I’m a Ph.D. candidate at the University of Central Florida, advised by Dr. Niloofar Yousefi in the Complex Adaptive Systems Lab, with three years of industry ML experience before my Ph.D. I build inference-time methods that make deployed LLMs, LLM judges, and agents reliable without retraining them, and I test when that reliability is real and when it is only apparent. My first-author work appears at NeurIPS 2026 (Main Track) and EACL 2027 (Main Conference).

📢 Ph.D. candidate (expected May 2028) · actively seeking Summer 2027 research internships in reliable LLMs, LLM evaluation, and agent safety. Please reach out if you have an opening.

Recent News (See All)

Nov 2026🎉 MedBayes-Lite accepted at Frontiers in Digital Health: a retraining-free layer that cuts confident, high-severity errors from clinical language models. Watch the one-minute video.
Nov 2026🎉 UAT-LITE accepted at EACL 2027 (Main): inference-time uncertainty-aware attention for pretrained transformers.
Nov 2026🎤 NeurIPS 2026 poster and talk slides are on the NeurIPS page; watch the one-minute video.
Sep 2026🎉 When Policies Cannot Be Retrained accepted at NeurIPS 2026 (Main Track): closed-form steering of a frozen offline RL policy.
Sep 2026🎉 NEXUS accepted to Findings of AACL-IJCNLP 2026: runtime safety for tool-using LLM agents.
Sep 2026🔍 Reviewer for the Trustworthy AI for Good workshop at NeurIPS 2026.

Beyond Research

Before research, I built things for competitions: my team was a divisional champion and global nominee at the NASA Space Apps Challenge 2017 with the Drone for Green project, won the IEEE SS12 Innovation Challenge 2018 with a kidnap-prevention mobile app, and was a finalist for the Dr. Pradeep P. Thevannoor Innovation Awards 2019 in India.

I also like building models from the ground up: MiniHealthLM is a ~600M-parameter transformer I trained from scratch for healthcare text, and I release small tools such as mldriguard for data-drift detection (more on my open-source page). I write about AI on Medium.

Collaborations

University of Central Florida
New York University
UMBC
Mississippi State University
Univ. of Southern Queensland
Queensland Univ. of Technology
Purdue Fort Wayne
Meharry Medical College
Nemours Children’s Health
North South University
Ahsanullah Univ. of Sci. & Tech.

Featured Publications

Full list on Google Scholar. Each paper has a project page and a one-minute narrated video.

NeurIPS 2026

When Policies Cannot Be Retrained: A Unified Closed-Form View of Post-Training Steering in Offline Reinforcement Learning

E. Hossain, M. J. I. Basher, I. Garibay, O. Garibay, N. Yousefi
NeurIPS 2026, Main Track

In shortA robot's control policy is often frozen once trained. We show how to steer it toward a new goal without retraining. When the new guidance is poor, the policy falls back on its original behaviour instead of collapsing; when the guidance is good, how much it helps depends on how capable the original policy was.

Technical summary

Deployment objectives change after training, but the actor often cannot be retrained. Composing a frozen offline RL actor with a goal-conditioned prior as a product of experts yields the same policy as KL-regularized adaptation, with β = α/(1−α) for diagonal-Gaussian actors. The result is graceful degradation rather than universal gain: under a degraded prior the composition stays anchored to the frozen actor while additive and prior-only adaptation collapse, and across D4RL MuJoCo, AntMaze, and Adroit its effect is bounded by actor competence (HELP/FROZEN/HURT 4/4/4 over ten seeds).

EACL 2027 UAT-LITE architecture: Bayesian embedding calibration, uncertainty-weighted attention, MC-dropout uncertainty estimation, and a confidence-guided decision module Full figure

UAT-LITE: Inference-Time Uncertainty-Aware Attention for Pretrained Transformers

E. Hossain, S. R. Dipta, S. Neupane, R. Rana, R. Shwartz-Ziv, I. Garibay, N. Yousefi
EACL 2027, Main Conference

In shortLanguage models are often confidently wrong. Running an already-trained model about ten times with small random perturbations shows where its answers disagree, and using that disagreement makes its confidence scores more honest, with no retraining.

Technical summary

Inference-time method that uses Monte-Carlo dropout to make pretrained transformer attention uncertainty-aware. Improves calibration, selective prediction, and robustness without retraining.

Frontiers 2026 Harmful overconfident error rate by clinical risk category: the base model reaches about 0.2 on diagnosis, treatment, medication and high-severity questions, while MedBayes-Lite stays below 0.01 in every category Full figure

MedBayes-Lite: A Clinical Uncertainty Governance Layer for Risk-Aware Medical Decision Support

E. Hossain, M. M. H. Nipu, M. Sheikh, T. Nuzhat, R. Rana, S. Neupane, B. W. Schuller, N. Yousefi
Frontiers in Digital Health, 2026 (accepted)

In shortClinical language models are often confidently wrong, especially on high-stakes questions. A layer around an existing model measures how unsure it is, makes its confidence honest, and hands uncertain cases to a clinician, with no retraining.

Technical summary

A retraining-free, parameter-free governance layer combining Monte Carlo dropout, calibration of the predictive distribution, and a confidence-guided abstention gate. On MedMCQA and MedQA-USMLE it reduces ECE by 0.23 to 0.33 and drives harmful overconfident errors toward zero; under MedMCQA→MedQA-USMLE shift, confident errors on high-severity items fall from about 21% to near zero. It does not improve AURC, temperature scaling matches most of its calibration at lower cost, and it costs coverage.

Ten LLM judges vote 8-of-10 for response A, but three of the agreeing judges share the same blind spot; ignoring shared errors declares system A significantly better, while accounting for them finds no difference Full figure

Agreement Overstates Evidence: Error Dependence in LLM Judge Consensus

E. Hossain, N. Yousefi, S.-N. Lim
Under review at ICLR 2027

In shortWhen several AI models grade the same answers, their agreement looks like strong evidence. But they tend to make the same mistakes, so ten AI judges carry roughly the information of three or four independent ones, and ignoring this can change a study's conclusions.

Technical summary

Consensus among LLM judges assumes they fail independently. They do not: in a ten-judge bank the mean error correlation is 0.21, so the votes carry the evidence of about 3.5 independent judges, frontier judges from different providers are even more correlated (0.56), and ignoring the dependence flips significance conclusions in up to 28% of comparisons. Changing the voting rule is not a reliable fix, even when the error pattern is known, so the correction belongs in the statistical inference.

Under review Four panels: four inference actions, per-question outcomes, confidence detects wrong but not fixable answers, routers recover little of the oracle gain Full figure

Knowing When a VLM Is Wrong Is Not Enough: When Does Extra Compute Help?

Authors withheld during double-blind review
Under review, 2026

In shortVision-language models can look closer at an image or reason step by step, and adaptive methods pick one per question. Compared fairly against handing out the same actions at random at equal cost, trained routers gain at most 2.1 points, because confidence tells a router that an answer is wrong, not whether extra compute will fix it.

Technical summary

An evaluation protocol that isolates the value of the per-question decision: an equal-cost random allocation of the same fixed actions, same-weight baselines, and disjoint samples for selecting and scoring the oracle. On five small VLMs and 600 questions, held-out oracles gain 4.8–9.8 points over the best fixed action, yet trained routers recover under one-fifth of that, and VisionThink does not beat an equal-FLOPs mixture of its own fixed configurations. Confidence detects wrong answers (AUROC 0.77–0.84) but not fixable ones (0.55–0.64).

NEXUS architecture: tool-using LLM agents emit a structured plan that a runtime safety monitor screens with deterministic rules, argument inspection, a learned risk score, and a session manager before issuing one of four graded interventions Full figure

NEXUS: Structured Runtime Safety for Tool-Using LLM Agents

E. Hossain, M. M. H. Nipu, T. N. Ornee, R. Rana, N. Yousefi
Findings of AACL-IJCNLP 2026

In shortBefore an AI assistant runs a tool, such as sending an email or deleting a file, a fast checker reads its plan and decides: let it run, block it, ask the user, or ask the AI to revise.

Technical summary

A structured-plan safety monitor that chooses among four interventions, allow, block, request confirmation, or request revision, by combining deterministic rules, argument-level inspection, and a calibrated risk score for graded escalation. Reaches 0.949 F1 while adding 0.205 ms median latency, under 0.1% overhead on a typical agent loop.

ChannelGuard system diagram: the threat surface, a multi-agent pipeline with six information-bottleneck gates, and the pass, compress, or block decision made at each gate Full figure

ChannelGuard: From Borrowed Safety to Application-Controlled Defenses in Multi-Agent LLMs

E. Hossain, M. M. H. Nipu, F. T. J. Faria, T. N. Ornee, M. Sheikh
Under review at TMLR

In shortAI systems built from several cooperating models pass messages to each other. Many such pipelines look safe only because the cloud provider filters content; we add checks on every hand-off that the application controls itself.

Technical summary

Across 2,100 test cases and eight attack families, the apparent safety of multi-agent LLM pipelines often rests on opaque provider filters rather than the system itself. ChannelGuard places training-free information-bottleneck gates on every inter-agent channel, blocking tool-output poisoning outright and halving prompt-injection success while holding task accuracy at 0.867.

Parametric temporal conflict overview: a frozen LLM baseline returning the outdated fact, conflict localization by comparing standard and date-prefixed prompts, and residual-stream steering that recovers the newer fact Full figure

Right Knowledge, Wrong Answer: Characterizing Parametric Temporal Conflict in Open-Weight Language Models

E. Hossain, S. Saha, T. N. Ornee, S. S. Jennifer, U. C. Biswas, S. R. Dipta, R. Rana, N. Yousefi
Preprint, 2026

In shortA language model can store an up-to-date fact, such as who holds an office now, yet still answer with the outdated one. We show the newer fact is present and locate where inside the model the old answer wins.

Technical summary

Introduces Parametric Temporal Conflict, where a model holds the newer fact but its default forward pass still prefers the outdated one. On a verified benchmark of 8,746 Wikidata transitions, a date-prefix prompt recovers the newer fact in 61–81% of cases and activation patching flips 72–85%, localizing the preference to model-specific upper layers rather than missing knowledge.

More papers (3)
MPV pipeline: raw preference data passes through four verifier agents, a k-of-n consensus filter, and into a verified preference dataset used for DPO training Full figure

Fault-Tolerant Preference Alignment via Multi-Agent Verification

E. Hossain, M. Rahimimovassagh, S. Neupane, M. J. I. Basher, I. Garibay, N. Yousefi
ICLR 2026 Workshop on Principled Design for Trustworthy AI

In shortAI models learn from examples of better and worse answers, and some of those examples are wrong. Four independent checker agents vote on every example, and only the ones they agree on are used for training.

Technical summary

RLHF and DPO both assume the preference data is trustworthy. MPV drops that assumption: four verifier agents check every preference pair for factuality, safety, ethics, and trust, and only pairs clearing a k-of-n vote reach training. Consensus provably suppresses corrupted supervision at an exponential rate, and on Qwen2-7B it cuts the summarization refusal rate from 0.148 to 0.005 while holding win-rate at parity.

Safe and Scalable Collaboration in Multiagent LLM Systems: A Comprehensive Review

E. Hossain, M. M. H. B. Nipu, M. S. Mahmood, M. J. Hossen, M. F. Mridha
IEEE Transactions on Systems, Man, and Cybernetics: Systems, 2026

In shortA map of how teams of AI agents coordinate, communicate, and reason together, and where safety and scale break down.

Technical summary

A review of large language model based multiagent ecosystems, covering how agents coordinate, communicate, and reason across distributed settings, and what fails as those systems scale. Organized around four foundational pillars of multiagent LLM design, it maps the safety and scalability gaps that separate working demonstrations from deployable systems.

BIOGEN architecture: RNA-seq gene clusters pass through a retrieval layer over PubMed and UniProt, an interpreter agent, a critic ensemble with a consensus layer, and a persistent context memory before biological interpretation is produced Full figure

BIOGEN: Evidence-Grounded Multi-Agent Reasoning for Transcriptomic Interpretation in Antimicrobial Resistance

E. Hossain, M. Shoeibi, I. Garibay, N. Yousefi
Frontiers in Bioinformatics, vol. 6, 2026

In shortAn AI system that explains which genes help bacteria resist antibiotics, where every claim must cite a real, checkable database entry.

Technical summary

A multi-agent framework that interprets RNA-seq gene clusters by pairing retrieval over PubMed and UniProt with an interpreter agent and a three-way critic ensemble. Every claim carries a verifiable identifier: the non-verifiable identifier rate is 0.000 against 0.100 for an LLM-only baseline, and BIOGEN is the only system holding that rate at zero across all five bacterial datasets, including under cross-organism and cross-condition shift.

Talks

  • Dec 2026When Policies Cannot Be Retrained (poster). NeurIPS 2026, Atlanta. 1-min video · poster & slides
  • Jan 2026Multi-agent verification for safe preference alignment (poster). Amazon Trusted AI Symposium, New York.
  • Oct 2025CITE (presentation). Molecular Machine Learning Conference, MIT Jameel Clinic. paper
  • 2025Analyzing post-vaccination heart attack risk using machine learning. Guest lecture. video
  • 2024AI in healthcare and what it can mean for rural America (invited talk). Mississippi Health & Innovation Conference.
  • RecordedUsing AI for personalized care and nutrition for rural Americans. video
  • RecordedAnalysing and identifying COVID-19 risk factors using machine learning. video

In the Press

Research Mentorship

Md Mehedi Hasan Nipu B.Sc., North South University, Bangladesh · 2025 – Present
Md Ahadul Haque Suvo Undergraduate Student, Aviation and Aerospace University, Bangladesh
Rifat Bin Reza Undergraduate Student, BRAC University, Bangladesh
  • Multilingual Small Language Models on Bangla OS Commands - Ongoing, 2026 – Present
Ilmi Tabassum Graduate Student, Macquarie University · 2022 – 2023
Oishee Ghosh Graduate Student, Jahangirnagar University · 2020 – 2022

Peer Review

WorkshopsNeurIPS 2026 AI4GOODICML 2026 GenBioICLR Trustworthy AI
JournalsScientific ReportsPLOS ONEComputers in Biology and MedicineCMBBE