Jiaming Zhang

Balanced education research proposals yield higher publication rates but paradoxically struggle with citation impact, challenging traditional metrics of academic success.

ViTs can achieve robust generalization through adversarial training even when overfitting, mirroring a phenomenon previously observed only in CNNs.

Projector fine-tuning, commonly used for aligning MLLMs, unexpectedly introduces backdoor vulnerabilities with activation mechanisms distinct from those in text-only LLMs.

Panoramic vision-language models can achieve a level of holistic scene understanding and robustness in adverse conditions that's impossible for traditional pinhole-based VLMs.

Agentic LLMs are far more vulnerable to indirect prompt injection attacks than previously thought: AdapTools achieves over 2x improvement in attack success while significantly degrading system utility, even against strong defenses.

LLM agents can now defend against indirect prompt injection attacks without sacrificing task performance, thanks to a new method that surgically manipulates attention based on latent space analysis.

Data valuation for LLMs doesn't need backpropagation: a single forward pass is enough to match gradient-based methods, unlocking massive speedups.