Large language models are increasingly used as agents for multi-step decision-making, but they transfer poorly to unfamiliar environments. One common fix is to train agents to predict future observations — a world-model approach that adds training cost and compounds errors when predictions feed into planning.
Researchers behind a new paper, EVOKE, argue that for LLM agents in digital environments much of this world knowledge is already internalized during pretraining. The problem, they say, is not acquiring it but eliciting it. Typical post-training offers little pressure to do so: supervising a single goal at each visited state can push policies toward superficial contextual habits that fail to generalize.
EVOKE supplies that pressure through goal diversity at fixed states. At collected states, the environment state, interaction history, and available actions are held constant while alternative goals are swapped in. The policy must rank the same candidate actions under each goal, so action preferences have to change — a policy relying on habits or single-goal correlations cannot order them correctly. This implicitly draws on the policy's pretrained world knowledge.
The method requires no world-model module and no inference-time planning. According to the paper, EVOKE achieves the best average results on ALFWorld, WebShop, and search-based QA across three backbones, and reaches 91.1% / 93.1% / 84.6% on unseen ALFWorld games with Qwen2.5-3B, Qwen2.5-7B, and Qwen3-1.7B respectively.
Controlled analyses suggest the backbone already encodes action consequences before any ALFWorld training, and that EVOKE reduces habitual errors — choosing the action that is correct for a different goal in the same context. The authors frame the work as a new perspective on eliciting internalized world knowledge for transferable action through direct decision supervision.