Emergent inabilities? Inverse scaling over the course of pretraining
This highlights a critical issue for AI researchers and developers, as it shows that training can inadvertently degrade model capabilities on certain benchmarks, necessitating comprehensive testing.
The study investigated whether language models can exhibit decreased performance on specific tasks during training, even as general performance improves, and found that Pythia 12B showed such inverse scaling on 8 tasks, with 5 tasks demonstrating that larger models experience greater performance declines with more training.
Does inverse scaling only occur as a function of model size, or can it also occur over the course of training? We carry out an exploratory study investigating whether the performance of language models on specific tasks can decrease (while general performance remains high) during training on the language modeling task. We find 8 tasks on which Pythia 12B (Biderman et al., 2023) shows decreased performance over the course of training. Five of these tasks (TruthfulQA-MC1, TruthfulQA-MC2, Hindsight Neglect, Memo Trap, and Pattern Match Suppression) additionally show a consistent relationship whereby larger language models show a greater decrease in performance the more they are trained, despite showing standard (positive) scaling overall. This highlights the importance of testing performance at all relevant benchmarks any time models are trained on additional data, even if their overall performance improves