Language as an interface to intelligence
Our research uses language in two complementary ways: to teach intelligent systems and to understand their behavior. This program grew out of Shashank’s PhD dissertation (2018) on learning from natural-language supervision, and now spans three thrusts: interpreting and explaining model behavior, learning through language and interaction, and studying how language and context shape cognition and social behavior.
Interpreting and Explaining Model Behavior
How can we explain what a model has learned, and predict how it will behave?
We develop methods for generating natural-language explanations of model behavior, discovering systematic patterns in model errors, and testing whether explanations are faithful. We also investigate what LLM reasoning traces reveal about the computations that produce their answers, including causal interventions on internal mechanisms such as attention heads.
Selected workUNMASK 2026OPeX 2026CoT faithfulness 2026Deceptive commitment 2026SLED 2025Causal Diagnosticity 2025DISCERN 2024MaNtLE 2023Learning Through Language and Interaction
How can machines learn new concepts and behaviors from explanations, instructions, and questions?
People learn new concepts through instructions, explanations, and conversation, often without extensive direct experience. We investigate how machines can learn in similar ways, including by asking questions and using language to acquire new knowledge. This is the lab’s longest-running thrust.
Selected workINTERACT 2025CLUES 2022Learning to Ask 2019PhD thesis 2018Zero-shot from NL quantification 2018Joint concept learning 2017Language, Cognition and Social Behavior
How does language shape reasoning, decision-making, and social behavior, in AI systems and in people?
Language does more than communicate information: it shapes how people and intelligent systems reason, make decisions, and interact. We study these effects in LLMs, including how conversational context, personas, and narrative framing shape model behavior, and what LLMs mean for human communication. We also look for cognitive regularities in language itself, such as sound symbolism across languages.
Selected workNarrative priors 2026Chameleon LLMs 2025Epistemic foundations 2025SocialGaze 2024Open questions
- Can we explain a complex model well when we can only observe what it does, and cannot inspect its internals?
- Can we tell when a language model becomes committed to a decision while it reasons, and intervene before that decision shows up in its answer?
- How can a model ask for the explanations and information it needs, instead of relying only on static training data?
Working with us
We welcome collaborations in domains where understanding an AI system matters as much as measuring its accuracy. Our work on interpretability, behavioral explanation, faithful reasoning, and interactive learning can be useful wherever AI systems affect consequential decisions or scientific inquiry.
Our current collaborations span education, public health and HIV prevention, and AI for governance and civic deliberation. We are also interested in new partnerships in health, science, and other domains where questions about explanation, reliability, interaction, or social context can motivate new machine-learning research.
- Interpretable AI in consequential settings. Identifying, explaining, and testing the failures of AI systems used in sensitive settings.
- Interactive learning from domain experts. Systems that learn from questions, explanations, and feedback.
- Language, context, and social behavior in AI systems. How framing, persuasion, and participation shape, and are shaped by, AI.
Faculty and researchers who would like to explore a project are welcome to write to Shashank at ssrivastava@cs.unc.edu.
Code and data
- CLUES ACL 2022 codedata Benchmark and code for learning classifiers from natural-language explanations.
- INTERACT ACL 2025 code Code for teaching language models to learn interactively, by asking a teacher questions.
- MaNtLE EMNLP 2023 code Code for a model-agnostic explainer that describes how a classifier behaves, in natural language.
- DISCERN EMNLP 2024 code Code for explaining a text classifier's systematic errors in natural language, and using the explanations to improve it.
- Causal Diagnosticity EMNLP 2025 code Code for a testbed that compares faithfulness metrics for natural-language explanations.
All code and data
- Chameleon LLMs EMNLP 2025 code
- Causal Diagnosticity EMNLP 2025 code
- Unreliable narrators ACL 2025 code
- INTERACT ACL 2025 code
- SocialGaze EMNLP 2024 code
- Fuse to Forget EMNLP 2024 code
- DISCERN EMNLP 2024 code
- Leveraging Multiple Teachers for Test-Time... EMNLP 2023 code
- Pragmatic Reasoning Unlocks Quantifier Semantics... EMNLP 2023 code
- MaNtLE EMNLP 2023 code
- LaSQuE ACL 2023 code
- Beyond the Imitation Game TMLR 2023 codedata
- What do Large Language Models... EMNLP 2022 code
- Compositional Generalization for Kinship Prediction... WNU @ NAACL 2022 code
- Predicting Difficulty and Discrimination of... ACL 2022 codedata
- ePiC ACL 2022 codedata
- CLUES ACL 2022 codedata
- Mapping Language to Programs using... EMNLP 2021 code
- Adversarial Scrubbing of Demographic Information... EMNLP 2021 code
- Improving and Simplifying Pattern Exploiting... EMNLP 2021 code
- How Helpful is Inverse Reinforcement... ACL 2021 code
- PRover EMNLP 2020 code
- Learning Web-based procedures by Reasoning... ACL 2020 data