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..
REMI: Reconstructing Episodic Memory During Internally Driven Path Planning
Authors: Zhaoze Wang, Genela Morris, Dori Derdikman, Pratik Chaudhari, Vijay Balasubramanian
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Here, we propose a system-level theory of MEC-HC wiring that explains how grid and place cell patterns could be connected to enable cue-triggered goal retrieval, path planning, and reconstruction of sensory experience along planned routes. We show analytically that grid-based planning permits shortcuts through unvisited locations and generalizes local transitions to long-range paths. During planning, intermediate grid states trigger place cell pattern completion, reconstructing sensory experiences along the route. Using a single-layer RNN modeling the HC-MEC loop with a planning subnetwork, we demonstrate these effects in both biologically grounded navigation simulations using Ratatou Gym and visually realistic navigation tasks using Habitat Sim. Codes for experiments, simulations, and vision encoder are available at 1,2,3. |
| Researcher Affiliation | Academia | 1Dept. of Electrical and Systems Eng., Univ. of Pennsylvania 2Tel Aviv Sourasky Medical Center 3Gray Faculty of Medical and Health Sciences, Tel Aviv University 4Rappaport Faculty of Medicine, Technion Israel Institute of Technology 5Dept. of Physics, Univ. of Pennsylvania 6Santa Fe Institute 7Rudolf Peierls Centre for Theoretical Physics, University of Oxford |
| Pseudocode | No | The paper includes mathematical formulations like Equation (1) for RNN dynamics and logical steps described in prose for planning (e.g., 'If the animal selects the phase with the highest value after each update and renormalizes the vector, this process traces a smooth trajectory toward the target'), but it does not present these as clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Codes for experiments, simulations, and vision encoder are available at 1,2,3. 1Project Page: https://zhaozewang.github.io/remi 2Ratatou Gym (Simulation Suite): https://ratatougym.github.io 3Bottleneck MAE: https://github.com/grasp-lyrl/btnk mae |
| Open Datasets | Yes | To this end, we replaced the simulated SMC rate maps Rsmc RLx Ly N with visual feature embeddings extracted from the Habitat Synthetic Scenes Dataset (HSSD) [66] in the photorealistic simulator Habitat-Sim [45 47]. ... Pretraining Btnk MAE on Image Net 1k. |
| Dataset Splits | Yes | To verify that the network indeed learned path integration, we first trained a model containing only grid, speed, and head direction cells (Figure 1b). We train this model on short random trajectories (5 s) but accurately path-integrate over significantly longer trajectories (10 s, 120 s) during testing (Figure 1c). ... During training, we generate random 1-second trajectories to sample pairs of current and target locations... For testing, we generate longer 10 and 20 second trajectories to define start and target locations... |
| Hardware Specification | Yes | Yes, all our experiments were conducted on a single machine equipped with an RTX 5090 GPU with 36GB of VRAM. |
| Software Dependencies | No | We use the FAISS library [57, 58]. Specifically, we employ the Index IVFFlat structure... All other code and data were developed by us. Additional packages used in our implementation are standard Python libraries from Py PI, all of which are open-source. The paper mentions the FAISS library but does not provide a specific version number. It also refers to 'standard Python libraries from Py PI' without listing specific libraries and their versions. |
| Experiment Setup | Yes | Specifically, consider a standard RNN that updates its dynamics as: zt+1 = α zt + (1 α) Winut + Wrecf(zt) (1) where z Rdz is the hidden state, u is the input, α is the forgetting rate, and Win, Wrec are the input and recurrent weight matrices... We set α as a learnable vector in Rdz+d I+d O to allow different cells to have distinct forgetting rates... To model this, we trained nine identical networks differing only in masking ratio rmask (0.1 to 0.9), which specifies the maximum fraction of head direction, speed, GC, and SMC inputs and initial hidden states randomly set to zero during training... The planning subnetwork is trained to minimize the mean squared error between the current and target GC states for all timesteps. |