obed junias

I’m a graduate researcher at the BLAST Lab, CU Boulder, advised by Dr. Maria L. Pacheco. My work is at the intersection of natural language reasoning, trustworthy LLMs, and AI safety, with a focus on building systems and benchmarks that make language models and agentic pipelines more transparent, logically grounded, and reliable. Most recently, I introduced LOGICAL-COMMONSENSEQA (ACL 2026), a benchmark that probes how well language models compose plausibility judgments using logical operators.

My current research investigates model collapse in foundation models, studying how recursive training on synthetic data degrades model quality, and tracing where this degradation originates through mechanistic analysis.

I am also fortunate to have collaborated with Dr. Theodora Chaspari on algorithmic bias in mental health classification, with work published at IEEE-BHI 2025.

research interests

• LLM Reliability & Trustworthiness: I study model collapse, investigating how recursive training on synthetic data degrades model quality over generations through mechanistic analysis and measurement frameworks.

• Model Evaluation & Benchmarking: I design benchmarks that probe logical and faithful reasoning in LLMs, including LOGICAL-COMMONSENSEQA (ACL 2026), which examines how well models compose plausibility judgments using logical operators.

• Responsible AI & Fairness: I study demographic disparities in LLM behavior in high-stakes domains, including bias in clinical NLP (IEEE-BHI 2025).

• Agentic Systems: I build multi-agent LLM pipelines for complex real-world tasks, focusing on reliable reasoning, tool use, and scalable inference.

Of late, I have been reading into AI safety and alignment and human-AI interaction (HAI/HCI), and I am excited about the research questions at their intersection with reliable and trustworthy language systems.

I am actively looking for PhD positions in AI/NLP for 2027, with interests including but not limited to model reliability, reasoning, trustworthy language systems, and broader NLP/LLM research. Feel free to reach out if you’d like to discuss research.