GUI-CIDER: Mid-training GUI Agents via Causal Internalization and Density-aware Exemplar Reselection

Researchers introduced GUI-CIDER, a mid-training method for GUI agents that turns GUI trajectory knowledge into text.

It filters examples to preserve causal structure and reduce redundancy, then uses the refined data for training.

On two GUI knowledge benchmarks and three task completion benchmarks, it improved GUI operation understanding and success rates.

The approach addresses a key weakness in GUI agents by explicitly teaching GUI operation knowledge beyond standard fine-tuning or reinforcement learning.