Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Scaling Up Active Testing to Large Language Models
Authors: Gabrielle Berrada, Jannik Kossen, Freddie Bickford Smith, Muhammed Razzak, Yarin Gal, Thomas Rainforth
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 Experiments We now seek to empirically assess our proposed approach for active testing of LLMs (Section 3), as well as our proposed risk-error-estimation method (Section 4). We provide implementation details in Appendix B and code at github.com/gabrielleberrada/scaling-up-active-testing. |
| Researcher Affiliation | Academia | 1 OATML, Department of Computer Science, University of Oxford 2 Department of Statistics, University of Oxford |
| Pseudocode | Yes | Algorithm 1 Sampling-based active testing input Target model, f; loss function, ℓ; acquisition function, a; training set, Dtrain; pool set, Dpool; label budget, M 1: Compute f(xj) for all xj Dpool 2: Set Dtest = 3: for m (1, 2, . . . , M) do 4: Train πm (e.g., on Dtrain Dtest) 5: Compute am(xj) for all xj Dpool 6: Sample im qm(i) 7: Sample yim peval(y|xim) 8: Set Dtest Dtest {(xim, yim)} 9: end for 10: Compute ˆRLURE output Risk estimate, ˆRLURE |
| Open Source Code | Yes | We provide implementation details in Appendix B and code at github.com/gabrielleberrada/scaling-up-active-testing. |
| Open Datasets | Yes | Our experiments are based on five datasets: Stanford Sentiment Treebank 2 (SST2; Socher et al, 2013; unknown license), Subjectivity (Subj; Pang & Lee, 2004; Creative Commons Attribution 4.0 International License), Financial Phrase-bank (FPB; Malo et al, 2013; Creative Commons Attribution Non Commercial Share Alike 3.0 Unported License), Hate Speech (HS; De Gibert et al, 2018; Creative Commons Attribution-Share Alike 3.0 Spain License) and Massive Multitask Language Understanding (MMLU; Hendrycks et al, 2021; MIT license). |
| Dataset Splits | Yes | Table 1 Pool-set and test-set sizes for each dataset. Sets are randomly sampled so that they are disjoint. ... We randomly split the 10,000-sentence train set into a 6,000-sentence pool set and a 4,000-sentence test set. |
| Hardware Specification | No | Generating model outputs over the pool set is the main computational cost, typically requiring two GPUs for Llama 2 70B and one GPU for all other models used in this work. |
| Software Dependencies | No | The paper does not explicitly mention specific version numbers for software dependencies like Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | All few-shot models are given 50 in-context examples for all datasets, except for the MMLU dataset, for which models receive only 5 examples due to token limit. Examples are ordered randomly once and fixed for all evaluations, ensuring that all models receive the same context. The set of in-context examples is therefore fixed beforehand and is common to both the target model and the surrogate model, for label efficiency and fair comparison. ... To obtain deterministic, reproducible token generation, we set the maximum number of tokens to 1, do not set a top-k nor top-p value and output the logits directly, from which we compute probabilities. ... All acquisition probabilities below α = 0.1 times the probability corresponding to uniform-random acquisition are brought up to this limit value. |