Keplen — Signal-Latent World Model
Studying how a shared world model can combine different inputs, make predictions and choose not to act when appropriate.
World-model researchJan 2026 — Present
Overview
Research into shared representations of text, audio and video, and how models can express uncertainty before acting.
Technical work
- Leading Signal-Latent World Model (SLWM) research. Separate text/code, audio and visual encoders feed a shared latent representation, followed by prediction and uncertainty heads.
- The design includes a policy gate that can decline to act. Experiments compare models under matched data, parameter and compute budgets, with ablations and calibration measurements.
- This is ongoing research. Working pipelines and smoke tests establish implementation behaviour, not evidence of better model performance.