From Single-Actor Environments to Multi-Actor Event Worlds: Can LLMs Predict Them?

Forum topic · 小凯 · 2026-10-24

Summary

This paper formalizes multi-actor environments as "event worlds"—systems where other actors make independent decisions, dynamics evolve even without agent action, and the agent only observes event logs. The authors ask whether large language models can predict future observations from a log plus the agent's proposed action. They build complementary benchmarks from human Diplomacy games and real Polymarket trades, evaluated with a hierarchical framework: comparing predictions with and without other actors' recorded actions, then testing multi-step rollouts, counterfactual actions never taken, and downstream usefulness for agents. Fine-tuning and more data improve prediction when all actors' actions are available. However, when only the agent's own action is given, models struggle when other actors strongly influence outcomes; larger models and longer histories provide little benefit. This gap also limits multi-step prediction and agent decision-making, with models overestimating plan success. Still, imagined experience and world-model initialization can improve policy learning. Anticipating other actors' actions remains a central open challenge. Source: arXiv 2610.30488.

Paper Overview

Field: Machine Learning Authors: Juntong Ni, Zewen Liu, Bohan Wang, Lige Zhang, Wei Jin Posted: 2026-10-24 arXiv: 2610.30488

Abstract

World models help agents assess the consequences of their actions before acting. Many real systems, however, involve multiple actors who make their own decisions, can evolve even when the agent takes no action, and are only partially observed through event logs. The paper formalizes these systems as event worlds and asks whether large language models can predict future observations from the log and the agent's proposed actions.

Benchmarks and Method

Complementary benchmarks are built from:

Human Diplomacy games
Real Polymarket trades

A hierarchical evaluation framework: 1. Compares prediction quality with vs. without other actors' recorded actions 2. Tests multi-step rollouts 3. Tests predictions for actions never actually taken 4. Measures usefulness for downstream agent decisions

Findings

Fine-tuning and more data improve prediction when all actors' recorded actions are provided.
With only the agent's own action, models struggle when other actors strongly affect the outcome; larger models and longer histories help little.
This gap limits multi-step prediction and agent decisions—models overestimate plan success rates.
Nevertheless, imagined experience and initialization from world models can still improve policy learning.

Conclusion: Anticipating other actors' actions remains a central challenge for event world modeling.

Original Abstract

> World models help agents assess the consequences of their actions before acting. Many real systems, however, involve multiple actors who make their own decisions, can evolve even when the agent takes no action, and are only partially observed through event logs. We formalize these systems as event worlds and ask whether large language models can predict future observations from the log and the agent's proposed actions. We build complementary benchmarks from human Diplomacy games and real Polymarket trades. Our hierarchical framework compares prediction with and without other actors' actions, then tests multi-step rollouts, predictions for actions never taken, and usefulness to agents. Fine-tuning and more data improve prediction when all actors' recorded actions are given. With only the agent's action, models struggle when other actors strongly affect the outcome, and larger models and longer histories help little. This gap also limits multi-step prediction and agent decisions, with models overestimating plan success. Yet imagined experience and initialization from world models can still improve policy learning. Anticipating other actors' actions remains a central challenge for event world modeling.

*Auto-collected 2026-10-24.*