[cosc-grad-students-list] 2 Thesis Proposal Presentations - Abiodun Adedeji & Nyo Me Han

[Eulenfeld, Menda]

Presentation 1
Title: FINAGENT: A BENCHMARK FOR EVALUATING LARGE LANGUAGE MODELS IN FINANCIAL ANALYSIS TASKS
Presenter: Abiodun Adedeji
Supervisor: Dr. Wenlu Wang

Abstract
I present FinaGent, a benchmark for evaluating large language models (LLMs) on financial analysis tasks grounded in real U.S. Securities and Exchange Commission (SEC) filings. FinaGent comprises 41 expert-validated question-context-answer triples across five task categories: Risk Analysis, Comparative Analysis, Strategic Reasoning, Causal Reasoning, and Financial Interpretation, and four major companies: JPMorgan Chase, Apple, Pfizer, and Exxon Mobil. I evaluate four state-of-the-art LLMs using ROUGE, BERTScore, and LLM-as-Judge. Key findings are: (1) Task type, not model size, is the strongest predictor of LLM performance; (2) Risk Analysis is the most difficult task for all models; (3) LLM-as-Judge provides more reliable evaluation than lexical or semantic measures for financial reasoning. Additionally, human-assigned difficulty labels do not align with computational difficulty, suggesting future benchmarks must better reflect actual task challenges.


Presentation 2
Title: Reasoning Under Uncertainty: LLM-Based Coordination of Multi-Robot Systems with Intermittent Communication and Robot Failures
Presenter: Nyo Me Han
Supervisor: Dr. Bozhen Liu

Abstract
Multi-robot systems provide clear benefits for large-scale inspection, monitoring, and search-and-rescue tasks, but their effectiveness can be hindered by hardware issues, unreliable communication, and too much reliance on fixed coordination methods. Current coordination structures often struggle to balance scalability with the intelligence of individual robots and do not have ways to continuously assess trust in each robot. This makes it hard to spot performance drops and potential failures during missions. While large language models (LLMs) show promise for multi-robot planning, existing methods usually use static coordination patterns and do not effectively tackle the best way to coordinate when assessing bridge structures under communication limits and structural uncertainty. This thesis presents a hybrid agent framework that promotes proactive coordination strategy selection, ongoing trust monitoring, clear decision-making, and flexible replanning during missions. The framework includes four main components: (i) Proactive Mission Strategy Selection (PMSS), a pre-mission module that picks the best coordination strategy using a structured mission descriptor and rule-based scoring; (ii) a Trust Generalization and Personalization Mechanism (TGPM) that continuously calculates a Trust-Theoretic Score for each robot to detect reduced sensing, mobility, or communication capabilities; (iii) LLM-Augmented Justification (LAJ), which gives natural language explanations to support clarity and accountability; and (iv) LLM-Driven Adaptive Replanning (LDAR), which creates context-sensitive recovery plans in response to hardware, battery, and communication problems. The proposed model is used for autonomous assessments of bridge structures and is tested with CommDeg-Bridge, a new benchmark for trust-aware, agent-based multi-robot bridge inspections. This benchmark measures damage distribution, resilience to communication issues, and the accuracy of coordination strategy selection. Results show improved adaptability, resilience, and clarity compared to fixed-strategy methods. Although it is designed for bridge assessments, the framework's reasoning abilities allow it to be applied in other inspection and post-disaster assessment fields.

──────────
Join Zoom Meeting
https://tamucc.zoom.us/j/93413503585?pwd=x8zsNj2zcZkijfZWAYBNqEzPHYCboZ.1<https://nam12.safelinks.protection.outlook.com/?url=https%3A%2F%2Ftamucc.zoom.us%2Fj%2F93413503585%3Fpwd%3Dx8zsNj2zcZkijfZWAYBNqEzPHYCboZ.1&data=05%7C02%7Ccosc-grad-students-list%40listserv.tamucc.edu%7C5c2f3e458b5548d4dc3308dea5fb0596%7C34cbfaf167a64781a9ca514eb2550b66%7C0%7C0%7C639130694722543344%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=egRI7TsMFAqDXtq9REWw8%2FXI2noGl%2Bh5Sq0HtMyqS5Q%3D&reserved=0>

Meeting ID: 934 1350 3585
Passcode: 522135

---

One tap mobile
+13462487799,,93413503585#,,,,*522135# US (Houston)
+12532050468,,93413503585#,,,,*522135# US

---

Join by SIP
* 93413503585 at zoomcrc.com<mailto:93413503585 at zoomcrc.com>
Passcode: 522135

Join instructions
https://tamucc.zoom.us/meetings/93413503585/invitations?signature=XDAymK_t8Qh6aikaFfZ0I5Ihem26He37k4-JaqcQ-Nc<https://nam12.safelinks.protection.outlook.com/?url=https%3A%2F%2Ftamucc.zoom.us%2Fmeetings%2F93413503585%2Finvitations%3Fsignature%3DXDAymK_t8Qh6aikaFfZ0I5Ihem26He37k4-JaqcQ-Nc&data=05%7C02%7Ccosc-grad-students-list%40listserv.tamucc.edu%7C5c2f3e458b5548d4dc3308dea5fb0596%7C34cbfaf167a64781a9ca514eb2550b66%7C0%7C0%7C639130694722567481%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=aAYKPxUvcL9BsGHeW1QDuWvIRYJC3PQM7ZeDcV1GlN8%3D&reserved=0>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listserv.tamucc.edu/pipermail/cosc-grad-students-list/attachments/20260429/1c03317e/attachment-0001.htm>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: text/calendar
Size: 10631 bytes
Desc: not available
URL: <http://listserv.tamucc.edu/pipermail/cosc-grad-students-list/attachments/20260429/1c03317e/attachment-0001.ics>