Wang, Z. (CSE) – From Static Alignment to Adaptive Safety: Toward Reliable and Capable AI Systems

Modern AI systems are rapidly moving beyond static text generation toward capable models and agents that reason, use tools, store memories, and update persistent state, yet safety methods still often assume a fixed model whose behavior can be controlled by output-level refusal. This leaves critical gaps in understanding why aligned models fail under adversarial pressure, how to align reasoning models without suppressing their useful capabilities, and how to preserve safety once capability and control are externalized into editable agent state. My research proposes a static-to-adaptive safety framework for building reliable and capable AI systems: studying the mechanisms that shape behavior inside models, using reasoning capability as a substrate for safety alignment, and governing persistent state as agents learn and adapt over time. We instantiate this agenda through two completed works and three proposed directions. AttnGCG studies adversarial failures in aligned language models, showing how jailbreak attacks can manipulate model attention and expose limitations of output-level safety analysis. STAR-1 studies safety alignment for large reasoning models, showing that policy-grounded reasoning data can improve safety while largely preserving general reasoning capability. Building on these foundations, we further study when editable agent harnesses meaningfully affect future behavior, how persistent state creates new safety risks, and how adaptive agents can safely update state while preserving useful learning. Together, my research aims to move beyond static alignment alone, toward AI systems whose safety remains reliable as their capabilities expand through reasoning and adaptation.

Event Host: Zijun Wang, Ph.D. Student, Computer Science & Engineering

Advisor: Cihang Xie 

Zoom ID:  962 8317 0929

Passcode: 687715