PiCO: Peer Review in LLMs based on Consistency Optimization

Authors: Kun-Peng Ning, Shuo Yang, Yuyang Liu, Jia-Yu Yao, Zhenhui Liu, Yonghong Tian, Yibing Song, Yuan Li

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform experiments on multiple datasets with standard rank-based metrics, validating the effectiveness of the proposed approach. We perform experiments on multiple crowdsourcing datasets with standard rank-based metrics, the results demonstrate that the proposed Pi CO framework can effectively obtain a large language models leaderboard closer to human preferences.
Researcher Affiliation Academia 1School of Electrical and Computer Engineering, Peking University 2Peng Cheng Laboratory EMAIL, EMAIL, EMAIL
Pseudocode Yes Algorithm 1 Overall Framework Algorithm of Peer Review
Open Source Code Yes Our code is released at https://github.com/PKU-YuanGroup/PiCO.
Open Datasets Yes To validate the effectiveness of the proposed approach, we perform experiments on Chatbot Arena (Zheng et al., 2023), MT-Bench (Zheng et al., 2023), and Alpaca Eval (Li et al., 2023b).
Dataset Splits Yes The ratios of response sets D are 1, 0.7, and 0.4, respectively.
Hardware Specification No The paper does not provide specific hardware details used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details (e.g., library or solver names with version numbers) needed to replicate the experiment.
Experiment Setup Yes Initialize model weights vector w with Gaussian distribution, In the framework of the Elo mechanism, as specified by Equation 16, the BASE value is set to 10, and the SCALE factor is determined to be 400., in the context of the Rank mechanism, as outlined by Equation 17, rank(j) signifies the current ranking of model j, with the constant K assigned a value of 200., k is a hyper-parameter recommended to be set to 3 to 7, and we set k = 3 in this paper., it iteratively removes the lowest-scoring LLM from the reviewer queue for the next consistency optimization stage, until 60% of models are eliminated.