1 articles with this tag
LLMs collapse under simple lexical constraints, revealing fragility in instruction tuning and flawed evaluation methods.