LLM Human Response Alignment: A Multi-Sample Debiasing Framework

Li Jiang, Xiao Liu

Keywords: LLMs; synthetic human responses; debiasing

TL;DR: We propose debiasing LLM-as-human proxies by feeding the full multi-sample response vector to the debiasing module. Provably dominant over the baselines with most 50%+ lower error.

Abstract: Large Language Models (LLMs) are increasingly used as proxies for human respondents in social science, behavioral, and marketing experiments, yet their synthetic responses diverge systematically from human ones. We study the post-hoc debiasing route that keeps the LLM frozen and learns an external debiasing estimator to align the LLM responses with the human responses. We model both human and LLM responses as latent distributions to capture their inherent heterogeneity. However, the two distributions are observed asymmetrically, with the collection cost capping the human side at a few observations per individual, while the LLM is queryable at negligible cost. Existing pipelines collapse the LLM side to a scalar before correction and discard the distributional signal that repeated cheap queries could otherwise capture. Motivated by this missed signal, we employ the full multi-sample vector of LLM responses as the input feature to a debiasing module on both population-level and individual-level debiasing tasks. Under squared loss, we give an information-theoretic motivation for retaining the full vector, showing it weakly Bayes-dominates any compression and strictly so on supra-mean targets. Across three benchmarks, our method attains the best value on the majority of metric cells at both the population and individual levels, reducing prediction error over the uncorrected Base LLM by over 50% on multiple tasks, and yields better distributional alignment to the human responses against the baselines.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 66