My papers and research manuscripts in machine learning, language models, computer vision and educational assessment.
arXiv preprint · arXiv:2607.26117 · 2026
A controlled study of self-repair, execution feedback, and blind retry across 1.5B, 3B, and 7B code models, with preregistered hypotheses and replication across model families.
Research manuscript · arXiv submission version, 2 Oct. 2026 · 2026
A reproducibility audit and three-seed comparison of ordinary and inverse-frequency sampling with Qwen2-VL-2B and QLoRA. Balanced accuracy improves modestly while overall accuracy falls on IDRiD and RetinaMNIST+; peak reserved GPU memory is 3.684 GiB.
Research manuscript · 2026
An exact worst-reference binary Dice degradation bound computed from three observable mask counts, with certified partial outputs. Evaluated on ClinicDB, Massachusetts Buildings, and Kvasir-SEG; the bound limits loss relative to a baseline, not absolute segmentation quality.
Saved-model demonstration; predictions are not fact checks.
arXiv preprint · arXiv:2609.25006 · 2026
An audit of leakage and shortcut learning using DistilBERT, TF–IDF controls, and cross-corpus transfer to examine the gap between in-corpus scores and generalization.
Illustrative grading workflow, not a recorded scoring result.
Submitted to IEEE UPCON 2026 · Paper ID 1525 · 2026
A framework combining OCR, SBERT, CLIP, Pix2Tex, SymPy, and Gemini to evaluate text, diagrams, and mathematics. Reported Pearson correlation with human scores: r = 0.77.
Actual retinal images; scanning guides are illustrative.
Submitted to arXiv · Decision pending · 2026
Parameter-efficient adaptation of a 2B vision-language model using QLoRA under a 6 GB VRAM constraint, with resolution ablations, per-class error analysis, and explanation evaluation.