Reflective CoT Faithfulness Study (Think Twice)
I collaborated on a research study evaluating whether reflective chain-of-thought improves LLM reasoning faithfulness under controlled prompt bias.
Reflective CoT reduced bias susceptibility by about 7 percentage points over standard CoT across 650 cases
I co-developed a reproducible Python evaluation pipeline across Gemini 2.5 Flash Lite and LLaMA 3.2 3B, comparing direct answering, chain-of-thought, and reflective chain-of-thought across 13 BIG-Bench Hard categories in zero-shot and few-shot settings.
We found that Reflective CoT reduced bias susceptibility by about 7 percentage points over standard CoT, but models still produced plausible rationales that did not reveal the prompt cue driving biased answers. Our results show that structured reflection improved bias resistance but did not establish faithful reasoning.