EECS undergraduate at GIST · Undergraduate Research Intern at the GIST Data Science Lab (advisor: Prof. Sundong Kim)
I research test-time adaptation and self-correcting reasoning in AI models, using the ARC-AGI benchmark as my testbed. I also build and ship full-stack AI products.
-
Test-time training for Tiny Recursive Models on ARC-AGI-2 · 📄 poster
I trained a TRM checkpoint with the ARC-AGI-2 evaluation tasks held out, then compared full fine-tuning and LoRA for adapting it to tasks it had never seen. LoRA underfits, with a loss about 70× higher. Full fine-tuning reaches 4.07% pass@2 on the 120 held-out tasks; the original TRM scores 7.80%, but it saw these tasks during pre-training. Presented at the 10th GIST AX-Challenge (2026). -
Iterative self-correction in language models (undergraduate thesis, in progress)
A confidence ("doubt") head that predicts whether an LLM's answer is correct and decides when to revise it, with revision strength scaled to the model's uncertainty.
| Project | What it is | Highlights |
|---|---|---|
| Composed image retrieval | Find a garment from a photo plus a text edit ("make it sleeveless") | LoRA-tuned CLIP, cross-attention fusion, FAISS · Recall@10 25.8% |
| Tic-tac-toe RL | Value iteration vs. self-play Q-learning | Q-learning matches the optimal policy in all 4,520 positions |
| HabeshaSphere | Everyday web tools for Ethiopian users | Exchange rates, tax calculator, Ethiopian calendar · ~100–200 monthly users |
| GPA Calculator | Semester and cumulative GPA web app | React · live demo |
| Password Manager | Encrypted command-line password vault | scrypt key derivation, authenticated encryption, tests |
| Amharic TTS | Amharic text-to-speech web app | React frontend for an n8n speech pipeline |
🌐 berekety1.github.io · bereketyimolal8991@gmail.com · habeshasphere.com · Gwangju, South Korea
Languages: Amharic (native) · English (fluent) · Korean (elementary)