Multivariate Stochastic Dominance via Optimal Transport and Applications to Models Benchmarking

Authors: Gabriel Rioux, Apoorva Nitsure, Mattia Rigotti, Kristjan Greenewald, Youssef Mroueh

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We showcase our method in comparing and benchmarking Large Language Models that are evaluated on multiple metrics.
Researcher Affiliation Collaboration Gabriel Rioux Center for Applied Mathematics Cornell University Apoorva Nitsure MIT-IBM Watson AI Lab IBM Research Mattia Rigotti MIT-IBM Watson AI Lab IBM Research Kristjan Greenewald MIT-IBM Watson AI Lab IBM Research Youssef Mroueh MIT-IBM Watson AI Lab IBM Research
Pseudocode Yes A Algorithm Algorithm 1 Multivariate Stochastic Order Multi-testing (relative and absolute)
Open Source Code Yes Code for these experiments is available at https://github.com/IBM/stochastic-order-eval.
Open Datasets Yes For our first evaluation we use the dataset from Jiang et al. [2023] (MIT license) that consists of responses from 12 different instruction following LLMs
Dataset Splits No The data has a train (100K rows) and test (5k rows) split where each row consists of an instruction, input sentence, the expected output from users, as well as the responses of a set of different LLMs with their decoding parameters and evaluation scores on different metrics.
Hardware Specification Yes All experiments were run on NVIDIA A100 80GB GPUs
Software Dependencies Yes All experiments were run on NVIDIA A100 80GB GPUs using Py Torch [Ansel et al., 2024] (v.2.3.0, BSD-3 license) and the Python Optimal Transport package [Flamary et al., 2021] (v.0.9.3, MIT license)
Experiment Setup Yes We then compute the pairwise ratios for these empirical distributions using the logistic loss with β = 0.2, the regularization parameter λ = 0.1, and utilize the relative testing procedure from Section 4.2 to rank the 12 LLMS