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..
Universal Sequence Preconditioning
Authors: Annie Marsden, Elad Hazan
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We empirically validate that convolutional preconditioning with Chebyshev or Legendre coefficients yields significant online regret improvements across various learning algorithms and data types. |
| Researcher Affiliation | Collaboration | *Google Deepmind Princeton University |
| Pseudocode | Yes | Algorithm 1 General Sequence Preconditioning (Offline Version) |
| Open Source Code | No | A link to the code will be provided in the final version. However, we omit the Git Hub repository with all of our code at the moment in order not to violated the double blind policy and keep the paper anonymous. |
| Open Datasets | Yes | To evaluate whether our proposed preconditioning approach generalizes to real-world time series, we conduct experiments on the well-established ETTh1 dataset from the Electricity Transformer Temperature (ETT) benchmark [37]. |
| Dataset Splits | No | We generate N = 200 sequences of length T = 2000 via three mechanisms: (i) a noisy linear dynamical system, (ii) a noisy nonlinear dynamical system, and (iii) a noisy deep RNN. [...] We set the horizon to be T = 5000 and we sweep over a broader range of learning rates ϖ {10 j}j=0,1,2,3,4,5. As before we consider (i) no preconditioning (baseline), (ii) fixed Chebyshev coefficients, (iii) fixed Legendre coefficients, and (iv) coefficients learned jointly with model parameters. |
| Hardware Specification | No | The paper does not explicitly provide specific hardware details such as GPU/CPU models, memory, or compute worker types used for experiments. The NeurIPS checklist provides an affirmative answer but an empty justification for this question, indicating a lack of detail. |
| Software Dependencies | No | The paper mentions software components like 'LSTM' and 'Adam optimizer' but does not provide specific software dependencies with version numbers (e.g., 'Python 3.8, PyTorch 1.9, and CUDA 11.1'). |
| Experiment Setup | Yes | We test polynomial degrees n {2, 5, 10, 20}. To ensure fair comparison, for each algorithm and conditioning c variant we perform a grid search over learning rates ϖ {10 3, 10 2, 10 1}, selecting the one minimizing average regret across the N sequences. In the case of the learned coefficients, we sweep over the 9 pairs of learning rates (ϖmodel, ϖcoefficients) {10 3, 10 2, 10 1} {10 3, 10 2, 10 1}. [...] a 10-layer LSTM with hidden dimension 100 per layer using the Adam optimizer. We set the horizon to be T = 5000 and we sweep over a broader range of learning rates ϖ {10 j}j=0,1,2,3,4,5. |