Qi Li

Small visual perturbations can cause World-Action Models to execute harmful actions while still predicting a plausible future, revealing a critical vulnerability in their design.

LLMs can misinterpret benign text as physically dangerous actions, but a new probing method achieves over 99% accuracy in identifying these risks without relying on explicit unsafe keywords.

Predicting ALS progression with a digital twin model reveals that lower limb function is the strongest predictor of wheelchair access, transforming how we approach patient care.

Current MLLM benchmarks are missing the forest for the trees: Agentic-MME reveals that strong final-answer accuracy masks surprisingly poor tool use and planning in complex multimodal tasks.

Forget hand-crafted membership inference attacks - AutoMIA learns better strategies automatically, adapting to different models and eliminating the need for manual feature engineering.

SAM's impressive zero-shot segmentation abilities don't directly translate to medical imaging, but this new fine-tuning approach unlocks its potential for accurate nuclei instance segmentation with minimal added parameters.

LongCat-Next shatters the language-centric paradigm by unifying text, vision, and audio into a single autoregressive model with minimal modality-specific design, finally reconciling understanding and generation in discrete vision modeling.

VLMs' hallucinations aren't just errors, but traceable pathologies in their "cognitive trajectory," diagnosable via geometric anomalies in a learned state space.