Afrah Nazar / Research
Research
Five projects and one question: when a network makes a decision that matters, can we see how — and trust it?
-
P-01
Individual report · since January 2026
AI can judge whether a patient qualifies for a clinical trial — but not faithfully explain why. I searched all 1,152 attention heads of the open-source model Qwen3-8B and isolated a six-head circuit that appears to read the eligibility rule, retrieve the patient’s medications, cross-check them, and integrate the verdict. Ablating even one of the six heads sharply degrades accuracy across 100 test patients.
2026Individual report
Read the project page →
-
P-02
2nd Place (Software) · ACSEF · March 2026
Trials are delayed or abandoned when they cannot recruit; LLM screening is fast but unverifiable. I trained lightweight linear probes on Qwen3-8B’s internal activations to verify its eligibility decisions. Probes verified decisions at 89.9% across 100 unseen patients per trial — versus direct classification by GPT-4o-mini (70.4%) and Claude Haiku (84.0%) — and are trial-specific: pooling 500 trials drops accuracy to chance.
2026ACSEF · 2nd Place
Read the project page →
-
P-03
ACSEF · March 2025 · Abstract No. 1368, AAN 2026
Can a language model detect dementia from a transcribed picture description? Reading Llama 3.1-8B’s internals did it far better than prompting the very same model — 0.88 against 0.55 F1, 0.87 against 0.61 accuracy — across 231 Cookie Theft descriptions from DementiaBank’s Pitt Corpus. Co-authored with a school project partner; Abstract No. 1368, submitted to the American Academy of Neurology Annual Meeting, 2026.
2025ACSEF · AAN 2026 abstract
Read the project page →
-
P-04
4th Place (Biotechnology) · ACSEF · March 2025
A human brain has 86 billion neurons and a language model 500 billion — far too many to trace. A fruit fly has 186,000, and its wiring has been mapped in full. I won access to the FlyWire connectome, rebuilt the fly’s smell circuit as 120 simulated neurons in Brian2, and trained it with reward and punishment: the same connections learned to attract, then repel, then attract again — but never told a strong smell from a weak one.
2025ACSEF · 4th Place
Read the project page →
-
P-05
4th Place (Software) · ACSEF · March 2024
An individual project evaluating generative AI for ADHD screening and management. 4th Place, Software, at the 2024 Alameda County Science & Engineering Fair.
2024ACSEF · 4th Place
Read the project page →
-
P-06
Self-initiated · since 2021 · 4 hrs a week
Every project here was self-initiated, about four hours a week since 2021. This is the part that does not show up in a results table: the papers, the cold emails that found me two advisors, the tools I had to learn first, and the GPU time I paid for.
2021 –Method, not a project
Read the project page →