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..
RNNs perform task computations by dynamically warping neural representations
Authors: Arthur Pellegrino, Angus Chadwick
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Here, we test this hypothesis, we develop a Riemannian geometric framework that enables the derivation of the manifold topology and geometry of a dynamical system from the manifold of its inputs. By characterising the time-varying geometry of RNNs, we show that dynamic warping is a fundamental feature of their computations. We trained an RNN on a classic contextual binary evidence integration task with dynamics: dx = (Wϕ(x) − x + b1u1 + b2u2 + c11ctx=1 + c21ctx=2)dt + B d W, x(0) = 0 We tested these hypotheses in RNNs trained to perform a sequential working memory task. |
| Researcher Affiliation | Academia | Arthur Pellegrino The Gatsby Unit University College London EMAIL Angus Chadwick School of Informatics University of Edinburgh EMAIL |
| Pseudocode | No | The paper describes mathematical derivations and experimental setups but does not include any explicit pseudocode blocks or algorithms in a structured format. |
| Open Source Code | Yes | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: The code necessary to reproduce all the figures is provided. |
| Open Datasets | No | The paper does not provide concrete access information (links, DOIs, specific repositories, or formal citations for public availability) for the datasets used in its main experiments. For Section S8.6, it cites a paper where data was recorded but does not explicitly state that the specific data used is publicly available nor provides a direct access link within this paper. |
| Dataset Splits | No | The paper primarily describes training RNNs on tasks by generating inputs based on parameters (e.g., varying u1, u2 or angles θ) rather than using pre-existing datasets with defined splits. For the deep neural network in S8.1, it mentions discretizing the input range: "We evaluated these integrals by discretising [0, 2π) into 200 even bins." For the other tasks (contextual decision-making, working memory, parametric working memory, BCI data), there is no explicit mention of training/validation/test splits, specific percentages, or sample counts. |
| Hardware Specification | Yes | Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: The paper discloses the compute resources used in supplementary materials (a single RTX 4070). |
| Software Dependencies | No | The paper mentions several software components and tools such as "Adam" (optimizer), "Optax" (library), "Heun" (SDE solver), "Jax" and "Pytorch" (auto-differentiation frameworks). However, it consistently lacks specific version numbers for these software dependencies, which is required for a reproducible description. |
| Experiment Setup | Yes | Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? Answer: [Yes] Justification: The core model used in each analysis are provided in the main manuscript, while all the details necessary to train the models, including all hyperparameters are provided in supplementary materials S8. |