Introducing Interaction World Models
In the same way world models can be used to simulate possible futures of the real world, they can provide agents with a method to support counterfactual reasoning and long-horizon planning specifically in social interaction—relying on a world model that captures the latent dynamics of interaction.
The Need for Latent Dynamics: A Crisis Support Scenario
Consider the scenario where a crisis support agent is interacting with an incoming caller. The agent is tasked with the goal to reduce anxiety, lower arousal, decrease resistance, and increase trust. These internal human states do not evolve smoothly or monotonically. An intervention (like logical reframing) that successfully reduces confusion when a caller's resistance is low might trigger a severe, defensive "backfire effect" when their resistance and arousal are high. Planning under such complex, state-dependent dynamics requires a model that can look ahead. An Interaction World Model allows the agent to implicitly simulate, "If I use a validation strategy now, how will their trust and cognitive load shift over the next three turns?"
Social-JEPA: The First Interaction World Model
Our work explores how to best represent and forecast these social futures. Traditionally, generative world models approach this by narrating possible futures—generating the interlocutor's next response in text, then analyzing it. However, autoregressive text generation is both fundamentally lossy and far too slow for real-time planning within a conversational turn.
We introduce Social-JEPA, the first Interaction World Model for latent interaction dynamics. Social-JEPA operates on dyadic textual interaction histories and executes multi-step rollouts without ever decoding text:
Bypasses Text Rendering: It forecasts future interaction states directly as compact embeddings, avoiding the surface-level confounds and signal loss inherent to natural language generation.●
Self-Supervised Latent Learning: It learns interaction structure through Exponential Moving Average (EMA) targets and multi-step rollout training, functioning independently of oracle state labels.
Enables Real-Time MPC: It allows the agent to evaluate the downstream effects of multiple abstract strategies (e.g., Validate, Reframe) simultaneously, supporting continuous Model Predictive Control.
Shifting the Paradigm for Social Planning
Social-JEPA is the first latent-space world model for social interaction. To rigorously test it, we built LID-Bench, the first controlled testbed with oracle latent states for interaction dynamics.
Compared to existing generative and reconstructive approaches on our new benchmark, we observe Social-JEPA achieves massive leaps in both fidelity and efficiency: 2.8x to 3.9x higher accuracy in forecasting latent dynamics across multi-step horizons. 1,059x to 2,314x speed improvements during rollouts compared to evenly matched generative models spanning 117M to 350M parameters.
A structural bottleneck exists in observation-space models: rendering predictions into text destroys up to 100% of the predictive signal by step 5. Scaling up parameter counts by 3x does not close this gap.
This work significantly moves the needle toward real-time social interaction planning, proving that the immense efficiency of latent-space prediction translates successfully from continuous physical control to discrete, language-based social domains.