World modeling enables test-time scaling without environment replicas

The key insight is that this approach avoids the fork-and-simulate bottleneck entirely by treating inference as iterative refinement within a single execution trace, rather than requiring parallel rollouts. Most test-time scaling work assumes you can spin up cheap copies of the environment; this paper assumes you cannot.

This connects directly to the World Reward Models paper from late October, which also tackled the computational waste in agent search by avoiding full trajectory simulation. Where WRMs learned to predict comparative outcomes, Agentic World Modeling sidesteps simulation altogether by reframing the problem as evidence gathering within one path. Both papers are responses to the same constraint: real production systems (like Arca's multi-day financial workflows) cannot afford to fork execution or spin up test environments. The latent world models critique from the same period is also relevant here: it showed that learned world models often don't actually model environments faithfully. This paper avoids that trap by not building a world model at all, only reasoning over the trajectory it's actually executing.

If this technique shows comparable scaling benefits to parallel test-time search on standard benchmarks (GPQA, ARC-Challenge) without requiring environment access, that validates the core claim. More importantly, watch whether teams building production agents (like Arca or similar financial/enterprise systems) adopt this in their long-horizon workflows within the next six months. If adoption happens before a major benchmark win, that signals the operational constraint was real and this solves it.