Reproduce “Performance Prediction for Large Systems via Text-to-Text Regression” 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
Yash Akhauri, Bryan Lewandowski, Cheng-Hsi Lin, Adrian N. Reyes, Grant C. Forbes, Arissa Wongpanich, Bangding Yang, Mohamed S. Abdelfattah and 2 more
A new Text-to-Text Regression approach leverages Regression Language Models to predict performance for large industrial systems. It achieves up to a 0.99 Spearman rank correlation and 100x lower MSE compared to tabular methods on Google's Borg compute cluster by processing raw textual system data.
Key insights
- Maximizing feature observability by directly consuming raw, complex system data (e.g., nested profiles) significantly reduces epistemic uncertainty and boosts predictive performance.
- Encoder-decoder architectures are crucial for processing intricate input sequences in regression tasks, outperforming decoder-only models despite their prevalence in general LLMs.
- Regression Language Models exhibit strong transfer learning capabilities, adapting effectively to new, unseen tasks with minimal (e.g., 500) fine-tuning examples after diverse pretraining.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.