# Jacob Andreas Jacob Andreas is an assistant professor at MIT CSAIL researching language grounding and compositional semantics, with contributions to machine learning and natural language processing. Jacob Andreas is an assistant professor at the [MIT Computer Science and Artificial Intelligence Laboratory](https://www.wikiprompt.org/wiki/mit-csail) (MIT CSAIL). His research focuses on language grounding and compositional semantics, exploring how [AI](https://www.wikiprompt.org/wiki/artificial-intelligence) systems can connect linguistic structure to real-world meaning and action. He is affiliated with MIT's Department of Electrical Engineering and Computer Science. Andreas completed his PhD at the [University of California, Berkeley](https://www.wikiprompt.org/wiki/berkeley-ai-research), where he was advised by Dan Klein. His doctoral work examined how compositional structures in language can be learned and used for tasks such as visual question answering and instruction following. Before joining MIT, he held a postdoctoral position at [Google DeepMind](https://www.wikiprompt.org/wiki/google-deepmind). ## Research Contributions Andreas's early research introduced methods for learning modular neural networks that compose sub-tasks in a structured way. In a 2016 paper, he and colleagues proposed a model that learns to decompose a question into smaller sub-questions, each answered by a specialized module, and then combines the results. This approach demonstrated improved performance on visual question-answering benchmarks compared to monolithic [neural networks](https://www.wikiprompt.org/wiki/neural-network). His work on compositional semantics has also addressed how to induce latent linguistic structures from data without explicit annotations. He has studied the use of [large language models](https://www.wikiprompt.org/wiki/large-language-model) for semantic parsing and grounding, examining how these models can be steered to produce interpretable and compositional representations. ## Language Grounding and Instruction Following A central theme in Andreas's research is grounding language in interactive environments. He has investigated how agents can learn to follow natural language instructions in simulated worlds, using reinforcement learning and [curriculum learning](https://www.wikiprompt.org/wiki/curriculum-learning) to improve sample efficiency. His work has shown that compositional task descriptions can be leveraged to generalize to unseen combinations of instructions. Andreas has also explored the relationship between language and formal logic, proposing ways to translate natural language into executable programs. This line of research connects to broader efforts in [machine learning](https://www.wikiprompt.org/wiki/machine-learning) to build systems that reason explicitly about structure. ## Selected Publications and Recognition Andreas has published at major venues including NeurIPS, ICLR, ACL, and EMNLP. His papers have received multiple best-paper nominations and awards. He was named a Sloan Research Fellow in 2021, recognizing his early-career contributions to computer science. His teaching at MIT includes graduate courses on natural language processing and machine learning. He has also given invited talks at academic institutions and industry research labs, including [OpenAI](https://www.wikiprompt.org/wiki/openai) and [Anthropic](https://www.wikiprompt.org/wiki/anthropic). ## Broader Impact and Current Directions Andreas's recent work has examined the interpretability of [transformer](https://www.wikiprompt.org/wiki/transformer) models, particularly how attention mechanisms encode compositional information. He has argued for the importance of designing benchmarks that test systematic generalization, a challenge highlighted by the [meta-learning](https://www.wikiprompt.org/wiki/meta-learning) community. He is also interested in the intersection of language and probabilistic programming, aiming to build models that can reason about uncertainty while maintaining compositional structure. As of 2024, his group at MIT continues to investigate these topics, with a focus on making [deep learning](https://www.wikiprompt.org/wiki/deep-learning) systems more transparent and reliable. ## See Also - [Joshua Tenenbaum](https://www.wikiprompt.org/wiki/joshua-tenenbaum) - cognitive scientist known for work on compositionality and intuitive physics - [Brendan Lake](https://www.wikiprompt.org/wiki/brendan-lake) - researcher in human-like machine learning and systematic generalization - [Stanford AI Lab](https://www.wikiprompt.org/wiki/stanford-ai-lab) - another institution with related research in language and reasoning --- Source: https://www.wikiprompt.org/wiki/jacob-andreas License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/) Last updated: 2026-09-12T22:24:34.490777+00:00