obed junias

I’m a graduate researcher in Computer Science at the University of Colorado Boulder, where I work within the BLAST Lab under the supervision of Dr. Maria L. Pacheco. My work sits at the intersection of natural language reasoning, LLM safety, and agentic systems.

Currently, a primary focus of my research is the long-term reliability of foundation models. I am exploring data quality and model collapse, investigating how synthetic data proliferation affects model reliability over successive training generations. I specifically study how this collapse manifests during foundation model pretraining, approaching it from a systems perspective to understand how robust pretraining pipelines can be implemented and evaluated at scale in open, transparent, and sustainable model ecosystems. I aim to extend these investigations into reflective post-training frameworks, including reinforcement learning (RL)-based approaches.

This research builds upon my established foundation in natural language processing. I develop interpretable reasoning systems and benchmarks for commonsense and logical inference [ACL 2026], exploring structured frameworks as a way to make machine reasoning more transparent, structured, and logically grounded.

In parallel, I work with Dr. Theodora Chaspari on bias detection and fairness evaluation in LLMs, particularly within the mental health domain—work that reflects my broader commitment to building AI systems that remain trustworthy and equitable across diverse populations.

Looking ahead, my overarching vision is to advance the frontiers of interpretable reasoning and LLM safety. I strive to develop methodologies that ensure AI systems are not only highly capable, but remain transparent, logically grounded, and deeply aligned with positive human outcomes.

research interests

Interpretable Reasoning & Natural Language Understanding: Commonsense and logical inference, structured and neuro-symbolic reasoning
LLM Safety & Model Collapse: Fundamental reliability of pretraining, synthetic data proliferation, recursive training degradation
Responsible AI & Fairness in ML: Bias mitigation in healthcare, social and moral alignment, ethical development of AI
Agentic Systems & Post-Training Frameworks: Self-reflective agents, reinforcement learning (RL) for alignment, transparent model ecosystems

I’m actively seeking opportunities in NLP and related areas.

Feel free to reach out for research collaborations or other opportunities.

news

Apr 07, 2026	Thrilled to announce that our work, “LOGICAL-COMMONSENSEQA: A Benchmark for Logical Commonsense Reasoning,” has been accepted for publication at the ACL 2026 Main Conference. See you in person! arXiv:2601.16504
Jan 27, 2026	New preprint alert! We introduce LOGICAL-COMMONSENSEQA, a benchmark designed to isolate multi-fact and compositional inference in LLMs using logical operators (AND, OR, NEITHER/NOR). Our evaluation reveals that fluent reasoning often masks underlying logical failures, especially in negation-based tasks. Check the paper on arXiv!
Dec 08, 2025	Our paper, Assessing Algorithmic Bias in Language-Based Depression Detection, has been officially published in the 2025 IEEE EMBS International Conference on Biomedical and Health Informatics (BHI) proceedings. You can find the full paper on IEEE Xplore.
Oct 20, 2025	I will be in Atlanta for the IEEE-EMBS BHI Conference from October 26-29, 2025, where I will be presenting my work on evaluating and mitigating bias in LLMs.
Oct 15, 2025	I have been selected by IEEE-BHI as an NSF–EMBS–Google Young Professional NextGen Scholar as an early researcher.

selected publications

Assessing Algorithmic Bias in Language-Based Depression Detection: A Comparison of DNN and LLM Approaches

Obed Junias, Prajakta Kini, and Theodora Chaspari

In 2025 IEEE EMBS International Conference on Biomedical and Health Informatics (BHI), 2025

DOI arXiv Bib HTML PDF

@inproceedings{junias2025assessing,
  title = {Assessing Algorithmic Bias in Language-Based Depression Detection: A Comparison of DNN and LLM Approaches},
  author = {Junias, Obed and Kini, Prajakta and Chaspari, Theodora},
  booktitle = {2025 IEEE EMBS International Conference on Biomedical and Health Informatics (BHI)},
  pages = {1--7},
  year = {2025},
  organization = {IEEE},
  doi = {10.1109/BHI.2025.11269509},
  url = {https://ieeexplore.ieee.org/abstract/document/11269509},
}

LOGICAL-COMMONSENSEQA: A Benchmark for Logical Commonsense Reasoning

Obed Junias and Maria Leonor Pacheco

arXiv preprint arXiv:2601.16504, 2026

Abs arXiv Bib PDF

Commonsense reasoning often involves evaluating multiple plausible interpretations rather than selecting a single atomic answer, yet most benchmarks rely on single-label evaluation, obscuring whether statements are jointly plausible, mutually exclusive, or jointly implausible. We introduce LOGICAL-COMMONSENSEQA, a benchmark that re-frames commonsense reasoning as logical composition over pairs of atomic statements using plausibility-level operators (AND, OR, NEITHER/NOR). Evaluating instruction-tuned, reasoning-specialized, and fine-tuned models under zero-shot, few-shot, and chain-of-thought prompting, we find that while models perform reasonably on conjunctive and moderately on disjunctive reasoning, performance degrades sharply on negation-based questions. LOGICAL-COMMONSENSEQA exposes fundamental reasoning limitations and provides a controlled framework for advancing compositional commonsense reasoning.
@article{junias2026logical, title = {LOGICAL-COMMONSENSEQA: A Benchmark for Logical Commonsense Reasoning}, author = {Junias, Obed and Pacheco, Maria Leonor}, journal = {arXiv preprint arXiv:2601.16504}, year = {2026}, }