Understanding Emergent Abilities of Language Models from the Loss Perspective
This work addresses the challenge of predicting emergent abilities in AI models for researchers and practitioners, offering a more reliable metric than model size, though it is incremental in redefining the concept.
The paper tackles the problem of understanding emergent abilities in language models by proposing to study them from a pre-training loss perspective rather than model size, demonstrating that models with the same pre-training loss achieve similar performance on downstream tasks regardless of size, and discovering that emergent abilities manifest when pre-training loss falls below a specific threshold, with performance jumping from random guessing to higher levels.
Recent studies have put into question the belief that emergent abilities in language models are exclusive to large models. This skepticism arises from two observations: 1) smaller models can also exhibit high performance on emergent abilities and 2) there is doubt on the discontinuous metrics used to measure these abilities. In this paper, we propose to study emergent abilities in the lens of pre-training loss, instead of model size or training compute. We demonstrate that the Transformer models with the same pre-training loss, but different model and data sizes, generate the same performance on various downstream tasks, with a fixed data corpus, tokenization, and model architecture. We also discover that a model exhibits emergent abilities on certain tasks -- regardless of the continuity of metrics -- when its pre-training loss falls below a specific threshold. Before reaching this threshold, its performance remains at the level of random guessing. This inspires us to redefine emergent abilities as those that manifest in models with lower pre-training losses, highlighting that these abilities cannot be predicted by merely extrapolating the performance trends of models with higher pre-training losses.