EVOKE: Eliciting World Knowledge in Agents for Transferable Decision-Making
EVOKE uses goal diversity to elicit LLM agents' pretrained world knowledge for better transfer.
EVOKE is a post-training method for LLM agents that holds environment state and history fixed while ranking the same candidate actions under alternative goals. The authors argue this pressure elicits pretrained world knowledge that single-goal supervision leaves unused, without training a separate observation predictor. Across diverse tasks and three backbones, they report better task performance, generalization to unseen environments, and improved data efficiency. Controlled analyses examine what drives those gains.
- EVOKE ranks the same actions under alternative goals at a fixed state
- Method aims to elicit pretrained world knowledge rather than train a predictor
- Reported gains in performance, unseen-environment transfer, and data efficiency
- Evaluated across diverse tasks using three model backbones
From the original article
1 min read at huggingface.co
Large language models (LLMs) are increasingly deployed as agents for multi-step decision-making, yet transfer poorly to unseen environments. World-model methods address this by training agents to predict future observations, at the cost of additional training and errors that compound when predictions are used for planning.
By Yuhan Guo, Jinming Liu, Liang Xu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.