An Empirical Study and Theoretical Explanation on Task-Level Model-Merging Collapse

Researchers have identified a phenomenon called 'merging collapse' where combining independently fine-tuned large language models leads to catastrophic performance degradation. The study reveals that representational incompatibility between tasks, rather than parameter conflicts, is the primary cause of merging failures.

• →Model merging can fail catastrophically when combining certain task-specialist LLMs, a phenomenon termed 'merging collapse'.
• →Representational incompatibility between tasks is strongly correlated with merging collapse, challenging conventional wisdom about parameter-space conflicts.
• →The failure occurs consistently across different merging methods when certain task combinations are attempted.
• →Researchers provide theoretical explanation through rate-distortion theory establishing fundamental limits on task mergeability.
• →The findings suggest that not all independently developed AI models can be successfully merged, regardless of the merging methodology used.