Social world models predict multiple futures to improve outcomes
Anticipating Social Consequences with Distributional World Models
Artificial Intelligence
Summary
Some situations involve many possible results from one choice, and just guessing one future can miss important outcomes. The authors made a new model called DSWM that predicts various possible reactions and their chances all at once instead of one at a time. This helps an agent pick better actions by knowing many likely consequences. They trained a large model on many simulated examples and tested it in several different situations, showing it does better than other methods.
What this means in practice
- •For game developers: Generate multiple possible non-player character responses at once to create more dynamic social interactions in games.
- •For robotic system designers: Enable robots to anticipate multiple consequences of their actions to improve cooperation and negotiation with humans.
Authors
Viet-Thanh Pham, Thuy-Trang Vu, Zhuang Li, Lizhen Qu, Gholamreza Haffari
Abstract
Social world models predict how interactions develop after an action, helping agents negotiate, cooperate and mediate. In these situations, one action can lead to several plausible futures, hence predicting only one future can overlook consequential alternatives, while exploring them through repeated simulations increases inference cost. We propose DSWM, a Distributional Social World Model that predicts alternative responses, their probabilities and immediate and final outcome values in one output per candidate action. We train a 4B model on repeated simulations using supervised fine-tuning and GRPO. At inference time, the actor uses these possible consequences to choose an action. Across three seen and two unseen environments, with interaction backbones absent from training collection, DSWM achieves the best average performance on 13 of 15 metrics. Ablations show that retaining multiple futures and their predicted probabilities improves outcomes in all five environments.