Finding Side-Effect Loopholes in AppWorld and WorkArena Verifiers
Richard Abrich
原文摘要
Language-model agents act for people inside software, sending payments or editing records. Benchmarks such as AppWorld and WorkArena grade them with automated verifiers, whose verdicts also become reinforcement learning (RL) rewards, so a lenient verifier rewards mistakes. Verifier audits mostly hunt for runs that pass without doing the task; we ask about runs that do it and also change something nobody asked for. We mutation-test both verifiers with hand-written edits to correct runs, designed from the verifier source; an end-state check confirms each false positive. All confirmed false positives were unrequested changes; planted omissions and wrong values produced none. AppWorld's tests passed a duplicate write in all six tasks of two templates, and three RL methods that train on one of these templates would fully reward it; one count assertion per template rejected all six duplicates while passing every correct run. WorkArena's form verifier, called once after submission, passed all 36 edited runs that filled an unrequested field. Checked after every field, it caught extra fields filled first or in drop-downs but still passed all 8 runs that filled a text or reference field last. Both verifiers check that the requested change happened but not reliably that nothing else changed.
cs.AI · cs.SE