Large language models (LLMs) are becoming increasingly useful for programming and robotics tasks, but for more complicated reasoning problems, the gap between these systems and humans looms large. Without the ability to learn new concepts like humans do, these systems fail to form good abstractions — essentially, high-level representations of complex concepts that skip less-important details — and thus sputter when asked to do more sophisticated tasks.

Luckily, MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) researchers have found a treasure trove of abstractions within natural language. In three papers to be presented at the International Conference on Learning Representations this month, the group shows how our everyday words are a rich source of context for language models, helping them build better overarching representations for code synthesis, AI planning, and robotic navigation and manipulation.

The three separate frameworks build libraries of abstractions for their given task: LILO (library induction from language observations) can synthesize, compress, and document code; Ada (action domain acquisition) explores sequential decision-making for artificial intelligence agents; and LGA (language-guided abstraction) helps robots better understand their environments to develop more feasible plans. Each system is a neurosymbolic method, a type of AI that blends human-like neural networks and program-like logical components.

LILO: A neurosymbolic framework that codes

Large language models can be used to quickly write solutions to small-scale coding tasks, but cannot yet architect entire software libraries like the ones written by human software engineers. To take their software development capabilities further, AI models need to refactor (cut down and combine) code into libraries of succinct, readable, and reusable programs.

Refactoring tools like the previously developed MIT-led Stitch algorithm can automatically identify abstractions, so, in a nod to the Disney movie “Lilo & Stitch,” CSAIL researchers combined these algorithmic refactoring approaches with LLMs. Their neurosymbolic method LILO uses a standard LLM to write code, then pairs it with Stitch to find abstractions that are comprehensively documented in a library.

LILO’s unique emphasis on natural language allows the system to do tasks that require human-like commonsense knowledge, such as identifying and removing all vowels from a string of code and drawing a snowflake. In both cases, the CSAIL system outperformed standalone LLMs, as well as a previous library learning algorithm from MIT called DreamCoder, indicating its ability to build a deeper understanding of the words within prompts. These encouraging results point to how LILO could assist with things like writing programs to manipulate documents like Excel spreadsheets, helping AI answer questions about visuals, and drawing 2D graphics.

“Language models prefer to work with functions that are named in natural language,” says Gabe Grand SM ’23, an MIT PhD student in electrical engineering and computer science, CSAIL affiliate, and lead author on the research. “Our work creates more straightforward abstractions for language models and assigns natural language names and documentation to each one, leading to more interpretable code for programmers and improved system performance.”

When prompted on a programming task, LILO first uses an LLM to quickly propose solutions based on data it was trained on, and then the system slowly searches more exhaustively for outside solutions. Next, Stitch efficiently identifies common structures within the code and pulls out useful abstractions. These are then automatically named and documented by LILO, resulting in simplified programs that can be used by the system to solve more complex tasks.

The MIT framework writes programs in domain-specific programming languages, like Logo, a language developed at MIT in the 1970s to teach children about programming. Scaling up automated refactoring algorithms to handle more general programming languages like Python will be a focus for future research. Still, their work represents a step forward for how language models can facilitate increasingly elaborate coding activities.

Ada: Natural language guides AI task planning

Just like in programming, AI models that automate multi-step tasks in households and command-based video games lack abstractions. Imagine you’re cooking breakfast and ask your roommate to bring a hot egg to the table — they’ll intuitively abstract their background knowledge about cooking in your kitchen into a sequence of actions. In contrast, an LLM trained on similar information will still struggle to reason about what they need to build a flexible plan.

Named after the famed mathematician Ada Lovelace, who many consider the world’s first programmer, the CSAIL-led “Ada” framework makes headway on this issue by developing libraries of useful plans for virtual kitchen chores and gaming. The method trains on potential tasks and their natural language descriptions, then a language model proposes action abstractions from this dataset. A human operator scores and filters the best plans into a library, so that the best possible actions can be implemented into hierarchical plans for different tasks.

“Traditionally, large language models have struggled with more complex tasks because of problems like reasoning about abstractions,” says Ada lead researcher Lio Wong, an MIT graduate student in brain and cognitive sciences, CSAIL affiliate, and LILO coauthor. “But we can combine the tools that software engineers and roboticists use with LLMs to solve hard problems, such as decision-making in virtual environments.”

When the researchers incorporated the widely-used large language model GPT-4 into Ada, the system completed more tasks in a kitchen simulator and Mini Minecraft than the AI decision-making baseline “Code as Policies.” Ada used the background information hidden within natural language to understand how to place chilled wine in a cabinet and craft a bed. The results indicated a staggering 59 and 89 percent task accuracy improvement, respectively.

With this success, the researchers hope to generalize their work to real-world homes, with the hopes that Ada could assist with other household tasks and aid multiple robots in a kitchen. For now, its key limitation is that it uses a generic LLM, so the CSAIL team wants to apply a more powerful, fine-tuned language model that could assist with more extensive planning. Wong and her colleagues are also considering combining Ada with a robotic manipulation framework fresh out of CSAIL: LGA (language-guided abstraction).

Language-guided abstraction: Representations for robotic tasks

Andi Peng SM ’23, an MIT graduate student in electrical engineering and computer science and CSAIL affiliate, and her coauthors designed a method to help machines interpret their surroundings more like humans, cutting out unnecessary details in a complex environment like a factory or kitchen. Just like LILO and Ada, LGA has a novel focus on how natural language leads us to those better abstractions.

In these more unstructured environments, a robot will need some common sense about what it’s tasked with, even with basic training beforehand. Ask a robot to hand you a bowl, for instance, and the machine will need a general understanding of which features are important within its surroundings. From there, it can reason about how to give you the item you want. 

In LGA’s case, humans first provide a pre-trained language model with a general task description using natural language, like “bring me my hat.” Then, the model translates this information into abstractions about the essential elements needed to perform this task. Finally, an imitation policy trained on a few demonstrations can implement these abstractions to guide a robot to grab the desired item.

Previous work required a person to take extensive notes on different manipulation tasks to pre-train a robot, which can be expensive. Remarkably, LGA guides language models to produce abstractions similar to those of a human annotator, but in less time. To illustrate this, LGA developed robotic policies to help Boston Dynamics’ Spot quadruped pick up fruits and throw drinks in a recycling bin. These experiments show how the MIT-developed method can scan the world and develop effective plans in unstructured environments, potentially guiding autonomous vehicles on the road and robots working in factories and kitchens.

“In robotics, a truth we often disregard is how much we need to refine our data to make a robot useful in the real world,” says Peng. “Beyond simply memorizing what’s in an image for training robots to perform tasks, we wanted to leverage computer vision and captioning models in conjunction with language. By producing text captions from what a robot sees, we show that language models can essentially build important world knowledge for a robot.”

The challenge for LGA is that some behaviors can’t be explained in language, making certain tasks underspecified. To expand how they represent features in an environment, Peng and her colleagues are considering incorporating multimodal visualization interfaces into their work. In the meantime, LGA provides a way for robots to gain a better feel for their surroundings when giving humans a helping hand. 

An “exciting frontier” in AI

“Library learning represents one of the most exciting frontiers in artificial intelligence, offering a path towards discovering and reasoning over compositional abstractions,” says assistant professor at the University of Wisconsin-Madison Robert Hawkins, who was not involved with the papers. Hawkins notes that previous techniques exploring this subject have been “too computationally expensive to use at scale” and have an issue with the lambdas, or keywords used to describe new functions in many languages, that they generate. “They tend to produce opaque ‘lambda salads,’ big piles of hard-to-interpret functions. These recent papers demonstrate a compelling way forward by placing large language models in an interactive loop with symbolic search, compression, and planning algorithms. This work enables the rapid acquisition of more interpretable and adaptive libraries for the task at hand.”

By building libraries of high-quality code abstractions using natural language, the three neurosymbolic methods make it easier for language models to tackle more elaborate problems and environments in the future. This deeper understanding of the precise keywords within a prompt presents a path forward in developing more human-like AI models.

MIT CSAIL members are senior authors for each paper: Joshua Tenenbaum, a professor of brain and cognitive sciences, for both LILO and Ada; Julie Shah, head of the Department of Aeronautics and Astronautics, for LGA; and Jacob Andreas, associate professor of electrical engineering and computer science, for all three. The additional MIT authors are all PhD students: Maddy Bowers and Theo X. Olausson for LILO, Jiayuan Mao and Pratyusha Sharma for Ada, and Belinda Z. Li for LGA. Muxin Liu of Harvey Mudd College was a coauthor on LILO; Zachary Siegel of Princeton University, Jaihai Feng of the University of California at Berkeley, and Noa Korneev of Microsoft were coauthors on Ada; and Ilia Sucholutsky, Theodore R. Sumers, and Thomas L. Griffiths of Princeton were coauthors on LGA. 

LILO and Ada were supported, in part, by ​​MIT Quest for Intelligence, the MIT-IBM Watson AI Lab, Intel, U.S. Air Force Office of Scientific Research, the U.S. Defense Advanced Research Projects Agency, and the U.S. Office of Naval Research, with the latter project also receiving funding from the Center for Brains, Minds and Machines. LGA received funding from the U.S. National Science Foundation, Open Philanthropy, the Natural Sciences and Engineering Research Council of Canada, and the U.S. Department of Defense.

In order for natural language to be an effective form of communication, the parties involved need to be able to understand words and their context, assume that the content is largely shared in good faith and is trustworthy, reason about the information being shared, and then apply it to real-world scenarios. MIT PhD students interning with the MIT-IBM Watson AI Lab — Athul Paul Jacob SM ’22, Maohao Shen SM ’23, Victor Butoi, and Andi Peng SM ’23 — are working to attack each step of this process that’s baked into natural language models, so that the AI systems can be more dependable and accurate for users.

To achieve this, Jacob’s research strikes at the heart of existing natural language models to improve the output, using game theory. His interests, he says, are two-fold: “One is understanding how humans behave, using the lens of multi-agent systems and language understanding, and the second thing is, ‘How do you use that as an insight to build better AI systems?’” His work stems from the board game “Diplomacy,” where his research team developed a system that could learn and predict human behaviors and negotiate strategically to achieve a desired, optimal outcome.

“This was a game where you need to build trust; you need to communicate using language. You need to also play against six other players at the same time, which were very different from all the kinds of task domains people were tackling in the past,” says Jacob, referring to other games like poker and GO that researchers put to neural networks. “In doing so, there were a lot of research challenges. One was, ‘How do you model humans? How do you know whether when humans tend to act irrationally?’” Jacob and his research mentors — including Associate Professor Jacob Andreas and Assistant Professor Gabriele Farina of the MIT Department of Electrical Engineering and Computer Science (EECS), and the MIT-IBM Watson AI Lab’s Yikang Shen — recast the problem of language generation as a two-player game.

Using “generator” and “discriminator” models, Jacob’s team developed a natural language system to produce answers to questions and then observe the answers and determine if they are correct. If they are, the AI system receives a point; if not, no point is rewarded. Language models notoriously tend to hallucinate, making them less trustworthy; this no-regret learning algorithm collaboratively takes a natural language model and encourages the system’s answers to be more truthful and reliable, while keeping the solutions close to the pre-trained language model’s priors. Jacob says that using this technique in conjunction with a smaller language model could, likely, make it competitive with the same performance of a model many times bigger.  

Once a language model generates a result, researchers ideally want its confidence in its generation to align with its accuracy, but this frequently isn’t the case. Hallucinations can occur with the model reporting high confidence when it should be low. Maohao Shen and his group, with mentors Gregory Wornell, Sumitomo Professor of Engineering in EECS, and lab researchers with IBM Research Subhro Das, Prasanna Sattigeri, and Soumya Ghosh — are looking to fix this through uncertainty quantification (UQ). “Our project aims to calibrate language models when they are poorly calibrated,” says Shen. Specifically, they’re looking at the classification problem. For this, Shen allows a language model to generate free text, which is then converted into a multiple-choice classification task. For instance, they might ask the model to solve a math problem and then ask it if the answer it generated is correct as “yes, no, or maybe.” This helps to determine if the model is over- or under-confident.

Automating this, the team developed a technique that helps tune the confidence output by a pre-trained language model. The researchers trained an auxiliary model using the ground-truth information in order for their system to be able to correct the language model. “If your model is over-confident in its prediction, we are able to detect it and make it less confident, and vice versa,” explains Shen. The team evaluated their technique on multiple popular benchmark datasets to show how well it generalizes to unseen tasks to realign the accuracy and confidence of language model predictions. “After training, you can just plug in and apply this technique to new tasks without any other supervision,” says Shen. “The only thing you need is the data for that new task.”

Victor Butoi also enhances model capability, but instead, his lab team — which includes John Guttag, the Dugald C. Jackson Professor of Computer Science and Electrical Engineering in EECS; lab researchers Leonid Karlinsky and Rogerio Feris of IBM Research; and lab affiliates Hilde Kühne of the University of Bonn and Wei Lin of Graz University of Technology — is creating techniques to allow vision-language models to reason about what they’re seeing, and is designing prompts to unlock new learning abilities and understand key phrases.

Compositional reasoning is just another aspect of the decision-making process that we ask machine-learning models to perform in order for them to be helpful in real-world situations, explains Butoi. “You need to be able to think about problems compositionally and solve subtasks,” says Butoi, “like, if you’re saying the chair is to the left of the person, you need to recognize both the chair and the person. You need to understand directions.” And then once the model understands “left,” the research team wants the model to be able to answer other questions involving “left.”

Surprisingly, vision-language models do not reason well about composition, Butoi explains, but they can be helped to, using a model that can “lead the witness”, if you will. The team developed a model that was tweaked using a technique called low-rank adaptation of large language models (LoRA) and trained on an annotated dataset called Visual Genome, which has objects in an image and arrows denoting relationships, like directions. In this case, the trained LoRA model would be guided to say something about “left” relationships, and this caption output would then be used to provide context and prompt the vision-language model, making it a “significantly easier task,” says Butoi.

In the world of robotics, AI systems also engage with their surroundings using computer vision and language. The settings may range from warehouses to the home. Andi Peng and mentors MIT’s H.N. Slater Professor in Aeronautics and Astronautics Julie Shah and Chuang Gan, of the lab and the University of Massachusetts at Amherst, are focusing on assisting people with physical constraints, using virtual worlds. For this, Peng’s group is developing two embodied AI models — a “human” that needs support and a helper agent — in a simulated environment called ThreeDWorld. Focusing on human/robot interactions, the team leverages semantic priors captured by large language models to aid the helper AI to infer what abilities the “human” agent might not be able to do and the motivation behind actions of the “human,” using natural language. The team’s looking to strengthen the helper’s sequential decision-making, bidirectional communication, ability to understand the physical scene, and how best to contribute.

“A lot of people think that AI programs should be autonomous, but I think that an important part of the process is that we build robots and systems for humans, and we want to convey human knowledge,” says Peng. “We don’t want a system to do something in a weird way; we want them to do it in a human way that we can understand.”