Foundations of AI
https://allenai.org/foundations-of-ai (opens in a new tab)
Understanding LLM training data
Large text corpora are the backbone of today’s language models, but we have a limited understanding of the content of these datasets. Our research in this area seeks to uncover important facts about these popular corpora including general statistics, quality, social factors, and potential contamination by evaluation data.
What's In My Big Data? (opens in a new tab)
Retrieval-augmented generation
Retrieval-augmented generation (RAG) improves the responses of large language models by allowing them to access an additional authoritative knowledge source outside of their training data. Our work in this field offers ways to make RAG more scalable, performant, and more respectful of data concerns like copyright.
- SILO Language Models: Isolating Legal Risk In a Nonparametric Datastore (opens in a new tab)
- Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection (opens in a new tab)
- Scaling Retrieval-Based Language Models with a Trillion-Token Datastore (opens in a new tab)
Human-AI interaction
To realize the full promise of AI, it is critical to design user interfaces that support effective collaboration with human users. This research area explores a variety of novel interfaces for humans that maximize the helpfulness of AI when engaging with scientific literature, supporting better access, rapid and deep interactive information gathering, and more.
- The Semantic Reader Project: Augmenting Scholarly Documents through AI-Powered Interactive Reading Interfaces (opens in a new tab)
- FigurA11y: AI Assistance for Writing Scientific Alt Text (opens in a new tab)
- Qlarify: Recursively Expandable Abstracts for Directed Information Retrieval over Scientific Papers (opens in a new tab)
Theoretical insights about LLMs
While language models are somewhat opaque, it is possible to apply theoretical techniques to analyze their intrinsic capabilities and limitations, yielding important fundamental insights about such systems.
- The Expressive Power of Transformers with Chain of Thought (opens in a new tab)
- A Logic for Expressing Log-Precision Transformers (opens in a new tab)
- The Illusion of State in State-Space Models (opens in a new tab)
Intelligent language agents
As well as answering questions, language models can also act as intelligent agents, interacting autonomously with an external environment to perform complex tasks. Our research focuses on having such agents plan and learn in these environments in order to rapidly improve their performance.
- CLIN: A Continually Learning Language Agent for Rapid Task Adaptation and Generalization (opens in a new tab)
- Skill Set Optimization: Reinforcing Language Model Behavior via Transferable Skills (opens in a new tab)
- ADaPT: As-Needed Decomposition and Planning with Language Models (opens in a new tab)
Systematic reasoning with language
While language models are innately good at question-answering, our research has developed new methods for enabling them to reason systematically and to arrive at conclusions in a sound and explainable way.