Spurious Tool Use: When RL Agents Learn the Wrong Reason to Act
Understanding and mitigating shortcut tool selection in RL-trained agents through rewards for tool necessity.
•
1 min read
Read more
Understanding and mitigating shortcut tool selection in RL-trained agents through rewards for tool necessity.
Benchmarking LVLM robustness to spurious correlations and studying generalization beyond the SpuriVerse.