Is It Worth the (Environmental) Cost? Limited Evidence for Temporal Adaptation via Continuous Training
This work addresses the sustainability problem for AI practitioners by showing incremental evidence against the assumed benefits of continuous training, highlighting a lack of suitable benchmarks.
The paper investigates whether continuous training of language models to adapt to evolving language provides measurable benefits that justify the environmental cost, finding that temporally adapted models for English social media data do not improve downstream task performance over time, with pretrained models being more effective and efficient.
Language is constantly changing and evolving, leaving language models to become quickly outdated. Consequently, we should continuously update our models with new data to expose them to new events and facts. However, that requires additional computing, which means new carbon emissions. Do any measurable benefits justify this cost? This paper looks for empirical evidence to support continuous training. We reproduce existing benchmarks and extend them to include additional time periods, models, and tasks. Our results show that the downstream task performance of temporally adapted English models for social media data do not improve over time. Pretrained models without temporal adaptation are actually significantly more effective and efficient. However, we also note a lack of suitable temporal benchmarks. Our findings invite a critical reflection on when and how to temporally adapt language models, accounting for sustainability.