Jing-Yu Zhang

LLM agents can achieve high task accuracy while exhibiting profound epistemic *im*humility, failing to acknowledge uncertainty even when presented with contradictory evidence.

Latent communication between LLMs was largely an illusion of interface shortcutting until Draft-KV, which routes genuine drafting KV caches to let a frozen 0.5B model leap from 37% to 78% MMLU-Redux accuracy by tapping an 8B model's latent thoughts.

Leading models fall short in emotional intelligence, with EmoS achieving 83.8% accuracy and nearing human-level performance.

Combining text, images, and audio from academic papers can dramatically boost keyword extraction accuracy, revealing the untapped potential of multimodal data.

SelfCompact reveals that language models can autonomously manage context decay, achieving up to 18.1 points improvement in performance while cutting token costs by 30-70%.

MiniMax-M2 proves that massive parameter counts don't always translate to better agentic performance; strategic activation of a smaller subset can unlock frontier-level intelligence.

Even state-of-the-art LLMs struggle to follow complex instruction hierarchies, achieving only ~40% accuracy when navigating conflicts across a dozen privilege levels in agentic tasks.

LLMs struggle to navigate the nuances of real-world rules, achieving only ~45% accuracy on a new benchmark of legal and policy reasoning tasks.