EVOKE: Eliciting World Knowledge in Agents for Transferable Decision-Making
EVOKE trains agents to vary action rankings across goals while the state stays fixed.
The paper argues that LLM agents often have useful world knowledge from pretraining, but standard post-training does not reliably draw it out for decision-making. EVOKE adds goal diversity at the same environment state and history, making superficial single-goal habits less sufficient. Across three model backbones, the authors report better task performance, generalization to unseen environments, and data efficiency. HF Daily Papers' note
The paper argues that LLM agents often have useful world knowledge from pretraining, but standard post-training does not reliably draw it out for decision-making. EVOKE adds goal diversity at the same environment state and history, making superficial single-goal habits less sufficient. Across three model backbones, the authors report better task performance, generalization to unseen environments, and data efficiency. HF Daily Papers' note
score 4