework decodes language into latent subgoals for long-horizon planning

Planning with world models has long been plagued by compounding prediction errors and the challenge of defining optimizable goals. Visual targets offer precise local gradients but fail at distant guidance, while language is flexible yet noisy due to cross-modal misalignment or reliance on large generative models. The new LAGO (Latent Goal Prediction from Language) framework tackles this by predicting both sequences of intermediate goal states from language instructions and action-conditioned rollouts—all within a single latent space. Instead of optimizing toward a single global objective, LAGO dynamically decomposes instructions into explicitly predicted, locally tractable latent subgoals. By updating these subgoals online and using a soft minimum trajectory cost during planning, the agent can follow coherent latent trajectories over long horizons.

In evaluations across multiple environments with varying planning horizons, LAGO avoids the sharp degradation seen in prior methods, achieving robust and precise long-horizon planning purely from language. This approach bridges the precision of visual goals with the flexibility of text-guided control, enabling agents to execute complex instructions without requiring perfect cross-modal alignment or expensive generative models at inference time. The framework is detailed in a 9-page preprint under review, with code expected to follow.