LLMs are increasingly used to simulate and study human behavior. Such studies assume richer personas improve predictions of specific people's answers. We test this premise across four public datasets. First, we split each person's answer to an item into the item mean (the average human answer to that item) and the person's deviation from the item mean. Much of the accuracy that surrogate studies report comes from predicting item means. LLM predictions explain little of the deviations. We call this pattern item-mean surrogacy. Second, we split each deviation into stable signal and noise using a dataset in which the same participants answered the same questions twice. Stable signal is the part of a deviation that repeated answers share. Noise is the part that changes. In this dataset, more than half of the variance in deviations persists across occasions. Most of what the LLM misses is stable signal rather than noise. Third, we split stable signal into the person signal, a person's average deviation across questions, and the person-by-item signal, their remaining question-specific deviations. The person-by-item signal accounts for about 8.5 times as much variance as the person signal. The LLM recovers about a tenth of it. Richer personas, different models, and fine-tuning on human answers still leave most of the person-by-item signal unrecovered. Both fine-tuned models recover less of this signal than their untuned versions. These findings call for closer empirical scrutiny of LLMs as human surrogates.
John Seon Keun Yi Boston University, Boston University, Aaron Mueller Boston University, Dokyun Lee Boston University, Boston University
原文摘要
Multi-agent debate has been shown to improve reasoning in large language models (LLMs). However, it is compute-intensive, requiring generation of long transcripts before answering questions. To address this inefficiency, we develop a framework that distills multi-agent debate into a single LLM through a two-stage fine-tuning pipeline combining debate structure learning with internalization via dynamic reward scheduling and length clipping. Across multiple models and benchmarks, our internalized models match or exceed explicit multi-agent debate performance using up to 93% fewer tokens. We then investigate the mechanistic basis of this capability through activation steering, finding that internalization creates agent-specific subspaces: interpretable directions in activation space corresponding to different agent perspectives. We further demonstrate a practical application: by instilling malicious agents into the LLM through internalized debate, then applying negative steering to suppress them, we show that distillation makes harmful behaviors easier to localize and control with smaller reductions in general performance compared to steering base models. Our findings offer a new perspective for understanding multi-agent capabilities in distilled models and provide practical guidelines for controlling internalized reasoning behaviors.
cs.CL · cs.MA
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表