Skip to main content
Living Intelligence

Continual learning research

Living Intelligence Lab

Building biologically inspired AI systems that learn continuously about the world.

Read our researchGet in touch

Publications & Preprints

Selected Research

Exploring biologically inspired architectural principles for building continuous, dynamic learning platforms and systems.

Model RoutingMulti-LLM SystemsBio-Inspired Router
arXiv:2308.11601

Tryage: Real-time, intelligent Routing of User Prompts to Large Language Models

Surya Narayanan Hari, Matt Thomson

Inspired by the thalamic router in the brain, Tryage employs a perceptive router to predict downstream model performance on prompts and make routing decisions using an objective function that integrates performance predictions with user goals and constraints (such as model size, recency, security, verbosity, and readability).

Abstract →View PDF 📄
Prompt EngineeringControl TheoryLLM Controllability
arXiv:2310.04444

What's the Magic Word? A Control Theory of LLM Prompting

Aman Bhargava, Cameron Witkowski, Shi-Zhuo Looi, Matt Thomson

Formalizes LLM systems as discrete stochastic dynamical systems and studies prompting through control theory. Prompts of 10 tokens or fewer make the "correct" next token reachable at least 97% of the time, and can turn even the least likely tokens into the most likely ones.

Abstract →View PDF 📄
Predictive CodingCognitive MapsSpatial Navigation
Nat. Mach. Intell. 2024

Automated construction of cognitive maps with visual predictive coding

James Gornet, Matt Thomson

An agent navigating a virtual environment learns to predict its next visual frame and, in doing so, builds an internal map that reflects spatial distances, letting it locate itself relative to landmarks from vision alone. Proposes predictive coding as a unified algorithm for building cognitive maps.

Abstract →View PDF 📄
Continual LearningDifferential GeometryNetwork Adaptation
Nat. Mach. Intell. 2024

Engineering flexible machine learning systems by traversing functionally invariant paths

Guruprasad Raghavan, Bahey Tharwat, Surya Narayanan Hari, Dhruvil Satani, Rex Liu, Matt Thomson

Treats a network's weight space as a curved Riemannian manifold and adapts trained models by moving along geodesic paths that preserve prior knowledge. Matches or exceeds methods such as LoRA on continual learning, sparsification and adversarial robustness for BERT, vision transformers and CNNs.

Abstract →View PDF 📄

Learning

Systems that keep learning

Models that absorb new experience after deployment, without retraining from scratch or overwriting what they already know.

Architecture

Biology as blueprint

Memory, plasticity and consolidation borrowed from living systems, then rebuilt as architecture rather than metaphor.

Practice

Grounded in the world

Research measured against changing, open environments — the conditions real deployments actually meet.