Reproduce “Rethinking the Outlier Distribution in Large Language Models: An In-depth Study” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Rahul Raman, Khushi Sharma, Sai Qian Zhang
This study from New York University empirically investigates the numerical origins and propagation of outliers in large language models at a fine-grained, operator level, revealing that most massive activations can be removed without accuracy loss, and pinpoints normalization rescaling and specific weight matrix channels as primary sources of channel-wise outliers.
Key insights
- Most massive activations observed across LLM layers are 'Fake MAs', propagated through residual connections from earlier layers, and can be safely eliminated without impacting model accuracy.
- Channel-wise outliers primarily emerge from the learnable rescaling operation (γX + β) within normalization layers, rather than the standardization step.
- Specific 'Outlier Triggering Channels' within the Query (Wq) and Key (Wk) weight matrices disproportionately contribute to the generation of new channel-wise outliers and are critical for LLM performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.