Quinn Arnold
Applied Mathematics & Statistics · Bryant University
I work on reinforcement learning, mechanistic interpretability, LLM safety, and trustworthy machine-learning systems. Current work includes deep reinforcement learning for combinatorial auctions and research on mechanistic understanding of language models.
Selected
Current Research · Deep Reinforcement Learning
FastCombo treats auction price discovery as a sequential decision problem. A recurrent bidder–item graph policy uses public price and demand histories to produce new item prices, moving price-selection computation into an offline-trained deep reinforcement learning policy. A preprint is forthcoming.
Current Research · Mechanistic Interpretability
A conceptual and empirical project asking what it means to understand a learned model with respect to a particular behavior. It develops a structured evidence framework and tests bounded mechanistic claims using causal interventions, held-out predictions, generalization, and coverage.
LLM Safety · Bryant University · 2026
Built a security-hardened campus assistant around a QLoRA fine-tune of Qwen3-32B, hybrid retrieval, and layered prompt-injection defenses. The evaluation covered 15 attack vectors; the defense pipeline eliminated successful attacks in that test set, while a RoBERTa classifier achieved 95.3% injection recall.
Open Source · 2025–2026
Designed and built an open-source iOS and backend research system for comparing shortest and risk-weighted walking routes. The project combines SwiftUI, secure services, graph routing, time-binned KDE surfaces, and reproducible multicity evaluations with strict temporal holdout. Those evaluations compared KDE, XGBoost, and graph neural networks using corridor-level holdout incidents; KDE produced the strongest downstream routing outcomes in every evaluated city.
Open Source · Current
I contribute to public projects across reinforcement learning, model fairness, geospatial machine learning, agentic AI, and LLM systems.
Industry · Summer 2026
Worked on ML model governance through bias assessment, mitigation analysis, reliability review, and production-readiness evaluation. Also developed LLM-assisted image-analysis workflows that compared visual hazard descriptions with inspection reports for underwriting review.
Industry · Summer 2025
Built forecasting models for inventory and revenue planning, an image-classification pipeline for hierarchical product taxonomies, and a semantic-search system that translated natural-language questions into SQL over product and customer data.