Read: Compare action choices, possible outcomes, and expected values in a Markov decision tree grounded in the agent’s Actor, World Model, and Critic.

Revise: Edit action probabilities, outcome likelihoods, or values. CoChoice translates these corrections into supervised training targets for the corresponding model components.

Reinforce: Continue reinforcement learning while retaining user edits as training constraints, balancing task performance with human alignment. Users can inspect the updated explanation and repeat the cycle.

In a study with 39 participants collaborating with an agent in Overcooked, CoChoice improved user understanding and alignment compared with read-only and manually editable explanations. Stakeholder interviews explored opportunities in transportation, medicine, and cognitive modeling. The work introduces Editable XRL as a paradigm for bidirectional human–agent alignment.

Branch and merge: Break open-ended problems into parallel explorations, then consolidate complementary ideas. A directed acyclic graph makes conversation turns and inherited context explicit.

Maintain an overview: AI-generated context briefs summarize evolving lines of inquiry, helping users keep track of the larger task while exploring individual branches.

Refactor the structure: Preview, accept, or reject bounded AI suggestions for reorganizing the graph and clarifying relationships between ideas.

A user study found Trellis helpful for exploring alternatives, managing divergent hypotheses, integrating ideas across branches, and maintaining global task awareness. The project explores how interpretable, user-controlled context structures can support iterative sensemaking with LLMs.

Tree-based initialization: A sparse neural backbone follows the source tree’s depth and parent–child structure. Soft split conditions and partial decision paths transfer its rules, while leaf predictions initialize the output.

Calibrated split softness: Feature margins, sample support, and impurity reduction determine how decisively each split behaves at initialization, giving less reliable rules greater flexibility.

Path-dependent adaptation: Route-State passes context along the tree’s edges and learns adjustments to downstream decisions. Its routing adjustment starts at zero, preserving the tree-derived initial predictor before training.

Evaluation spans eight classification and regression datasets in single-tree and random-forest settings. TreePINN performs comparably to or better than competing tree-informed methods on most evaluated tasks, with additional analyses of limited-data learning, source-tree fidelity, and distilled-tree complexity.

While Explainable AI (XAI) helps users understand AI decisions, misalignment in domain knowledge can lead to disagreement. This inconsistency hinders understanding, and because explanations are often read-only, users lack the control to improve alignment. We propose making XAI editable, allowing users to write rules to improve control and gain deeper understanding through the generation effect of active learning. We developed CoExplain, leveraging a neural network for universal representation and symbolic rules for intuitive reasoning on interpretable attributes. CoExplain explains the neural network with a faithful proxy decision tree, parses user-written rules as an equivalent neural network graph, and collaboratively optimizes the decision tree. In a user study (N=43), CoExplain and manually editable XAI improved user understanding and model alignment compared to read-only XAI. CoExplain was easier to use with fewer edits and less time. This work contributes Editable XAI for bidirectional AI alignment, improving understanding and control.

Traditional vehicle reposition techniques used in ride-hailing contexts turn out to have little effect due to their ignorance on drivers' preference and response behaviors. This project first demonstrates that drivers do have personalized cruising preferences and carried out experiments to prove it necessary to consider drivers' preference when repositioning. Our on-field User Study of 106 professional drivers further stressed that drivers do have preference and their preference is a key factor in their decision making process of whether to accept a reposition or not. Made up with three key modules: an LSTM predictor of drivers' preferences, a decision model on drivers and a dual-agent DRL framework, our solution can both satisfy driver preferences and demand-supply gap.

Current medical decision support systems (MDSS) provide fixed guidelines to doctors, while each doctor has an individual approach to diagnosis, e.g. thresholds of indexes, sequence of examinations. Existing medical prediction tools also offer poor interpretability which is confusing to doctors. This project focuses on enabling doctors to build diagnostic models with their individual preference. With our toolkit, doctors can use visual programming to customize a prediction model that is highly interpretable and precise in practice.

Virtual Reality is trending, while there are devices like Xbox Adaptive Controller to provide accessibility for the people in need, they are still hard to use in VR context because people cannot see through a VR Headset. When people have to click a button, they can hardly find it without seeing it. This project detects ear motion and uses it as an input to the VR devices as a replacement of traditional controllers to provide accessibility to the people with special needs. Our user study on 15 volunteers showed that this method is effective and easy to use. This device gives everyone the access to control a VR device, even if they are not good at moving their ears.