Junk DNA Hypothesis: Pruning Small Pre-Trained Weights $\textit{Irreversibly}$ and $\textit{Monotonically}$ Impairs ``Difficult" Downstream Tasks in LLMs

We presentJunk DNA Hypothesisby adopting a noveltask-centricangle for the pre-trained weights of large language models (LLMs). It has been believed that weights in LLMs contain significant redundancy, leading to the conception that a considerable chunk of the parameters can be removed bypruningwithout compromising performance. Contrary to this belief, this paper presents acounter-argument: small-magnitude weights of pre-trained model weights encode vital knowledge essential for tackling difficult downstream tasks - manifested as themonotonic relationshipbetween the performance drop of downstream tasks across the difficulty spectrum, as we prune more pre-trained weights by magnitude. Moreover, we reveal that these seemingly inconsequential weights can result inirreparable lossof knowledge and performance degradation in difficult tasks, even when downstream continual training is allowed. Interestingly, our evaluations show that the other popular compression, namelyquantizationfailto exhibit similar ``monotonic" effect and does not as convincingly disentangle this task-difficulty information. To study formally, we introduce several quantifiable metrics togauge the downstream task difficulty: (a) within the same task category, and (b) across different task categories. Our extensive experiments substantiate the Junk DNA Hypothesis across a diverse range of model sizes, tasks, datasets, and even pruning methods. Codes are available at https://github.com/VITA-Group/JunkDNAHypothesis.git.