Symbol tuning improves in-context learning in language models

Abstract: We present symbol tuning - finetuning LLMs 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.

• The paper demonstrates that symbol tuning improves in-context learning by forcing models to rely on input-label mappings instead of semantic cues.
• The paper shows that the methodology enhances algorithmic reasoning, with symbol-tuned models achieving better performance on tasks like list functions and Turing problems.
• The paper finds that symbol tuning helps models override pre-existing semantic knowledge, reducing dependency on prompt engineering for robust task execution.

Symbol Tuning and Improved In-Context Learning in LLMs

The paper introduces the concept of symbol tuning as a finetuning technique for LLMs, especially those with substantial parameters, such as Flan-PaLM. This approach involves substituting natural language labels with abstract symbols during the finetuning process of LLMs on input-label pairs. This method is motivated by the hypothesis that models, when deprived of semantic cues, would need to rely on the input-label mapping to decode task requirements instead of depending on inherent semantic knowledge or explicit instructions.

Experimental Findings

Upon applying symbol tuning to various Flan-PaLM models, including configurations with up to 540 billion parameters, notable improvements were observed:

1. Performance in In-Context Learning Tasks: Models achieved better accuracy in tasks where traditional semantic labels or instructions were unavailable at the prompt. Symbol-tuned models exhibited substantial gains, especially in circumstances devoid of clear task framing, exemplifying their enhanced ability to leverage in-context information for task resolution.
2. Algorithmic Reasoning: Symbol-tuned models showed improved performance in algorithmic reasoning tasks, such as those involving list functions and simple Turing concepts, without specific prior tuning for numeric or logic-based tasks. This suggests increased versatility and adaptation, even on tasks differing significantly from the finetuning data.
3. Overriding Prior Knowledge: These models demonstrated significant ability to override their pre-existing semantic knowledge when tasked with following flipped-labels during evaluations, an ability diminished or lost in standard instruction-tuned models.

Practical and Theoretical Implications

Symbol tuning provides a pathway to reduce dependency on prior semantic knowledge and instruction engineering, potentially minimizing prompt engineering complexities while increasing the task adaptability and generalization of LLMs. The superior performance of symbol-tuned models, despite arbitrary labeling, underscores their potential for executing tasks without explicit semantic prompts. Practically, this can streamline applications in AI systems, making them robust to variations in labeling and framing which can occur across real-world scenarios.

Future Directions in AI Development

Future work could explore expanding the breadth of tasks and datasets employed in the symbol-tuning process, potentially integrating elements of natural language processing with algorithmic and numerical data. Further research might also investigate how these models can leverage the symbol-tuning approach for multi-modal or complex reasoning tasks that involve logical deduction intertwined with semantic comprehension. Additionally, analyzing the role of model architecture in enhancing or limiting the efficacy of symbol tuning remains a vital area of exploration.

In summary, symbol tuning constitutes a promising avenue to refine in-context learning in LLMs, enhancing their flexibility and reducing reliance on predefined semantic structures, ultimately broadening the scope and efficiency of large-scale LLM applications.