Stratified Prediction-Powered Inference for Effective Hybrid Evaluation of Language Models

Authors: Adam Fisch, Joshua Maynez, R. Hofer, Bhuwan Dhingra, Amir Globerson, William W. Cohen

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

Reproducibility Variable Result LLM Response
Research Type Experimental We compare our stratified estimator, Strat PPI, to two baselines: (i) the classical estimate, which uses only the labeled data, Sn; and (ii) PPI++, which uses both Sn and Sn. All of our experiments focus on 1-d mean estimation. We explore three different allocation strategies for Strat PPI: the first is to set ρk = wk to be data proportional (Strat PPI Prop.), the second is to set ρk optimally via the oracle ρk = ρ k (Strat PPI Opt.), and the third is to use the approximation, ρk wkˆσk, in Example 2 for λk = 1 when confidence scores are available (Strat PPI Heur.). We use λ-tuning for both PPI++ and Strat PPI, as outlined in 4.2. Additional experimental results are given in Appendix C.
Researcher Affiliation Industry Adam Fisch , Joshua Maynez , R. Alex Hofer Bhuwan Dhingra Amir Globerson William W. Cohen Google Deep Mind Google Research {fisch,joshuahm,rofer,bdhingra,amirg,wcohen}@google.com
Pseudocode Yes Algorithm 1 Stratified prediction-powered inference for general M-estimators (Strat PPI)
Open Source Code No Code may be made available at a future date.
Open Datasets Yes Seahorse. The Seahorse dataset [11] focuses on multilingual summarization.
Dataset Splits Yes For each experiment, we sample N = 10,000 total predictions f( X) using ρ1 = ρ2 = 0.5, i.e., proportional to masses of the two hypothetical, equal-weight strata. We then vary the total number n of labeled examples Y , where the allocation is chosen according to ρ (which differs depending on if we are using Strat PPI Prop. or Strat PPI Opt.).
Hardware Specification No Compute resources required are very light, as no model training is performed.
Software Dependencies No The paper does not specify version numbers for any software, libraries, or frameworks used in the experiments.
Experiment Setup Yes We assume that predictions are formed as f(Xik) = Yik + µk + σkϵik, where ϵik N(0, 1). ... We test three different scenarios: (i) where the two strata are homogeneous with µ1 = µ2 and σ1 = σ2; (ii) where the two strata have different prediction biases, µ1 = µ2; and (iii) where the two strata have different prediction noise levels, σ1 = σ2.