This repository began in March 2026 as a research project measuring positional bias in LLM multiple-choice answering and testing whether prompting interventions reduce that bias.
It has since been split into two standalone paper repos:
- cotenthusiast/two-stage-prompting-paper — sourced from the
two-stage-promptingbranch - cotenthusiast/model-generalization-paper — sourced from the
model-generalizationbranch
This repo is now being repurposed as mcq-framework, a reusable MCQ evaluation framework for LLM robustness research. It is currently early stage and not yet ready for use.
Existing branches are preserved as historical record.
For current research work, see the two paper repos linked above.