Shu Wu

CRANE achieves a remarkable 96.9% Grounded Success in knowledge editing for reasoning MLLMs, overcoming traditional failure modes that plague existing methods.

EAPO enables agents to learn when to forgo tool use, achieving a remarkable 10.45% performance boost while slashing tool calls by over 18%.

Achieve state-of-the-art results in agentic knowledge base question answering by distilling gold-action policies into on-policy student rollouts, bridging the gap between sparse rewards and weakly supervised intermediate actions.

LLM unlearning via counterfactual tuning can backfire, increasing hallucination rates in unexpected areas due to inconsistencies in the "fake" knowledge it's trained on.

LLMs can withstand 3,000 sequential knowledge edits without catastrophic forgetting, thanks to a new sparse editing framework that surgically manipulates knowledge circuits.

Get robust and interpretable models by combining adversarial training with explanation-guided learning, achieving a whopping +37% boost in adversarial accuracy on OOD data.