Stephen Peng, Vineet Jagadeesan Nair, Noah Broestl, Anne Friederike Kleppe, Steven Mills, Adi Zolotov
原文摘要
AI agents can pass checks on individual tasks while retaining memory or code changes that alter their later behavior. We introduce a control-theoretic framework that couples a fast loop controlling proposed actions with a slow loop reviewing persistent changes and reassessing configurations after model or tool updates. When measurement errors and changes between checks are bounded and controllers can enforce their decisions, we derive a deployment-level bound and a rule for dividing review effort. In simulation, dividing a fixed review budget between both loops reduces mean maximum departure from intended behavior roughly twelvefold compared with review confined to individual runs. In coding stress tests with authored memories and five matched seeds, combined controls reduce mean severity on a 0-3 scale from 0.325 without gates to 0.117, resolving 33 of 120 issues versus 35 without gates; in trading stress tests, combined controls reduce mean shortfall in low-emission purchases from 0.892 without gates to 0.107. The overall framework helps developers and auditors compare and enforce oversight designs, while the agent studies test component controls rather than the full deployment-level guarantee.
Stephen Peng, Vineet Jagadeesan Nair, Noah Broestl, Anne Friederike Kleppe, Steven Mills, Adi Zolotov
原文摘要
The number of domain fine-tuned models, particularly in the biomedical domain, is rapidly expanding. Fine-tuning promises high-performing models at a fraction of the cost of training a base model, but it can also produce unintended behavioral changes. Multiple studies have shown that safety guardrails can erode with fine-tuning, yet large-scale studies across model architectures, data mixtures, and fine-tuning methodologies remain limited. Where prior work in this space has studied value-centric refusal, this work focuses on knowledge-centric abstention. We audit 25 biomedical domain fine-tuned large language models against 11 paired base checkpoints on MedAbstain and ClinDet-Bench to evaluate the impact of fine-tuning on knowledge-centric abstention. We find that defined appropriate abstention measured by these benchmarks, already low in base models (median 8.3% on MedAbstain), falls further after fine-tuning (median 4.2%, decreasing in 21 of 25 pairs on MedAbstain), and that the magnitude and direction of change are heterogeneous across pairs. These results motivate fine-tuning methods that improve domain-specific capability while preserving appropriate caution.