One Lens, Many Worlds : A Capability-Typed Interface for World-Model Interpretability
TLDR
Proposes WorldModelLens, a capability-typed interface for interpretability of diverse world models, unifying methods across architectures.
Reasoning
The paper identifies a genuine problem of fragmented interpretability tooling for world models and offers a principled interface solution. However, the abstract lacks empirical validation or real-world experiments, making it unclear how effective the proposed substrate is in practice.
Read-first score
Read-first score 60.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 44.
Field roles
Rank sensitivity
Stability: volatile; rank range: 453.
Keyword Scores
Deep Analysis
Innovations
- Capability-typed interface that unifies interpretability across diverse world model architectures (latent recurrent, token-based, joint-embedding).
- WorldModelLens substrate with four required methods and optional heads, allowing RL and self-supervised models to be first-class without imitating each other.
- Single hook and cache layer exposing time-indexed activations, imagination rollouts, and intervention replay over the interface.
Methodology
WorldModelLens defines a typed interface with four required methods (encode, transition, initial state, sample) and optional heads (decode, reward, continue, actor, critic) declared via a capability descriptor. A hook and cache layer is built over this interface to expose time-indexed activations, imagination rollouts, and intervention replay. This design allows interpretability methods to be implemented once and applied across different world model architectures.
Key Results
No experimental results are mentioned in the abstract; the paper presents the interface design and its intended benefits.