Current research
Trust-Calibrated Reinforcement Learning for Adversarially Robust Multimodal Agents
Current manuscript: Learning When To Trust: Training-Time Calibration for Reinforcement Learning under Adversarial Observation Corruption — manuscript in preparation.
Modern multimodal agents often assume that their observations are reliable. In realistic environments, individual modalities can become noisy, missing, corrupted, contradictory, shifted, or adversarially manipulated. My research investigates whether an agent can estimate the reliability of its observations and adapt its sequential decision-making accordingly.
Ph.D. research in Computer Science at the University of Southern Mississippi, advised by Dr. Rabab Abdelfattah.
Research questions
Open questions this work investigates — none of these are claimed as solved.
- Q1.Can an agent identify when an observation or modality should not be trusted?
- Q2.Can calibrated reliability estimates improve sequential decision-making?
- Q3.How should reliability estimates affect an RL policy?
- Q4.How does reliability estimation behave under unseen corruption and distribution shift?
- Q5.What happens when individual modalities disagree?
- Q6.How robust is the trust estimator itself to adversarial manipulation?
- Q7.Can an agent abstain, request another observation, or defer to a human when uncertainty is high?
Current mathematical formulation
I am currently developing a reliability estimator of the general form
rt = σ( fφ( õt, ht−1 ) )
- õ_t
- potentially corrupted observation at time t
- h_{t−1}
- prior agent/context state
- f_φ
- learned reliability estimator
- σ
- maps the estimator output to [0, 1]
- r_t
- estimated reliability signal
The sigmoid output rt is a candidate learned reliability/trust score, not yet a calibrated probability — its calibration properties must be established empirically and theoretically before it can be treated as one.
A later multimodal formulation may estimate reliability separately for individual modalities, so that an agent can down-weight or discard a single corrupted channel rather than distrusting the entire observation.
Experiments
Experiments are developed and published in the open on TraceVox Research, which records complete per-timestep evidence, replays runs exactly, and exports provenance-carrying research bundles. Published experiments are browsable in the Public Research Library.
Gate experiment: corruption robustness (LavaGapS7)
Recorded pilot · published August 2026
- Research question
- PPO on MiniGrid-LavaGapS7: how does performance degrade under observation corruption, and does corruption-trained robustness transfer to held-out corruption families?
- Environment / task
- MiniGrid-LavaGapS7-v0 (grid navigation with lava hazards)
- Agent / model
- PPO (ppo_minigrid_baseline)
- Training conditions (arms)
- Clean-trained PPO vs. corruption-trained PPO (mask corruption applied with probability 0.25 during training)
- Evaluation conditions
- Greedy evaluation under observation corruption at probability up to 0.5, including corruption families held out from training (e.g., mask-trained policy evaluated under shuffle)
- Threat / corruption model
- Random observation corruption (mask, shuffle families). This is corruption robustness, not adversarial robustness — no adversary optimizes the perturbation.
- Scale
- Pilot / gate experiment: small number of seeds; 20–30 evaluation episodes per recorded condition
- Metrics
- Success rate, mean/median return, episode length, safety-failure rate, timeout rate, realized corruption rate
- Artifacts
- Per-timestep recorded traces (research.v1 schema), exact replays, provenance manifest, machine-readable JSON, citation metadata
Limitations & scope
- This is a pilot-scale gate experiment: few seeds and small episode counts. It establishes the recording/replay pipeline and baseline corruption sensitivity, not final conclusions.
- The perturbations here are random corruptions (mask, shuffle). They are not adversarial attacks — no adversary optimizes or selects the perturbation under a threat model. The recorded pilot demonstrates corruption robustness, not adversarial robustness; adversarial evaluation is a separate, planned stage of the research.
- No trust estimator is deployed in these runs yet; the recording schema reserves trust fields, which remain null until the estimator exists.
Publications & manuscripts
Learning When To Trust: Training-Time Calibration for Reinforcement Learning under Adversarial Observation Corruption
Olanrewaju Muili · Manuscript in preparation
Evaluating hybrid guardrail architectures for prompt injection defense in large language models
Olanrewaju Muili · SSRN preprint · 2026
Sinkhole susceptibility analysis using machine learning for west central Florida
Olanrewaju Muili, Hassan A. Babaie · Applied Computing and Geosciences, vol. 27, 100262 · 2025
Research software
TraceVox Research — an open research environment for reproducible experiments in trustworthy AI, reinforcement learning, multimodal agents, corruption robustness, adversarial robustness, evaluation, and AI safety. Distributed publicly via PyPI (pip install "tracevox-ai[rl]"); experiments run on the researcher's own compute.
Prior scientific ML research
LLM safety: hybrid guardrail evaluation (2026, sole-authored preprint)
Systematic evaluation of prompt-injection defenses on a two-tier 625-prompt benchmark spanning 18 attack categories. The hybrid regex + LLM-classifier configuration reached a 98.9% attack block rate with 0.0% benign false positives on the adversarially hard tier; regex-only filtering degraded to 24.6%.
Simulation-to-ML for marine energy anchoring (U.S. Department of Energy Marine Energy Postgraduate Fellowship)
Research engineering with Deep Anchor Solutions: led development of finite-element workflows for marine-energy anchoring systems, including soil-structure interaction, parametric modeling, and simulation-to-ML dataset generation — 5,000+ numerical simulation runs, in collaboration with engineering teams.
Geospatial ML for karst hazards (peer-reviewed, 2025)
Machine-learning sinkhole susceptibility analysis for west central Florida (Applied Computing and Geosciences), extended into the Karst Intelligence Agent — a multimodal decision-support system combining XGBoost susceptibility modeling, imagery analysis, and natural-language interpretation.
Collaboration interests
I'm interested in collaborating on trustworthy AI, reinforcement learning under unreliable observations, multimodal agents, adversarial robustness and its evaluation, uncertainty and calibration, AI safety evaluation, and scientific/simulation-driven ML — and in research internships in these areas.