When Agents Imagine The Next Step Before Taking It
Ronni Holmvig Strøm · 2026-06-29
The reasoning models took the spring headlines, and they deserved most of them. Underneath that story, a quieter cluster of June papers converged on a different capability: they give the model an internal account of what its actions will do before it does them. An agent that simulates the next move
The reasoning models took the spring headlines, and they deserved most of them. Underneath that story, a quieter cluster of June papers converged on a different capability: they give the model an internal account of what its actions will do before it does them. An agent that simulates the next move tends to make better ones.
An agent that reasons well in the moment is still, in the technical sense the literature now uses, reactive. It reads the state and picks an action, then finds out what happened by looking at the next state. The work this quarter is about installing the missing organ: a world model the agent can consult privately, asking what happens if I do this before committing to anything.
The Reactive Agent and Its Limits
"Large language model agents have demonstrated strong capability in sequential decision-making, yet they remain fundamentally reactive in long-horizon tasks." That is the opening claim of Internalizing the Future, and it names the problem precisely. A human planning a multi-step task runs cheap simulations in the head. Move the knight here, and the bishop is exposed. Book the connecting flight first, and the visa becomes the bottleneck. The agent, lacking that faculty, commits to step one and discovers the consequence by living through it.
On a three-step task this is tolerable. The recovery is cheap and the error stays small. On a forty-step task it is not, because errors do not wait politely at the step that produced them. They travel. The agent that mispredicts the state at step seven is still planning against that wrong state at step twelve, and by then the plan is fiction built on fiction.
What the agent needs is foresight it can trust, which turns out to be a harder thing to build than foresight it merely performs. The distinction is the whole game this quarter.
Two Families of World Model
Grounded Iterative Language Planning, from Xinyuan Song and Zekun Cai, draws the cleanest map of the territory. World models for language agents, they argue, come in two useful forms, and each fails in a way the other does not.
An agent-based world model is an LLM reasoning in language. You ask it to predict the next state, it calls an API and writes a description, flexibly and with all the linguistic range that makes these systems useful. Its errors arrive as hallucinated state changes. The model confidently reports that a file was written, a task completed, a dependency satisfied, when none of that happened. These errors are nearly impossible to score with ordinary regression losses, because a hallucinated sentence is not numerically far from a true one. It is just false.
A parameterized world model is a trained transition predictor, a dedicated network that maps state and action to next state. Its errors are measurable, and the field has the instruments to measure them: NodeMSE, delta accuracy, validity accuracy. You can watch the number move. The catch is that a parameterized model is usually a weaker planner on its own, without the broad reasoning the language model brings.
So the trade is flexibility you cannot measure against measurability that cannot plan. Stated that way, the resolution is almost obvious, and three of this quarter's papers reach for it: keep both, and make them check each other.