Xing-Jun Ma

HazardAuditor is introduced, an execution-grounded framework that runs heterogeneous agents in controlled environments and normalizes their interactions into a canonical event representation for cross-framework supervision, and observes that token-level post-training objectives create a structural mismatch for generative guards, causing longer rationales to dominate gradient updates.

Sci-MMR is introduced, a benchmark for multi-step evidence-grounded scientific reasoning built on structured argument graphs linking scientific claims, citation-grounded knowledge, visual evidence, and supporting regions, and it is found that current answer-centric benchmarks substantially overestimate the evidence-grounded reasoning capabilities of multimodal research agents.

Task-specific image editing can boost MLLM performance by over 29%, but not all tasks benefit equally—some are left behind.

Adversarial techniques traditionally seen as threats are now being repurposed by content owners to proactively safeguard their visual assets from misuse.

Environment evolution can reveal 17% more safety failures in complex tasks compared to static benchmarks, reshaping our understanding of agent vulnerabilities.

Vera reveals that existing LLM agents exhibit up to 93.9% vulnerability to multi-channel attacks, highlighting a significant gap in current safety evaluations.

ShutterMuse redefines photography guidance by combining composition and pose recommendations, outperforming existing models while cutting inference costs.

Self-evolving LLMs can amplify adversarial threats, making every known attack lineage-persistent and exposing critical vulnerabilities that static defenses can't address.

SentGuard detects 90.5% of unsafe content within two sentences, revolutionizing real-time moderation for large language models.

Guard models trained with BraveGuard can detect safety threats in computer-use agents with over 82% accuracy, a significant leap from conventional methods.

Even state-of-the-art multimodal LLMs like GPT-5.2 and Claude 4.5 can be jailbroken nearly half the time using OpenRT's diverse suite of attacks, revealing a critical lack of generalization across attack paradigms.