The cathode of the Large Plasma Device (LAPD) is operating outside of its expected temperature range of 1600 to 1800 C. As the cathode thermionically emits electrons to ionize argon gas into plasma, anomalous temperatures lead to unclear data and less uniform, unstable plasma columns. The purpose of this project is to diagnose the cathode’s temperature using spectroscopy. We constructed an automated grating spectrometer to collect a continuous spectrum of the cathode’s light emissions up to wavelengths of 2000 nm. To automate the spectrum collection, we connected a stepping motor, oscilloscope, and wave detector to the spectrometer. Treating the cathode as a blackbody emitter, we diagnosed its temperature by fitting its emission spectrum on a Planckian radiation curve. Diagnosing the cathode’s true temperature ensures data integrity for future experiments, including the Alfvén wave experiments and space physics replication studies the LAPD’s Basic Plasma Science Facility is widely regarded for.

As a whole, translation services between English and Mayan languages is underserved across the country, despite large numbers of Mayan speaking residents throughout the United States, from Los Angeles to Urbana Champaign, and Mayan languages frequently being some of the most spoken languages in immigration court. In prior work, we compiled the largest known English–Q'anjob'al (a Mayan language) dataset and showed that fine-tuning open-source language models can produce initial translation capabilities. However, models still do not understand the language well, and data scarcity remains a problem. Thus, in this work, we build on those findings by asking whether data from related Mayan languages can be reused to boost translation performance across the language family when data for any single language is scarce. Particularly, we tested structured task learning, starting from "easier" to "harder" tasks to improve performance while using the same dataset, as well as strategies that expose models to broader Mayan language data before specializing on individual target languages. We evaluate across a few Mayan languages, showing that training must be carefully structured in order to not overfit or confuse models.

Accurate depth perception and intuitive control are imperative for the safety of robotic surgery. The Teleoperated Enhanced Robotic Surgical Assistant (TERESA) project addresses these needs by integrating immersive virtual reality to provide surgeons with natural 3D vision and head-synchronized camera movement. This work presents a simulation framework developed using the MuJoCo physics engine to model and analyze the system’s behavior. This simulation allows for the evaluation of motion and control strategies while isolating them from the noise and mechanical instabilities inherent in hardware prototypes. The development process utilizes a simplified sliding joint model to establish a direct relationship between input commands and resulting motion. This phase is critical for validating how control signals—originating from user inputs via the headset interface—translate into physical actuation for the system's motors. This framework is subsequently extended into a 3D environment to model the spatial movement of the robotic arm, ensuring an accurate representation of the TERESA platform’s kinematics. Results indicate that the simulation provides a robust platform for verifying control loops and optimizing latency. This infrastructure facilitates systematic testing and refinement of control strategies prior to physical deployment. By bridging the gap between virtual commands and physical execution, this project enhances the reliability and precision of teleoperated surgical systems.

Large language model agents increasingly rely on external APIs to accomplish real-world tasks, yet identifying which tools to use remains a critical bottleneck. A fundamental semantic gap separates how users describe tasks from how tools are documented, and as API ecosystems scale to tens of thousands of endpoints, static one-shot retrieval cannot bridge it. We introduce FitText, a training-free dynamic tool retrieval framework that embeds retrieval directly into the agent's reasoning loop. Rather than querying the tool database with the raw user query, FitText prompts the LLM to generate pseudo-tool descriptions — natural-language hypotheses about needed functionality — that serve as semantically aligned retrieval probes. These probes are iteratively refined through retrieval feedback, diversified via stochastic generation, and evolved through a multi-generation memetic algorithm combining population-based exploration, crossover, mutation, and a tool memory that steers search away from previously explored regions. We evaluate FitText on ToolRet (43k tools) and StableToolBench (16,464 APIs). On ToolRet, average retrieval rank improves from 6.88 to 2.88. On StableToolBench, our Memetic strategy achieves a 0.73 average pass rate — a 24-point absolute gain over static retrieval — with the largest improvements on ambiguous multi-tool tasks where deterministic methods fail. These results establish adaptive, in-loop retrieval as a general-purpose interface for evolving tool ecosystems.