Jev’s resistance to attacks changes across tasks
Questa pagina non ha ancora una traduzione verificata. È mostrata la versione inglese. Leggi la versione inglese
In parole semplici
The paper compares Jev, a safety judge that answers preset questions with probabilities, with language-model judges and safety classifiers. The authors sent them attacks across tasks and used the same rules to turn their answers into pass, review, or block decisions.
Jev resisted one forged safety message on harmful-request moderation, but that same kind of forgery got 19% of the phishing emails Jev had flagged through. Its resistance depended on the task.
La ricerca
Before this comparison, Jev had been tested under attack without other judges alongside it.
The open question was whether its typed, probability-based judgments changed safety decisions more or less often than language-model judges or purpose-built classifiers facing the same attacker-controlled content across different tasks.
The authors compare Jev with language-model judges and safety classifiers under shared attacks, finding that its resistance depends on the task.
The authors compared eight systems across four tasks, adapting the same attacker-controlled content to each input format and routing outputs through one policy. On clean phishing emails, Jev identified 62% as phishing, similar to two GPT judges but below three GLM judges.
A forged safety message passed none of Jev’s clean-flagged harmful requests in moderation, yet got 19% of its clean-flagged phishing emails through. The contrast supports the authors’ view that Jev is more robust on clear-rule tasks than on overall judgments.
The judgment-based conclusion rests on one English phishing task and corpus, with single-turn calls. The attack measure records whether a judge passes a message, not what an assistant writes afterward or whether harm follows.
Analisi
Dettagli
- Autori del paper originale
- Xiang Li, Aichun Huang, Mingming Zhang, Zheli Liu
- Inviato ad arXiv
- Licenza
- arXiv non-exclusive