Sloth: scaling laws for LLM skills to predict multi-benchmark performance across families
This work addresses the challenge for researchers and practitioners in efficiently benchmarking and scaling LLMs across diverse families, though it is incremental as it builds on existing scaling law concepts by incorporating latent skills.
The paper tackles the problem of predicting large language model (LLM) performance across different model families, where existing scaling laws fail to generalize due to variations in training configurations, by proposing Skills Scaling Laws (Sloth) that use publicly available benchmark data and assume performance is driven by low-dimensional latent skills, resulting in more accurate and interpretable predictions without needing to train multiple models per family, as demonstrated on 12 benchmarks from the Open LLM Leaderboard.
Scaling laws for large language models (LLMs) predict model performance based on parameters like size and training data. However, differences in training configurations and data processing across model families lead to significant variations in benchmark performance, making it difficult for a single scaling law to generalize across all LLMs. On the other hand, training family-specific scaling laws requires training models of varying sizes for every family. In this work, we propose Skills Scaling Laws (SSLaws, pronounced as Sloth), a novel scaling law that leverages publicly available benchmark data and assumes LLM performance is driven by low-dimensional latent skills, such as reasoning and instruction following. These latent skills are influenced by computational resources like model size and training tokens but with varying efficiencies across model families. Sloth exploits correlations across benchmarks to provide more accurate and interpretable predictions while alleviating the need to train multiple LLMs per family. We present both theoretical results on parameter identification and empirical evaluations on 12 prominent benchmarks, from Open LLM Leaderboard v1/v2, demonstrating that Sloth predicts LLM performance efficiently and offers insights into scaling behaviors for complex downstream tasks and increased test-time compute.