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:
- Oceania
- New Zealand (0.04)
- Marshall Islands (0.04)
- Micronesia (0.04)
- Australia (0.04)
- North America
- United States
- Arizona (0.04)
- Florida (0.04)
- Utah (0.04)
- Oklahoma (0.04)
- Texas
- Travis County > Austin (0.04)
- Dallas County > Dallas (0.04)
- Massachusetts > Middlesex County
- Cambridge (0.04)
- Missouri > Jackson County
- Kansas City (0.04)
- Alaska > Denali Borough
- Mt Mckinley (0.04)
- Illinois
- Cook County > Chicago (0.04)
- Champaign County > Urbana (0.04)
- Washington > King County
- Seattle (0.04)
- Kansas > Wyandotte County
- Kansas City (0.04)
- New Jersey > Monmouth County
- Asbury Park (0.04)
- California
- San Francisco County > San Francisco (0.14)
- Los Angeles County > Los Angeles (0.04)
- New York > New York County
- New York City (0.04)
- Wisconsin > Milwaukee County
- Milwaukee (0.04)
- Cuba > La Habana Province
- Havana (0.04)
- Canada
- Ontario > Toronto (0.04)
- Manitoba > Winnipeg Metropolitan Region
- Winnipeg (0.04)
- United States
- Europe
- United Kingdom (0.27)
- Russia (0.14)
- Germany (0.04)
- Middle East > Cyprus (0.04)
- Eastern Europe (0.04)
- Spain > Galicia
- Madrid (0.04)
- Switzerland > Zürich
- Zürich (0.14)
- France
- Île-de-France > Paris
- Paris (0.04)
- Pays de la Loire > Loire-Atlantique
- Nantes (0.04)
- Île-de-France > Paris
- Belgium > Flanders
- Flemish Brabant > Leuven (0.04)
- Italy > Emilia-Romagna
- Metropolitan City of Bologna > Bologna (0.04)
- Asia
- Russia (0.14)
- Taiwan (0.04)
- Afghanistan (0.04)
- Kazakhstan (0.04)
- Pakistan (0.04)
- China
- Middle East
- Republic of Türkiye (0.04)
- Iran (0.04)
- Iraq > Baghdad Governorate
- Baghdad (0.04)
- India > Karnataka
- Bengaluru (0.04)
- South Korea > Daejeon
- Daejeon (0.04)
- Japan > Honshū
- Kantō > Tokyo Metropolis Prefecture > Tokyo (0.14)
- Africa
- South Africa (0.04)
- Nigeria (0.04)
- Kenya (0.04)
- Côte d'Ivoire > Abidjan
- Abidjan (0.04)
- Oceania
- Genre:
- Research Report > New Finding (0.67)
- Industry:
- Law Enforcement & Public Safety > Crime Prevention & Enforcement (1.00)
- Consumer Products & Services (1.00)
- Information Technology (1.00)
- Semiconductors & Electronics (1.00)
- Materials (0.92)
- Automobiles & Trucks (0.92)
- Education (0.68)
- Law > Statutes (0.67)
- Banking & Finance > Trading (0.67)
- Health & Medicine
- Pharmaceuticals & Biotechnology (1.00)
- Consumer Health (0.92)
- Therapeutic Area
- Immunology (1.00)
- Infections and Infectious Diseases (0.67)
- Government
- Military (1.00)
- Immigration & Customs (1.00)
- Regional Government
- North America Government > United States Government (1.00)
- Europe Government (0.67)
- Media
- Leisure & Entertainment > Sports
- Baseball (0.92)
- Technology: