Learning to Model the World with Language
TLDR
Dynalang learns a multimodal world model from diverse language to predict future observations and rewards, enabling game-playing and navigation in photorealistic homes.
Reasoning
The paper presents a novel perspective unifying language understanding with future prediction via a self-supervised world model. Strengths include demonstrated success on multiple tasks and pretraining on text-only data; weaknesses are limited comparison to baselines and lack of explicit real-world deployment details.
Read-first score
Read-first score 72.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 58.
Field roles
Rank sensitivity
Stability: volatile; rank range: 157.
Keyword Scores
Deep Analysis
Innovations
- Interpreting diverse language as a signal for future prediction, unifying language understanding with future prediction as a self-supervised learning objective.
- Dynalang agent that learns a multimodal world model to predict future text and image representations and acts from imagined model rollouts.
- Ability to pretrain on text-only data and generate language grounded in an environment.
Methodology
Dynalang learns a multimodal world model that predicts future text and image representations. The agent learns to act by performing imagined rollouts from the world model, using a self-supervised objective that unifies language understanding with future prediction.
Key Results
Dynalang outperforms language-conditioned policies on tasks ranging from game-playing to navigating photorealistic home scans when using diverse types of language (environment descriptions, game rules, instructions). It also enables pretraining on text-only data and generation of grounded language.