Experience · 05
Algoverse — cognitive bias in LLMs
By Haider Ali · AI/ML Researcher, Algoverse AI Research ·
Haider Ali leads a team at Algoverse AI Research, on a full scholarship, investigating anchoring bias in LLMs applied to medical diagnosis.
AI/ML researcher at Algoverse on a full scholarship, leading a team investigating anchoring bias in LLMs applied to medical diagnosis.
May 2025 → · ai/ml researcher (scholarship) · active
I lead a scholarship-funded team at Algoverse AI Research studying anchoring bias in LLMs used for medical diagnosis, and developing a novel fine-tuning approach to reduce it.
Key facts
- Full academic scholarship
- Research team lead
- Anchoring bias in medical LLMs
- Fine-tuning on patient narratives
The problem
- Anchoring bias: fixing on the first plausible diagnosis and no longer weighing alternatives, as human doctors also do.
- Approach: measure how often current models do it on diagnostic prompts, then fine-tune to reduce it.
What I did
- Lead the team on a full Algoverse academic scholarship, coordinating a cross-disciplinary collaboration.
- Designed the experiments and evaluation metrics.
- Fine-tuning data: netnographic patient data, meaning real patient narratives, not clinical summaries.
- Early experiments: GPT-4 prompted for structured, chain-of-thought diagnoses, scored against ground truth.
Stack
- python
- openai api
- llm fine-tuning
- chain-of-thought prompting
- evaluation metrics
Links
#ai #llm #research #medical