How Do You Know It Followed the Instruction?

SIMA’s action loop can produce a plausible movement even when the agent has not used the language it was given. This chapter examines how the project tries to separate instruction following from scene-driven behavior, especially when commercial games lack a common goal state or privileged evaluation API. The practical value is a reading rule: every success signal has a scope, a cost, and a failure mode, so no single score can stand for general embodied competence.

Why a plausible action is not enough

A controller can look intelligent for the wrong reason. If a scene contains one obvious object and one obvious interaction, the agent may perform the expected action because the visual arrangement strongly suggests it. The words may have contributed little or nothing. In SIMA, that distinction matters because the project is not only asking whether an agent can move through a 3D world. It is asking whether a user’s instruction changes what the agent selects and does.

The problem becomes sharper as instructions become open-ended. “Cut the carrot on the cutting board” may correspond to a recognizable goal state in a controlled environment. “Make a pile of rocks to mark this spot” or “see if you can jump over this chasm” may be meaningful to a human without mapping to a state variable that the environment records. Commercial games were built for play, not for checking every possible natural-language intention. They may not expose internal state, and even when a save can be loaded, restoring exactly the same situation can be slow or expensive.

This makes offline action prediction an incomplete test. SIMA can be shown a static image and asked to predict a valid response such as “jump,” or evaluated by the likelihood it assigns to a held-out human action. Those signals are useful for simple correspondences and early progress. They do not reliably predict success on a prolonged task, where the policy must keep a target in view, respond to the environment’s changes, and avoid irrelevant actions. A high likelihood for one next input is not the same as completing a command.

The distinction is between a local correspondence and an online episode. A local test asks, “Given this frame and this phrase, is this action plausible?” An online evaluation asks, “After the agent acts, does the world change in the intended way, and can the agent continue until the task is complete?” The second question includes the errors that accumulate between observations, actions, and consequences. It is more expensive, but it tests the capability SIMA actually claims to study.

A fair task must therefore make language matter. One useful strategy is to issue multiple instructions from the same initial state while allowing several possible actions. If the agent responds differently when the command changes, that is stronger evidence of language conditionality than success in a scene where only one behavior makes sense. This design does not prove perfect comprehension: an agent can still misinterpret the wording or use a superficial cue. It reduces the easier explanation that the environment itself dictated the answer.

There is a second complication. A task may have an objective outcome without having a single canonical path.