Yahan Yu

Contextual biases in INER can lead to catastrophic forgetting, but a novel Type-Balanced Contextual Learning method effectively mitigates this issue, enhancing model retention of old knowledge.

Projector-level forgetting in MLLMs can be mitigated, leading to significant performance gains in continual instruction tuning without excessive parameter growth.

LLMs don't need "wait, let me think..." to reason—in fact, dropping the cutesy anthropomorphic markers can actually *improve* their performance.