TrustSwap Reveals Source-Trust Shortcuts in Fact-Checking RL Agents

Testing shows that RL-trained fact-checkers often rely on source reliability labels rather than evidence content, with label changes altering up to 50% of verdicts. Standard GRPO fine-tuning worsens this shortcut in 8B models, though trust-swap augmentation (TSA) is proposed as a mitigation.