Symbol tuning improves in-context learning in language models
Wei, Jerry, Hou, Le, Lampinen, Andrew, Chen, Xiangning, Huang, Da, Tay, Yi, Chen, Xinyun, Lu, Yifeng, Zhou, Denny, Ma, Tengyu, Le, Quoc V.
–arXiv.org Artificial Intelligence
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 Flan-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 semantic knowledge.
arXiv.org Artificial Intelligence
Dec-30-2023
- Country:
- Africa (1.00)
- Asia > Middle East
- Iraq (0.67)
- Europe > Switzerland
- North America > United States
- California > San Francisco County > San Francisco (0.14)
- Genre:
- Research Report > New Finding (0.67)
- Industry:
- Leisure & Entertainment > Sports
- Baseball (0.92)
- Media
- Banking & Finance > Trading (0.67)
- Automobiles & Trucks (0.67)
- Education (0.68)
- Materials (0.92)
- Semiconductors & Electronics (1.00)
- Government
- Immigration & Customs (1.00)
- Military (1.00)
- Regional Government
- Europe Government (0.67)
- North America Government > United States Government (1.00)
- Health & Medicine
- Consumer Health (0.92)
- Pharmaceuticals & Biotechnology (1.00)
- Therapeutic Area
- Immunology (1.00)
- Infections and Infectious Diseases (0.67)
- Law (1.00)
- Energy (0.67)
- Information Technology (1.00)
- Consumer Products & Services (1.00)
- Law Enforcement & Public Safety > Crime Prevention & Enforcement (1.00)
- Leisure & Entertainment > Sports
- Technology: