EVOKE: Eliciting World Knowledge in Agents for Transferable Decision-Making

EVOKE. Eliciting World Knowledge in Agents for Transferable Decision-Making. By Yuhan Guo and ten others, from Shanghai Jiao Tong University, the Hong Kong Polytechnic University, and one other institution. Published on the twenty-ninth of September. Its code repository is public but slight: nine source files.

Large language models are increasingly put to work as agents, making decisions over many steps. They transfer poorly, though: an agent that does well in one environment tends to falter in one it has not seen. A common remedy is the world model. The agent is trained to predict future observations, and that prediction helps it plan. The authors note two costs. It takes additional training, and when those predictions are used for planning, their errors compound.

Then comes the reframing. For agents working in digital environments, much of that world knowledge is already inside the model, internalized during pretraining. The problem is not acquiring it. It is getting it out. And typical post-training, the authors argue, puts little pressure on the model to do that. When supervision comes under a single goal at each visited state, the policy is inadvertently pushed toward superficial contextual habits instead of the knowledge it already has.

EVOKE is a post-training method designed to supply the missing pressure, and it does so through goal diversity at fixed states. The motivation is a theoretical one. An agent that is competent across diverse goals must encode a world model, one that can be recovered from its action preferences. If that is what competence across goals requires, then training for competence across goals should draw the world model out.

In practice, EVOKE holds the environment state and the interaction history fixed. It then takes the same set of candidate actions and ranks them under alternative goals. Because the goal changes while everything else stays put, the right ordering of the actions has to change too. A policy leaning on contextual habits, or on correlations that held under a single goal, cannot get those orderings right. That is how the method implicitly calls on the policy's pretrained world knowledge to inform its decisions, without ever training it to predict observations.

The evaluation spans diverse tasks and three backbones. EVOKE improves task performance, generalizes better to unseen environments, and is more data efficient. The authors also run controlled analyses aimed at understanding what drives those gains. The abstract does not give the figures.

Where a world-model method adds a prediction task, EVOKE changes the question asked at each state. Same state, same history, same actions, different goal: habits learned under one goal no longer give the right answer, and the knowledge from pretraining is what the policy is pushed to draw on instead. The authors present this as a new perspective, eliciting knowledge a model already has through direct supervision of its decisions, and as a route to action that transfers.