Research map / Large language models / Language, safety and interpretability
Leveraging Attractor Dynamics in Spatial Navigation for Better Language Parsing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PHE-trinity, a brain-inspired prefrontal-hippocampal-entorhinal model with a velocity-driven modular continuous attractor grid network and separate structure/content streams, evaluated on SCAN language command parsing.
Problem
Understanding the shared computational mechanism between hippocampal-entorhinal spatial navigation and language processing, and whether attractor dynamics can improve systematic generalization and data efficiency in language command parsing.
Approach
A neural model that pretrains a modular continuous attractor (grid-like) network on path integration, then co-trains it with an LSTM PFC decoder and relational memory HC module on SCAN, with word-derived velocity inputs representing grammatical roles and factorized structure-content streams.
Key results
- PHE-trinity reaches 100.0% on SCAN Add Jump across 8 runs, versus 78.4 ± 27.4% for Syntactic Attention.
- PHE-trinity achieves 99.71 ± 0.24% on Add Jump with only 2000 training examples.
- On the SCAN Length split, PHE-trinity reaches 21.48 ± 3.4%, versus 0.0 ± 0.0% for MLC (reproduced) and 15.2 ± 0.7% for Syntactic Attention.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add spatial path-integration pretraining to the attractor dynamics (ablation) | PHE-trinity · SCAN Add Jump | accuracy: 75.24 ± 42.86 -> 100.0 ± 0.0 | Helped |
| add spatial path-integration pretraining to the attractor dynamics (ablation) | PHE-trinity · SCAN Length | accuracy: 19.4 ± 1.1 -> 21.48 ± 3.4 | Helped |
| add path-integration training to a model trained exclusively on SCAN (ablation) | PHE-trinity · SCAN Add Jump | accuracy: 74.96 ± 42.92 -> 75.24 ± 42.86 | No clear effect |
| add pre-established attractor dynamics through spatial path-integration pretraining (ablation) | PHE-trinity · SCAN tasks with limited training data | test accuracy: 99.71±0.24% with only 2000 training examples | Helped |
| increase the number of attractor modules in the MEC network (ablation) | predefined-attractor PHE-trinity · SCAN Add Jump | accuracy | Helped |
Surprising findings
- MLC reproduced by the authors scores 0.0% on the SCAN Length split, while it scores 99.78% on Add Jump.
Limitations
- Evaluated only on the simplified SCAN dataset, which captures limited aspects of human language.
- Lacks HC-to-MEC feedback connections and other biologically important features.
- Length-split accuracy remains low (21.48%) despite improvement.
Experimental setup
Datasets: SCAN spatial path integration task (synthetic)
Benchmarks: SCAN
Metrics: test accuracy
Baselines: Syntactic Attention MLC Control model 1 Control model 2
Builds on: Tolman-Eichenbaum Machine Syntactic Attention Vector-HaSH LSTM
Closest papers by meaning
- A simple model of co-emergence of grid and place fields (NeurIPS 2026 · similarity 0.87)
- REMI: Reconstructing Episodic Memory During Internally Driven Path Planning (NeurIPS 2025 · similarity 0.87)
- TopoLM: brain-like spatio-functional organization in a topographic language model (ICLR 2025 · similarity 0.86)
- Time Makes Space: Emergence of Place Fields in Networks Encoding Temporally Continuous Sensory Experiences (NeurIPS 2024 · similarity 0.86)
- Binding in hippocampal-entorhinal circuits enables compositionality in cognitive maps (NeurIPS 2024 · similarity 0.86)
- Global Distortions from Local Rewards: Neural Coding Strategies in Path-Integrating Neural Systems (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.