Symbol tuning improves in-context learning in language models

Jerry Wei, Le Hou, Andrew Kyle Lampinen, Xiangning Chen, Da Huang, Yi Tay, Xinyun Chen, Yifeng Lu, Denny Zhou, Tengyu Ma, Quoc V Le

Submission Type: Regular Long Paper

Submission Track: Language Modeling and Analysis of Language Models

Submission Track 2: Theme Track: Large Language Models and the Future of NLP

Keywords: in-context learning, large language models, natural language processing

TL;DR: Symbol tuning is a simple method that emphasizing input–label mappings to improve robustness to prompts without instructions/relevant labels and boost performance on algorithmic reasoning tasks.

Abstract:

We present symbol tuning - finetuning language models on in-context input-label pairs where natural language labels (e.g., "positive/negative sentiment") are replaced with arbitrary symbols (e.g., "foo/bar"). Symbol tuning leverages the intuition that when a model cannot use instructions or natural language labels to figure out a task, it must instead do so by learning the input-label mappings.

We experiment with symbol tuning across PaLM models up to 540B parameters and observe benefits across various settings. First, symbol tuning boosts performance on unseen in-context learning tasks and is much more robust to underspecified prompts, such as those without instructions or without natural language labels. Second, symbol-tuned models are much stronger at algorithmic reasoning tasks, with up to 18.2% better performance on the List Functions benchmark and up to 15.3% better performance on the Simple Turing Concepts benchmark. Finally, symbol-tuned models show large improvements in following flipped-labels presented in-context, meaning that they are more capable of using in-context information to override prior knowledge.

Submission Number: 162