Latent Dynamics · 2026-10-23
Anticipating Social Consequences with Distributional World Models
Viet-Thanh Pham, Thuy-Trang Vu, Zhuang Li, Lizhen Qu, Gholamreza Haffari
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
DSWM (Distributional Social World Model) predicts, for each candidate action, multiple plausible social response clusters with probabilities plus immediate and final outcome values. The model is trained as a 4B world model from repeated simulations using supervised fine-tuning and GRPO. At inference, the actor uses these predicted alternatives to choose actions. The authors report best average performance on 13 of 15 metrics across three seen and two unseen environments, with interaction backbones absent from training.
Why it matters
The paper addresses a core decision issue in social settings: one action can lead to several consequential futures. DSWM makes alternative responses explicit in a single output per candidate action, including their likelihoods and outcome values, aiming to reduce the inference cost of exploring futures while improving action selection and transfer to unseen environments.
Method
- Action-conditioned distributional WM: for one candidate action, outputs up to K response clusters with probability p_k, representative reply text r_k, and conditional mean immediate/final outcome values per requested measure/horizon.
- Training data from repeated simulations: save full state at a decision point, execute the same candidate action N=50 times to sample other parties’ responses; construct distributional targets by partitioning responses, clustering semantically, and using empirical frequencies and cluster outcome means.
- Train with SFT then GRPO: SFT generates JSON targets from visible context + candidate action + query; GRPO rewards matching generated distributions to reference continuations with constraints on parse, measures, and cluster count.
Limitation
The authors note that predicting and using cluster-mean values can approximate expected utility for nonlinear objectives, because the representation omits within-cluster variation and covariance.
Abstract (from arXiv)
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.
Related papers
- From Single-Actor Environments to Multi-Actor Event Worlds: Can LLMs Be Event World Models?
- AnyGoal: On World Model Generalization Beyond Expert Rollouts
- AffectWorld: Benchmarking Emotion-Driven Future World Prediction
- Beyond Prediction: Steering VLM Agents with Retrospective World Modeling
- An Introduction to World Models for Generalist Agents
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-22 · Indexed by Pub-AI: 2026-10-23 UTC
Anticipating Social Consequences with Distributional World Models
Preprint posted: 2026-10-22 · Original record ↗
- arxiv
- 2610.30657
Explicit source versions (1)
- v1 · 2026-10-22
- Preprint available
Source date: 2026-10-22 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
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.