Bo Li

PULSE slashes communication overhead by 89% while boosting training throughput by up to 2.3x, revolutionizing how we scale diffusion models across GPU clusters.

FragFuse reveals how long-term memory in LLMs can be weaponized to bypass access controls, achieving an 86.3% success rate in evading restrictions.

Shapley-guided analysis reveals hidden vulnerabilities in multi-agent systems, enabling targeted and coordinated adversarial attacks that traditional methods miss.

Current LLM-based web agents are vulnerable to prompt-injection attacks, with no reliable defenses against any attack objective, revealing a critical oversight in security evaluations.

Skill-augmented AI agents may enhance the quality of biomedical research outputs, but the improvement is subtle and requires further validation.

Frontier models can't build playable games in one shot, but a closed-loop system using GUI agents to playtest and refine code achieves a 66.8% success rate, proving that game generation needs to be a conversation, not a translation.

Ditching modular architectures unlocks surprisingly competitive vision-language performance, proving that end-to-end pixel-to-word models can rival traditional approaches at scale.

LLaVA-OV-2's codec-stream tokenization lets it crush existing video-language models, especially in tasks requiring fine-grained temporal understanding of high-frequency motion.

LLMs can now translate text in web images with significantly improved accuracy and efficiency thanks to a novel visual-aware adaptation framework that bridges the gap between high-level semantics and fine-grained visual details.

LLMs can learn multilingual translation far more effectively by explicitly separating and routing language modeling and translation knowledge during fine-tuning.

RL fine-tuning LMMs for tool use can collapse structural formats due to strong pretrained tool priors, but a surprisingly simple fix of targeted format rewards and frame-budget randomization can restore stability and boost performance.

AI agents are shockingly easy to manipulate into leaking API keys, deleting user data, and initiating unauthorized transactions across a wide range of real-world applications.

Coordinating LLM agents with evolving knowledge graphs, rather than just text, unlocks superior scientific ideation, beating state-of-the-art systems on multiple benchmarks.