
Angelina Parfenova
When Models Agree Too Much
Angelina Parfenova researches reasoning and interpretability, looking at what happens inside the black box.
Outputs vs Inner Reasoning
A model's final answer does not always match its own reasoning or next-token probabilities.
Convergence Amplifies Bias
In multi-round discussions, three or more models converge after about five rounds, even when told to hold opposing views. A moderator agent did not help, since it shares the same training data. "If there is some bias and some models are sharing it, I'm afraid it's going to amplify."
Two Ways to Evaluate
First, measure agreement between human raters with metrics such as Krippendorff's alpha or Cohen's kappa, and treat it as the ceiling for human-model agreement. Second, compare the distribution of human answers with the distribution produced by multiple agents.
Context Shapes Reasoning
When she fed a model random reasoning for the same question, not only its answer but also its internal next-token predictions changed. That helps explain why models converge when they read each other's outputs.
What Is Fairness?
Her closing question: if we make an LLM fair, morally good or neutral, is that already a bias? "We talk so much about bias, fairness, discrimination, but we never gave a definition."
Treat human disagreement as the ceiling for evaluation, and compare distributions of views rather than single answers.