PriorZero: Bridging Language Priors and World Models for Decision Making
TLDR
PriorZero integrates LLM priors into world-model-based planning via root-prior injection in MCTS and decoupled training, tested on Jericho and BabyAI.
Reasoning
Strengths include addressing prior-dynamics mismatch with a novel decoupled design and empirical validation on two benchmarks. Weaknesses are limited task domains (text and gridworld) and lack of real-world physical experiments.
Read-first score
Read-first score 61.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 41.
Field roles
Rank sensitivity
Stability: volatile; rank range: 465.
Keyword Scores
Deep Analysis
Innovations
- Root-prior injection mechanism that incorporates LLM priors exclusively at the root node of Monte Carlo Tree Search (MCTS), focusing search on semantically promising actions while preserving the world model's deep lookahead capability.
- Decoupled rollout-training design that separates world-model learning from LLM adaptation, enabling continuous refinement of the world model on interaction data.
- Alternating optimization that uses the world model's value estimates to provide fine-grained credit assignment signals for stable LLM fine-tuning.
Methodology
PriorZero integrates LLM-derived conceptual priors into world-model-based planning via a decoupled rollout-training design. During rollout, a root-prior injection mechanism incorporates LLM priors exclusively at the root node of MCTS. During training, the world model is continuously refined on interaction data, and its value estimates are leveraged to provide credit assignment signals for stable LLM fine-tuning through alternating optimization.
Key Results
Experiments on text-based adventure games in Jericho and instruction-following gridworld tasks in BabyAI demonstrate that PriorZero consistently improves both exploration efficiency and asymptotic performance.