Hand the wheel to a brilliant LLM. An LLM — a large language model that takes questions in plain language and answers in kind — can be wrapped in software that lets it decide for itself: click around the web, run searches, and get work done. We call that software an agent. On the course it practiced on, the agent drives as smoothly as an F1 driver. But drop it into an alley it has never seen and it hits a wall almost immediately. This tendency for agents to lose their footing in new environments — is it really because they "don't know enough"?
EVOKE flips the question. The problem isn't that the agent doesn't know the world; it's that it can't reach for what it already knows at the very moment of deciding. So the remedy is different too. Instead of cramming more new knowledge into its head, EVOKE shakes the dormant knowledge awake.
When what an agent has learned is bound to one specific task and environment, it can't carry over to a new situation. One line of work fixes this by building a "world model." The idea is to give the agent an internal predictive model — a way to picture the consequences of its actions before it takes them. Before choosing an action, it's trained to predict "if I do this, how will the screen change?" Those consequences follow rules that hold across environments ("press Place Order and the purchase goes through on any shopping site"), so a decision reasoned out from consequences has a good chance of holding up in a new environment too — that's the logic.
But learning consequences through prediction comes at two costs. First, prediction is a separate objective that makes training more expensive, and much of what gets predicted has nothing to do with the decision actually at hand. Second, the moment you use a predicted future for planning, its error snowballs across many steps.
Here the question sharpens. Does an LLM agent really have to learn those consequences "anew" through prediction? In a physical environment, dynamics like contact and motion are hard to grasp without living through them with a body. But an LLM agent mostly operates in digital environments — websites, search engines, and the like. In places like that, much of what an action leads to is already contained in the knowledge the model absorbed during pretraining — the first stage of learning, where the model reads the world's vast written record to build up basic common sense. Prior work shows that pretrained models, after seeing just a few examples, predict fairly well how a text environment will respond to an action.
If so, the nature of the problem changes. It's no longer about "acquiring" new knowledge, but about getting the model to "put to use" knowledge it already has. The trouble is that ordinary post-training — the stage where a finished model is taught further to fit a specific job — puts no pressure on exactly this kind of retrieval. Under conventional methods, an agent can get the answer right by simple rote memorization: "on this screen, always this button." It's like a student who's had only cram-session tutoring — they ace everything within the exam's scope but fall apart on an application problem they've never seen. Memorized habits work only in the environments seen during training, and they break fast in a new one.
Back to driving lessons. If you always teach on the same road, drilling "turn right at this intersection" over and over, the student just memorizes the habits of that road. But if the instructor parks at that same intersection and keeps changing the destination — "this time, to the hospital," "this time, to the gas station" — the student can no longer answer from memorized habit. They have to actually know where each road leads in order to choose differently for each destination.
That is exactly what EVOKE does. It leaves the environment state, the history so far, and the list of currently available actions untouched, and swaps out only the goal. The paper's example is vivid. The agent is standing on a checkout page. If the goal is "buy now," it should press Place Order; if the goal is "revise the selection," it should go back to the cart. The situation is one and the same, yet because the goal changes, the correct action flips.
Put this flip in technical terms. Whether it's a right turn or the direction of the hospital, every road shares the same knowledge of consequence: "follow this road and you arrive there." This consequence that an action brings about is called a transition. What changes is only how useful that consequence is to the current goal — its value, which we'll call a score. When the instructor keeps changing the destination, it forces the student to think separately about "where do I arrive (transition)" and "is that where I need to go right now (score)." Judging from the situation alone, with memorized habit, you can't correctly rank both sides at once when the goal has been flipped. In this way, the knowledge of consequence lying dormant inside the policy — the rule that decides which action to take in which situation — is dragged out of "merely holding it" into "using it to decide."
This idea has theoretical backing. Richens et al. (2025) showed that any agent that reliably handles a sufficiently diverse set of multi-step goals must have a recoverable world model embedded in its goal-conditioned behavior. EVOKE runs this proposition in reverse: if you demand correct behavior across diverse goals, the policy has no choice but to mobilize the world knowledge sleeping inside it.
In practice the procedure is a loop of four interlocking stages.
Hidden here is the most decisive device of all: taking the "attractive wrong answer" that the agent is most tempted to grab first — "this looks like the answer, right?" — and deliberately turning it into a training problem (policy-favored hard negatives — the trap wrong answers the policy itself mistakenly thinks are good). Whatever the policy reflexively reaches for, if it isn't the right answer, you put it straight into the competitor slot. For example, if the goal is to set a clean cup on the table while holding a dirty one, the right answer is to head to the sink — and the "just set the dirty cup on the table" that the policy eagerly proposes is precisely that trap wrong answer. By placing the wrong answer the policy itself favors directly into the competitor slot, EVOKE corrects exactly the trap it is most likely to fall into — because it explicitly suppresses the tempting wrong answer. Training that only lifts positives, by contrast, walks right past this trap.
This whole process runs on a low-cost method (LoRA) that trains only a small add-on rather than the entire model. And by repeatedly correcting with fresh data gathered from redeploying the trained model into the field (DAgger), it gets smarter round by round. At deployment time, though, it runs as a single ordinary policy — no prediction model, no extra add-on.
Evaluation was carried out across three domains — a text-based household task, web navigation, and search-based QA — using three backbone (the underlying base LLM) models. The results interlock along three strands. With only a small amount of data, EVOKE improved the ability to apply what was learned to unseen environments (generalization), and ultimately raised overall problem-solving ability. The central claim of this work is that it drew out transferable behavior with neither a prediction objective nor a separate world model attached — using only direct supervision over decisions (a training signal that tells it the right answer). Per-domain and per-backbone details can be found in the paper's tables.
The most notable part is the controlled analysis that isolates "where this effect actually comes from." The key question is this: is the source of the gain the intervention of swapping in different goals itself, or is it simply a matter of feeding in more goal data? The reason the two must be told apart is clear: if it's the latter, anyone can catch up by pouring in more data, but if it's the former, the intervention itself is what earns its keep. After comparing against a control matched for the same amount of data, the paper's conclusion is unambiguous — the gain comes not from pouring in more data, but from the "re-asking with a changed goal" method itself. The analysis also separately teases apart whether the iteration really keeps improving the policy, whether that knowledge was inside the model from the start, and whether the gain comes from better decisions or merely from an increased number of attempts.
Data efficiency is especially valuable in the field. If you have to stack up thousands of fresh labels for every environment, a deployed AI product's cost balloons every time the environment changes. If you can get transferable behavior from little data, the threshold for moving to a new customer or a new environment drops accordingly.
The limits are just as clear, and worth naming. First, the whole premise of this approach is that "knowledge of consequence is already in pretraining." The paper draws the line too: this holds for digital environments, while physical dynamics like contact and motion are still hard to learn without direct experience. In an unfamiliar domain that pretraining never captured, the "knowledge to pull out" may itself be thin.
Second, label quality leans heavily on the helper model (the annotator). Proposing alternative goals and judging whether they're compatible with the scene both fall to the annotator. Settling positives versus competitors by actual execution results is a safeguard, but the bias in the earlier step of proposing and filtering goals remains an open question.
Suppose ONTLE's AI PM is looking at an urgent ticket from Customer A. One request is stuck as "waiting on a decision (on_hold)," and several buttons sit on the screen — press for a reply, offer an alternative, escalate to a supervisor, keep it on hold. The one formula it learned at Customer B is "when stuck, nudge." But right now Customer A is at a stage where trust matters more than speed. A PM drilled on a single goal easily hardens into the surface habit of "if on_hold, always nudge," and that very habit misfires the moment the customer changes.
EVOKE offers two hints here. One is to swap only the goal into that same unresolved state — "hit the deadline," "preserve trust," "minimize cost" — and re-rank the action candidates. Then the PM can only choose correctly in all three cases by connecting what each button "actually causes" to the goal. The other is to suppress the trap wrong answer. You feed in the reflexive choice the PM is most prone to make — say, "when in doubt, just escalate" — as a trap wrong answer and correct for it.
One thing must be made clear, though. The paper's "grounded action evaluation" judged by actually executing actions and seeing the results, but ONTLE's PM can't go executing things carelessly on a customer ticket. So it would use the progress records of past similar projects in place of an answer key — and here, you first have to check whether those records came from "a state like the one now." Copying answers from records of a different situation only manufactures one more surface habit. This paragraph is an application concept we're sketching, not a verified result, and how to precisely match for differences in situation remains homework left undone.
No. Swapping the goal and re-asking is a training technique used only "while teaching." Once that process has conditioned the policy to use knowledge of consequence in its decisions, at deployment it simply runs as a single ordinary policy — no prediction model, no extra add-on. By analogy, a driver who's finished lessons chooses the route on their own, with no more pop quizzes from the instructor.
It isn't pure self-grading, but the helper model's judgment doesn't drop out entirely either. Labels are grounded in the "observed outcome" obtained by actually executing a candidate action — but classifying that outcome as "advances the goal / falls short / uncertain" still falls to the annotator (the helper model). Actions classified as uncertain are excluded from training altogether. In other words, it differs from blind self-grading in that the grading criterion is tied to reality outside the model (actual execution results), yet the model's judgment in interpreting those results is still involved. That's why the paper leaves open not only the bias in the earlier step of proposing and filtering alternative goals but also the bias in this labeling step.