Adoption of recent tools and technologies occurs when users largely perceive them as reliable, accessible, and an improvement over the available methods and workflows for the price. Five PhD students from the inaugural class of the MIT-IBM Watson AI Lab Summer Program are utilizing state-of-the-art resources, alleviating AI pain points, and creating recent features and capabilities to advertise AI usefulness and deployment — from learning when to trust a model that predicts one other’s accuracy to more effectively reasoning over knowledge bases. Together, the efforts from the scholars and their mentors form a through-line, where practical and technically rigorous research results in more dependable and useful models across domains.
Constructing probes, routers, recent attention mechanisms, synthetic datasets, and program-synthesis pipelines, the scholars’ work spans safety, inference efficiency, multimodal data, and knowledge-grounded reasoning. Their techniques emphasize scaling and integration, with impact all the time in sight.
Learning to trust, and when
MIT math graduate student Andrey Bryutkin’s research prioritizes the trustworthiness of models. He seeks out internal structures inside problems, comparable to equations governing a system and conservation laws, to know leverage them to supply more dependable and robust solutions. Armed with this and dealing with the lab, Bryutkin developed a technique to see into the character of enormous learning models (LLMs) behaviors. Along with the lab’s Veronika Thost of IBM Research and Marzyeh Ghassemi — associate professor and the Germeshausen Profession Development Professor within the MIT Department of Electrical Engineering and Computer Science (EECS) and a member of the Institute of Medical Engineering Sciences and the Laboratory for Information and Decision Systems — Bryutkin explored the “uncertainty of uncertainty” of LLMs.
Classically, tiny feed-forward neural networks two-to-three layers deep, called probes, are trained alongside LLMs and employed to flag untrustworthy answers from the larger model to developers; nevertheless, these classifiers also can produce false negatives and only provide point estimates, which don’t offer much details about when the LLM is failing. Investigating protected/unsafe prompts and question-answer tasks, the MIT-IBM team used prompt-label pairs, in addition to the hidden states like activation vectors and last tokens from an LLM, to measure gradient scores, sensitivity to prompts, and out-of-distribution data to find out how reliable the probe was and learn areas of knowledge which can be difficult to predict. Their method also helps discover potential labeling noise. It is a critical function, because the trustworthiness of AI systems depends entirely on the standard and accuracy of the labeled data they’re built upon. More accurate and consistent probes are especially vital for domains with critical data in applications like IBM’s Granite Guardian family of models.
One other solution to ensure trustworthy responses to queries from an LLM is to enhance them with external, trusted knowledge bases to eliminate hallucinations. For structured data, comparable to social media connections, financial transactions, or corporate databases, knowledge graphs (KG) are natural suits; nevertheless, communications between the LLM and KGs often use fixed, multi-agent pipelines which can be computationally inefficient and expensive. Addressing this, physics graduate student Jinyeop Song, together with lab researchers Yada Zhu of IBM Research and EECS Associate Professor Julian Shun created a single-agent, multi-turn, reinforcement learning framework that streamlines this process. Here, the group designed an API server hosting Freebase and Wikidata KGs, which consist of general web-based knowledge data, and a LLM agent that issues targeted retrieval actions to fetch pertinent information from the server. Then, through continuous back-and-forth, the agent appends the gathered data from the KGs to the context and responds to the query. Crucially, the system uses reinforcement learning to coach itself to deliver answers that strike a balance between accuracy and completeness. The framework pairs an API server with a single reinforcement learning agent to orchestrate data-grounded reasoning with improved accuracy, transparency, efficiency, and transferability.
Spending computation properly
The timeliness and completeness of a model’s response carry similar weight to the importance of its accuracy. This is very true for handling long input texts and people where elements, just like the subject of a story, evolve over time, so EECS graduate student Songlin Yang is re-engineering what models can handle at each step of inference. Specializing in transformer limitations, like those in LLMs, the lab’s Rameswar Panda of IBM Research and Yoon Kim, the NBX Professor and associate professor in EECS, joined Yang to develop next-generation language model architectures beyond transformers.
Transformers face two key limitations: high computational complexity in long-sequence modeling on account of the softmax attention mechanism, and limited expressivity resulting from the weak inductive bias of RoPE (rotary positional encoding). Which means because the input length doubles, the computational cost quadruples. RoPE allows transformers to know the sequence order of tokens (i.e., words); nevertheless, it doesn’t do a great job capturing internal state changes over time, like variable values, and is proscribed to the sequence lengths seen during training.
To deal with this, the MIT-IBM team explored theoretically grounded yet hardware-efficient algorithms. As a substitute for softmax attention, they adopted linear attention, reducing the quadratic complexity that limits the feasible sequence length. Additionally they investigated hybrid architectures that mix softmax and linear attention to strike a greater balance between computational efficiency and performance.
Increasing expressivity, they replaced RoPE with a dynamic reflective positional encoding based on the Householder transform. This approach enables richer positional interactions for deeper understanding of sequential information, while maintaining fast and efficient computation. The MIT-IBM team’s advancement reduces the necessity for transformers to interrupt problems into many steps, as an alternative enabling them to handle more complex subproblems with fewer inference tokens.
Visions anew
Visual data contain multitudes that the human brain can quickly parse, internalize, after which imitate. Using vision-language models (VLMs), two graduate students are exploring ways to do that through code.
Over the past two summers and under the advisement of Aude Oliva, MIT director of the MIT-IBM Watson AI Lab and a senior research scientist within the Computer Science and Artificial Intelligence Laboratory; and IBM Research’s Rogerio Feris, Dan Gutfreund, and Leonid Karlinsky (now at Xero), Jovana Kondic of EECS has explored visual document understanding, specifically charts. These contain elements, comparable to data points, legends, and axes labels, that require optical character recognition and numerical reasoning, which models still struggle with. With a view to facilitate the performance on tasks comparable to these, Kondic’s group got down to create a big, open-source, synthetic chart dataset from code that may very well be used for training and benchmarking.
With their prototype, ChartGen, the researchers created a pipeline that passes seed chart images through a VLM, which is prompted to read the chart and generate a Python script that was likely used to create the chart in the primary place. The LLM component of the framework then iteratively augments the code from many charts to ultimately produce over 200,000 unique pairs of charts and their codes, spanning nearly 30 chart types, in addition to supporting data and annotation like descriptions and question-answer pairs in regards to the charts. The team is further expanding their dataset, helping to enable critical multimodal understanding to data visualizations for enterprise applications like financial and scientific reports, blogs, and more.
As an alternative of charts, EECS graduate student Leonardo Hernandez Cano has his eyes on digital design, specifically visual texture generation for CAD applications and the goal of discovering efficient ways to enable to capabilities in VLMs. Teaming up with the lab groups led by Armando Solar-Lezama, EECS professor and Distinguished Professor of Computing within the MIT Schwarzman College of Computing, and IBM Research’s Nathan Fulton, Hernandez Cano created a program synthesis system that learns to refine code by itself. The system starts with a texture description given by a user in the shape of a picture. It then generates an initial Python program, which produces visual textures, and iteratively refines the code with the goal of finding a program that produces a texture that matches the goal description, learning to look for brand spanking new programs from the info that the system itself produces. Through these refinements, the novel program can create visualizations with the specified luminosity, color, iridescence, etc., mimicking real materials.
When viewed together, these projects, and the people behind them, are making a cohesive push toward more robust and practical artificial intelligence. By tackling the core challenges of reliability, efficiency, and multimodal reasoning, the work paves the best way for AI systems that aren’t only more powerful, but in addition more dependable and cost-effective, for real-world enterprise and scientific applications.
