Publications

When a clock reads 2:50 in a mirror, humans instantly infer the true time is 9:10, yet current Vision-Language Models overwhelmingly report “2:50,” hallucinating the canonical orientation rather than reasoning about the reflection. We term this systematic failure reflection blindness and introduce MIRA, the first benchmark designed to diagnose it: 713 real-world images and 3,529 question-answer pairs spanning 10 scenario categories (licence plates in rear-view mirrors, analog clocks, hands, ears, shoes, chemical formulas, ambulance mirror-text, selfie cameras, water reflections and indoor mirrors) organised into five question types. Evaluating 11 VLMs in both single-turn and multi-turn settings, we find that all models achieve only 39–48% accuracy versus 94% for humans, with frontier systems clustered at 42–44%, and performance collapsing most severely on tasks requiring mental un-flipping of text and left–right chirality judgment. Multi-turn dialogue does not improve performance; most models degrade by 1–14 percentage points when questions are posed sequentially.

We introduce the Massive Video Embedding Benchmark (MVEB), a 23-task benchmark for video embeddings spanning classification, zero-shot classification, clustering, pair classification, retrieval and video-centric question answering. We evaluate 33 models and find that no single model dominates: MLLM-based embeddings lead on classification, clustering, pair classification and QA; multimodal binding leads on retrieval and zero-shot classification; generative MLLMs without contrastive adaptation collapse on cross-modal tasks. Paired video-only vs. audio+video evaluations show that audio's contribution depends on dataset annotation provenance: audio helps when labels were produced from both modalities and hurts when they were produced from visuals alone, a six-point gap consistent across model families. MVEB is derived from MVEB+, a 184-task pool, and integrates into the MTEB ecosystem for unified evaluation across text, image, audio and video.

Designed and evaluated 45 configurations across five small language models and three quantization levels for medical question answering under edge deployment constraints.

Software-defined networking decouples the control plane from the data plane, fostering centralised control and improved programmability, but exposes new attack surfaces against the centralised controller. Among these threats, DDoS, specifically SYN Flood, can overwhelm the controller with requests. The proposed system compares Random Forest, eXtreme Gradient Boosting, Artificial Neural Networks and Decision Tree classifiers, detecting brute force, DoS, DDoS, probe and user-to-root attacks using features readily available at the SDN controller, avoiding computationally expensive deep packet inspection. Benchmarked on CICIDS-2017, ensemble classifiers, particularly Random Forest, significantly outperform the alternatives on accuracy, precision, recall and F1-score. The resulting framework is lightweight and feasible for integration directly with the SDN control plane.

The prevalence of electronic transactions in financial systems has significantly increased the risk of credit card fraud, and traditional detection methods struggle to keep pace with evolving fraudulent techniques. This study explores outlier detection as a robust approach to identifying anomalous patterns indicative of fraud. Leveraging statistical and machine learning algorithms, the research develops a framework for distinguishing genuine from fraudulent transactions through analysis of transaction amount, location, time and user behaviour, aiming to enhance detection accuracy while minimising false positives.

We present TabuLM, the first language model pre-trained on Kinyarwanda tabular data. Kinyarwanda is a morphologically rich Bantu language spoken by over 12 million people in Rwanda, yet lacks any dedicated tabular representation learning resource. TabuLM extends KinyaBERT-large with additive row, column and cell-type embeddings and a learned table-structure attention bias. Pre-training uses two new objectives, Masked Cell Recovery and Column Type Prediction, over 172 Rwandan government tables (~35,000 cells), and introduces TabQA-kin, the first native Kinyarwanda table question-answering benchmark. TabuLM achieves 62.0% exact match, outperforming KinyaBERT-large by 5.7 EM points and all multilingual baselines by 11.7–12.7 points.

We present KinyaEmbed, the first dedicated sentence embedding model for Kinyarwanda. Existing multilingual embedding models such as LaBSE, mE5-large and OpenAI text-embedding-3-large perform poorly on Kinyarwanda due to severe under-representation in their pre-training corpora. KinyaEmbed is built on KinyaBERT-large and trained via a four-stage curriculum using MultipleNegativesRankingLoss, drawing on paraphrase pairs from the Official Gazette of Rwanda, NLLB-translated MNLI triplets, English–Kinyarwanda OPUS-100 alignment and KinyaCOMET-filtered pairs. We introduce Wiki-RW-STS, a contamination-free Kinyarwanda STS benchmark. A seven-checkpoint ensemble achieves Spearman ρ = 0.7298 on SemRel2024-rw, surpassing mE5-large by 20.9% and OpenAI text-embedding-3-large by 41.0%.

A reproducible, cross-platform evaluation of the data structures underpinning limit order books in low-latency matching engines, accompanying a C++ implementation sustaining roughly 3M orders per second.

SR-LoRA and CLAC address spatial and semantic conflicts that arise when multiple LoRA adapters are composed, through softmax routing and cross-adapter attention conditioning. Evaluation uses a 480-compositions-per-method GPU benchmark on Stable Diffusion v1.5 with rank-4 LoRAs, DPM-Solver++ and CLIP ViT-L/14, together with τ/α/δ and N=3 ablations. Extending the protocol from three seeds to ten revised the text-alignment gain from 6.8% down to 4.3%, and CLAC raises identity consistency from 0.839 to 0.915, 7.6 points above the strongest baseline.

A verifiable reinforcement learning environment for autonomous network troubleshooting agents, covering environment design, verification logic, and reward mechanisms that score agent actions against ground truth rather than relying on a judge model.

A verifiable agent-learning environment for API contract migration, applying the same ground-truth verification approach to software engineering tasks: benchmark design, verification logic and reward mechanisms evaluated against known-correct outcomes.

Decision analysis, with its formal influence-diagram representation, unambiguous ground truth and well-established pedagogy, is an unusually clean domain in which to test whether LLMs can act autonomously: the correct answer to every question is known exactly, in closed form, before any model is queried. We find that the order of fields in the response format alone changes the answer, producing $60, $70, $82.35 and $67.50 under answer-first elicitation against a stable $76.00 under reasoning-first: a measurement artifact that affects how LLM reasoning is evaluated generally. A code-generation setup solves 128 of 128 problems that in-context reasoning fails, and error feedback lifts GPT-4o from 10.2 to 23.0 of 29 tests without ever reaching perfection. The architectural conclusion: models should supply only genuinely uncertain judgments, while code performs any numerically sensitive computation.

Generating patient-specific 3D brain MRI volumes without paired training data, using diffusion models. The work tests the hypothesis that gradient guidance fails precisely where adapter conditioning succeeds, via an oracle experiment. Residual blur was traced to conditional-mean averaging over plausible anatomies, localised to high-frequency wavelet subbands, and addressed with a wavelet-domain loss that improved sharpness. Includes an ADNI data pipeline and an nnUNet-based segmentation evaluation pipeline.