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

Researchers introduced GUI-CIDER, a three-stage mid-training method for GUI agents.

It turns GUI trajectories into textual knowledge, then reselects informative, non-redundant exemplars before retraining.

The approach improves GUI understanding and task success across multiple benchmarks.

It addresses a core gap in GUI agents by teaching explicit interface-operation knowledge, not just post-training skills.