Xuan Liu, Haoyang Shang, Yuanjun Feng, Shao-Yu Chu, Qiyu Li, Yash Raj Shrestha, Haojian Jin
原文摘要
We introduce HumanStudyProgram, a framework that represents a documented human study as an executable program with configurable LLM-based silicon participants. Its computational abstraction separates study logic from participant implementation through standardized APIs for presenting information and collecting responses. We implement a researcher-facing authoring tool that reconstructs source-linked study descriptions from papers and supplementary materials, supports researcher review, and generates executable HumanStudyProgram artifacts. Researchers can then configure silicon participants, run the programs, and analyze results from aggregate findings to individual responses. We demonstrate the approach through study cases and evaluate the tool through a hands-on expert review with 12 social science researchers. All completed the workflow and identified potential uses in understanding published protocols, prototyping studies, and examining AI behavior. We aim to improve the reproducibility of human studies by reconstructing their documented procedures as programs that others can reuse and build upon.
Large language models (LLMs) can memorize and reproduce training sequences verbatim, raising privacy and copyright concerns. Existing mitigation methods apply interventions uniformly, often degrading predictions at the many positions unrelated to verbatim recall. We study a token-level log-likelihood ratio between a target model and a domain-competent reference with lower exposure to the evaluated sequences. Exposure-controlled experiments support its use for localizing positions associated with sequence-level verbatim recall. Across evaluated settings, a minority of positions carries a disproportionate share of the positive signal, motivating selective activation correction. From a first-order surrogate of the memorization-reduction objective, we derive a probe-steer framework, which decomposes intervention into a probe that flags high-signal activations and a steer that applies a correction only when the probe exceeds a threshold. We propose Gated Subspace Steering (GSS) as its practical instantiation: under a background normalization constraint, the probe-steer pairs that are optimal for this surrogate follow in closed form from the leading singular vectors of a whitened gradient-weighted activation matrix. Across four benchmarks, GSS achieves memorization suppression comparable to strong baselines, with favorable utility trade-offs in several settings, while requiring 10-100$\times$ less compute than optimization-based alternatives. Code is available at https://github.com/ubc-tea/Mitigating-Memorization-Where-It-Happens.