Leilani H. Gilpin

Leilani H. Gilpin, Andrew R. Paley, Mohammed A. Alam et al. · mit

There is broad agreement that Artificial Intelligence (AI) systems, particularly those using Machine Learning (ML), should be able to "explain" their behavior. Unfortunately, there is little agreement as to what constitutes an "explanation." This has caused a disconnect between the explanations that systems produce in service of explainable Artificial Intelligence (XAI) and those explanations that users and other audiences actually need, which should be defined by the full spectrum of functional roles, audiences, and capabilities for explanation. In this paper, we explore the features of explanations and how to use those features in evaluating their utility. We focus on the requirements for explanations defined by their functional role, the knowledge states of users who are trying to understand them, and the availability of the information needed to generate them. Further, we discuss the risk of XAI enabling trust in systems without establishing their trustworthiness and define a critical next step for the field of XAI to establish metrics to guide and ground the utility of system-generated explanations.

Li Liu, Jiaming Qu, Marc Jowell Bagaoisan et al. · mit

Most existing assistive navigation tools focus on providing real-time guidance for Blind and Low-Vision (BLV) people, but few support building a holistic spatial understanding of unfamiliar environments before travel. Such cognitive map construction (e.g., knowing that a fountain is south of a tower and west of a hotel) is important for pre-travel planning, yet remains underexplored in prior work. To address this gap, we present Touching Space, an end-to-end system that retrieves map data for a target place and loads it into a frontend interface for exploration. The system combines haptic and audio feedback: users explore spatial layouts through touch and ask spoken questions to a conversational agent during exploration. Touching Space contributes a conversational interface that supports BLV users in building cognitive maps on commodity hardware.

Shiyuan Huang, Li Liu, Jincheng He et al. · mit

When faced with complex spatial problems, humans naturally sketch layouts to organize their thinking, and the act of drawing further sharpens their understanding. In this work, we ask whether a similar principle holds for Large Language Models (LLMs): can learning to construct explicit visual layouts from spatial descriptions instill genuine spatial understanding? We introduce Text2Space, a dataset that pairs natural language descriptions with ground-truth ASCII grid layouts and spatial QA pairs, enabling us to separate failures in constructing spatial representations from failures in reasoning over them. We adopt ASCII because it is human-readable, operates entirely within the token space of language models, and encodes spatial relations in a structurally verifiable form. Our evaluation reveals a pronounced "Read-Write Asymmetry": LLMs interpret ASCII representations effectively but struggle to produce them from text, and these construction errors propagate to incorrect answers downstream. To address this limitation, we train models on layout construction (Text$\rightarrow$ASCII) and find that it significantly improves spatial reasoning from text alone, even without producing any ASCII at inference time. Combining construction with comprehension training further amplifies these gains. Crucially, these improvements transfer to three external spatial reasoning benchmarks, demonstrating that, much as sketching sharpens human spatial thinking, learning to construct explicit layouts instills spatial understanding that generalizes beyond the training format.

Adam Amos-Binks, Dustin Dannenhauer, Leilani H. Gilpin · mit

Anticipatory thinking drives our ability to manage risk - identification and mitigation - in everyday life, from bringing an umbrella when it might rain to buying car insurance. As AI systems become part of everyday life, they too have begun to manage risk. Autonomous vehicles log millions of miles, StarCraft and Go agents have similar capabilities to humans, implicitly managing risks presented by their opponents. To further increase performance in these tasks, out-of-distribution evaluation can characterize a model's bias, what we view as a type of risk management. However, learning to identify and mitigate low-frequency, high-impact risks is at odds with the observational bias required to train machine learning models. StarCraft and Go are closed-world domains whose risks are known and mitigations well documented, ideal for learning through repetition. Adversarial filtering datasets provide difficult examples but are laborious to curate and static, both barriers to real-world risk management. Adversarial robustness focuses on model poisoning under the assumption there is an adversary with malicious intent, without considering naturally occurring adversarial examples. These methods are all important steps towards improving risk management but do so without considering open-worlds. We unify these open-world risk management challenges with two contributions. The first is our perception challenges, designed for agents with imperfect perceptions of their environment whose consequences have a high impact. Our second contribution are cognition challenges, designed for agents that must dynamically adjust their risk exposure as they identify new risks and learn new mitigations. Our goal with these challenges is to spur research into solutions that assess and improve the anticipatory thinking required by AI agents to manage risk in open-worlds and ultimately the real-world.

Shiyuan Huang, Siddarth Mamidanna, Shreedhar Jangam et al.

Large language models (LLMs) such as ChatGPT have demonstrated superior performance on a variety of natural language processing (NLP) tasks including sentiment analysis, mathematical reasoning and summarization. Furthermore, since these models are instruction-tuned on human conversations to produce "helpful" responses, they can and often will produce explanations along with the response, which we call self-explanations. For example, when analyzing the sentiment of a movie review, the model may output not only the positivity of the sentiment, but also an explanation (e.g., by listing the sentiment-laden words such as "fantastic" and "memorable" in the review). How good are these automatically generated self-explanations? In this paper, we investigate this question on the task of sentiment analysis and for feature attribution explanation, one of the most commonly studied settings in the interpretability literature (for pre-ChatGPT models). Specifically, we study different ways to elicit the self-explanations, evaluate their faithfulness on a set of evaluation metrics, and compare them to traditional explanation methods such as occlusion or LIME saliency maps. Through an extensive set of experiments, we find that ChatGPT's self-explanations perform on par with traditional ones, but are quite different from them according to various agreement metrics, meanwhile being much cheaper to produce (as they are generated along with the prediction). In addition, we identified several interesting characteristics of them, which prompt us to rethink many current model interpretability practices in the era of ChatGPT(-like) LLMs.

Biagio La Rosa, Leilani H. Gilpin, Roberto Capobianco

Compositional Explanations is a method for identifying logical formulas of concepts that approximate the neurons' behavior. However, these explanations are linked to the small spectrum of neuron activations (i.e., the highest ones) used to check the alignment, thus lacking completeness. In this paper, we propose a generalization, called Clustered Compositional Explanations, that combines Compositional Explanations with clustering and a novel search heuristic to approximate a broader spectrum of the neurons' behavior. We define and address the problems connected to the application of these methods to multiple ranges of activations, analyze the insights retrievable by using our algorithm, and propose desiderata qualities that can be used to study the explanations returned by different algorithms.

Sharan Subramanian, Leilani H. Gilpin

The application of Artificial Intelligence in the medical market brings up increasing concerns but aids in more timely diagnosis of silent progressing diseases like Diabetic Retinopathy. In order to diagnose Diabetic Retinopathy (DR), ophthalmologists use color fundus images, or pictures of the back of the retina, to identify small distinct features through a difficult and time-consuming process. Our work creates a novel CNN model and identifies the severity of DR through fundus image input. We classified 4 known DR features, including micro-aneurysms, cotton wools, exudates, and hemorrhages, through convolutional layers and were able to provide an accurate diagnostic without additional user input. The proposed model is more interpretable and robust to overfitting. We present initial results with a sensitivity of 97% and an accuracy of 71%. Our contribution is an interpretable model with similar accuracy to more complex models. With that, our model advances the field of DR detection and proves to be a key step towards AI-focused medical diagnosis.

Olivia Peiyu Wang, Leilani H. Gilpin

The growing adoption of large language models in legal practice brings both significant promise and serious risk. Legal professionals stand to benefit from AI that can reason over contracts, draft documents, and analyze sources at scale, yet the high-stakes nature of legal work demands a level of rigor that current AI systems do not provide. The central problem is not simply that LLMs hallucinate facts and references; it is that they systematically draw inferences that go beyond what the source text actually supports, presenting assumption-laden conclusions as if they were logically grounded. This proposal presents a neuro-symbolic approach to legal AI that combines the expressive power of large language models with the rigor of formal verification, aiming to make AI-assisted legal reasoning both capable and trustworthy, thus reducing the burden of manual verification without sacrificing the accountability that legal practice demands.

Jonathan Wellington Morris, Vishrut Shah, Alex Besanceney et al.

Sim-to real gap in Reinforcement Learning is when a model trained in a simulator does not translate to the real world. This is a problem for Autonomous Vehicles (AVs) as vehicle dynamics can vary from simulation to reality, and also from vehicle to vehicle. Slug Mobile is a one tenth scale autonomous vehicle created to help address the sim-to-real gap for AVs by acting as a test-bench to develop models that can easily scale from one vehicle to another. In addition to traditional sensors found in other one tenth scale AVs, we have also included a Dynamic Vision Sensor so we can train Spiking Neural Networks running on neuromorphic hardware.

Li Liu, Diji Yang, Sijia Zhong et al.

In question-answering scenarios, humans can assess whether the available information is sufficient and seek additional information if necessary, rather than providing a forced answer. In contrast, Vision Language Models (VLMs) typically generate direct, one-shot responses without evaluating the sufficiency of the information. To investigate this gap, we identify a critical and challenging task in the Visual Question Answering (VQA) scenario: can VLMs indicate how to adjust an image when the visual information is insufficient to answer a question? This capability is especially valuable for assisting visually impaired individuals who often need guidance to capture images correctly. To evaluate this capability of current VLMs, we introduce a human-labeled dataset as a benchmark for this task. Additionally, we present an automated framework that generates synthetic training data by simulating ``where to know'' scenarios. Our empirical results show significant performance improvements in mainstream VLMs when fine-tuned with this synthetic data. This study demonstrates the potential to narrow the gap between information assessment and acquisition in VLMs, bringing their performance closer to humans.

Biagio La Rosa, Leilani H. Gilpin

While neurons are the basic units of deep neural networks, it is still unclear what they learn and if their knowledge is aligned with that of humans. Compositional explanations aim to answer this question by describing the spatial alignment between neuron activations and concepts through logical rules. These logical descriptions are typically computed via a search over all possible concept combinations. Since computing the spatial alignment over the entire state space is computationally infeasible, the literature commonly adopts beam search to restrict the space. However, beam search cannot provide any theoretical guarantees of optimality, and it remains unclear how close current explanations are to the true optimum. In this theoretical paper, we address this gap by introducing the first framework for computing guaranteed optimal compositional explanations. Specifically, we propose: (i) a decomposition that identifies the factors influencing the spatial alignment, (ii) a heuristic to estimate the alignment at any stage of the search, and (iii) the first algorithm that can compute optimal compositional explanations within a feasible time. Using this framework, we analyze the differences between optimal and non-optimal explanations in the most popular settings for compositional explanations, the computer vision domain and Convolutional Neural Networks. In these settings, we demonstrate that 10-40 percent of explanations obtained with beam search are suboptimal when overlapping concepts are involved. Finally, we evaluate a beam-search variant guided by our proposed decomposition and heuristic, showing that it matches or improves runtime over prior methods while offering greater flexibility in hyperparameters and computational resources.

Biagio La Rosa, Leilani H. Gilpin

Neurons are the fundamental building blocks of deep neural networks, and their interconnections allow AI to achieve unprecedented results. Motivated by the goal of understanding how neurons encode information, compositional explanations leverage logical relationships between concepts to express the spatial alignment between neuron activations and human knowledge. However, these explanations rely on human-annotated datasets, restricting their applicability to specific domains and predefined concepts. This paper addresses this limitation by introducing a framework for the vision domain that allows users to probe neurons for arbitrary concepts and datasets. Specifically, the framework leverages masks generated by open vocabulary semantic segmentation to compute open vocabulary compositional explanations. The proposed framework consists of three steps: specifying arbitrary concepts, generating semantic segmentation masks using open vocabulary models, and deriving compositional explanations from these masks. The paper compares the proposed framework with previous methods for computing compositional explanations both in terms of quantitative metrics and human interpretability, analyzes the differences in explanations when shifting from human-annotated data to model-annotated data, and showcases the additional capabilities provided by the framework in terms of flexibility of the explanations with respect to the tasks and properties of interest.

Olivia Peiyu Wang, Tashvi Bansal, Ryan Bai et al.

Large Language Models (LLMs) suffer from critical reasoning gaps, including a tendency to hallucinate and poor accuracy in classifying logical fallacies. This limitation stems from their default System 1 processing, which is fast and intuitive, whereas reliable reasoning requires the deliberate, effortful System 2 approach (Kahneman, 2011; Li et al., 2025). Since full System 2 training is often prohibitively expensive, we explore a low-cost, instruction-based intervention to bridge this gap. Our methodology introduces a novel stepwise instruction dataset that decomposes fallacy classification into a series of atomic procedural steps (simple binary questions). We further augment this with a final verification step where models consult a relational knowledge graph of related fallacies. This procedural, rule-based intervention yields a significant improvement in LLM logical fallacy classification. Crucially, the approach also provides enhanced transparency into the LLMs' decision-making, highlighting a practical pathway for Neuro-symbolic architectures to address LLM reasoning deficits.

Kargi Chauhan, Leilani H. Gilpin

Modern diffusion models generate realistic traffic simulations but systematically violate physical constraints. In a large-scale evaluation of SceneDiffuser++, a state-of-the-art traffic simulator, we find that 50% of generated trajectories violate basic physical laws - vehicles collide, drive off roads, and spawn inside buildings. This reveals a fundamental limitation: current models treat physical validity as an emergent property rather than an architectural requirement. We propose Validity-First Spatial Intelligence (VFSI), which enforces constraints through energy-based guidance during diffusion sampling, without model retraining. By incorporating collision avoidance and kinematic constraints as energy functions, we guide the denoising process toward physically valid trajectories. Across 200 urban scenarios from the Waymo Open Motion Dataset, VFSI reduces collision rates by 67% (24.6% to 8.1%) and improves overall validity by 87% (50.3% to 94.2%), while simultaneously improving realism metrics (ADE: 1.34m to 1.21m). Our model-agnostic approach demonstrates that explicit constraint enforcement during inference is both necessary and sufficient for physically valid traffic simulation.

Leilani H. Gilpin, Cecilia Testart, Nathaniel Fruchter et al.

There is a disconnect between explanatory artificial intelligence (XAI) methods and the types of explanations that are useful for and demanded by society (policy makers, government officials, etc.) Questions that experts in artificial intelligence (AI) ask opaque systems provide inside explanations, focused on debugging, reliability, and validation. These are different from those that society will ask of these systems to build trust and confidence in their decisions. Although explanatory AI systems can answer many questions that experts desire, they often don't explain why they made decisions in a way that is precise (true to the model) and understandable to humans. These outside explanations can be used to build trust, comply with regulatory and policy changes, and act as external validation. In this paper, we focus on XAI methods for deep neural networks (DNNs) because of DNNs' use in decision-making and inherent opacity. We explore the types of questions that explanatory DNN systems can answer and discuss challenges in building explanatory systems that provide outside explanations for societal requirements and benefit.

Leilani H. Gilpin, David Bau, Ben Z. Yuan et al.

There has recently been a surge of work in explanatory artificial intelligence (XAI). This research area tackles the important problem that complex machines and algorithms often cannot provide insights into their behavior and thought processes. XAI allows users and parts of the internal system to be more transparent, providing explanations of their decisions in some level of detail. These explanations are important to ensure algorithmic fairness, identify potential bias/problems in the training data, and to ensure that the algorithms perform as expected. However, explanations produced by these systems is neither standardized nor systematically assessed. In an effort to create best practices and identify open challenges, we provide our definition of explainability and show how it can be used to classify existing literature. We discuss why current approaches to explanatory methods especially for deep neural networks are insufficient. Finally, based on our survey, we conclude with suggested future research directions for explanatory artificial intelligence.