Research & Publications

My research work in applied AI: foundations, alignment, prediction

CPTAH: Does Continued Pre-Training Change How Alignment Affects African LLMs?

Balbino Tchoutzine• September 21, 2026Completed

Remote research internship: the same SFT-then-DPO pipeline applied to Qwen3.5-4B-Base and AfriqueQwen (CPT on 35.5B tokens, 50 African languages), evaluated in Hausa on the HHH axes with no judge. CPT buys language processing but no detectable truthfulness; on moderation (AfriHate), alignment degrades the non-CPT model and improves the CPT one.

alignmentCPTSFTDPOQLoRAHausa

Read report ↗

Leakage-Resistant Models for MCI-to-Dementia Conversion Risk Prediction

Balbino Tchoutzine• July 24, 2026In progress

Research internship at UMMISCO on predicting conversion from MCI to dementia. I designed a cross-cohort analysis plan (NACC, ADNI, Framingham) with every choice fixed before seeing the data, ran a pilot on OASIS-2 (AUC up to 0.75), and deployed Lumenscreen, a teaching demonstrator. Next step: data access.

mild cognitive impairmentdementia conversionanalysis plandata leakageexternal validationNACCADNIFraminghamOASIS-2pilot study

Read report ↗

T-Grokk: Dynamic Label Noise Favors Grokking

Balbino Tchoutzine• February 1, 2026Unpublished

Unlike static noise, corrupting labels at every training step (dynamic noise) acts as implicit regularization: at 50% noise, the model reaches 100% test accuracy versus a complete failure under static noise.

Read paper ↗

Analyzing Grokking Dynamics in the Presence of Noise (ξ)

Balbino Tchoutzine• November 1, 2025Unpublished

How the fraction of corrupted labels (ξ) affects memorization and generalization (grokking) in an MLP trained on modular addition, up to a critical threshold where the model collapses.

Read paper ↗

Research Interests

Foundations of deep learning and learning dynamics (grokking, robustness to label noise)
Alignment and fine-tuning of foundation models
Long-horizon clinical prediction (dementia 2 to 5 years out, from early diagnostics and MRI)
Computer vision and multimodal learning
Geospatial ML and foundation models for Earth observation
Low-resource NLP / speech (African languages as a demanding testbed)
Competitive and large-scale applied machine learning