Jeonghyun Park

Code-switching in multilingual models reveals a surprising grammar-frame effect that can significantly degrade performance, but a new intervention can effectively mitigate these issues.

LLMs are surprisingly susceptible to irrelevant framing details, flipping decisions nearly 30% of the time, and naive attempts to fix it only make things worse.

LLMs exhibit a surprising "structural blind spot" that causes them to fixate on initial actions and fail to explore diverse solutions when faced with pressure to converge.