Research reveals LLMs struggle significantly when processing multiple documents simultaneously, experiencing gradual performance decline followed by complete collapse.

LLMs show consistent performance degradation when handling multi-instance inputs, even on tasks where they normally excel individually

Performance decline occurs in two phases: gradual degradation at 20-100 instances, then sharp collapse at higher instance counts

Context length is associated with performance issues but is not the sole factor—instance count plays a crucial independent role

Common real-world applications like analyzing sentiment across multiple movie reviews are negatively impacted by this multi-instance processing limitation