Multilingual Latent Alignment for Strategic World Models

[Arsh123344423]

Motivation
Strategic World Models currently assume a single observation language, which limits their applicability in real-world settings where agents must reason and plan across multiple languages and cultural contexts. Enabling multilingual reasoning while preserving strategic consistency would significantly improve generalization and usability.

This feature proposes extending StrategicWM with language-agnostic latent state alignment, allowing identical strategic situations described in different languages to map to the same abstract world state.

Proposed Feature
Introduce Multilingual & Cross-Lingual World Model Alignment by:

1. Multilingual Observation Encoding

   1. Integrate a multilingual text encoder (e.g., XLM-R, mT5) to process language-based observations.
   2. Map observations from different languages into a shared latent space.

2. Strategic Latent Alignment Loss

   1. Enforce similarity between latent states representing the same world configuration across languages.
   2. Possible approaches:
      1. Contrastive loss
      2. KL-divergence
      3. InfoNCE-style objective

3. Language-Conditioned Decoding

   1. Decode plans, actions, or explanations in the input language without altering the core planner.
   2. Supports explainable and human-aligned planning.