Xin Song

The paper introduces OS-BLIND, a benchmark demonstrating that current safety evaluations fail to detect critical vulnerabilities in computer-use agents when user instructions are benign, showing high attack success rates even for safety-aligned models.

The paper theoretically analyzes the limitations of parameter-based knowledge editing and empirically demonstrates that these methods consistently damage core LLM capabilities compared to retrieval-based baselines.

The paper proposes a novel probabilistic globally constrained decoding (P-GCD) method that efficiently constructs proposals for locally constrained decoding, significantly improving convergence speed and performance compared to existing approaches.