Liquid Neural Networks (LNNs), also known as Liquid Time-Constant (LTC) Networks, are a class of continuous-time neural network models developed at MIT CSAIL that use nonlinear ordinary differential equations (ODEs) with adaptive time constants to enable real-time learning and adaptation after training.¹⁾ Inspired by the nervous system of the nematode C. elegans, LNNs employ fewer but more expressive neurons, achieving robust performance on time-series and control tasks with dramatically smaller model sizes than conventional architectures.²⁾

Traditional neural networks — including LSTMs, GRUs, and Transformers — operate in discrete time steps with fixed weights after training. Once deployed, their behavior is static; they cannot adapt to distribution shifts or novel conditions without retraining.

Liquid Neural Networks address this limitation by modeling neurons as continuous-time dynamical systems whose parameters vary based on input. The “liquid” metaphor reflects how the network's behavior flows and adapts in response to incoming data, much like a liquid conforming to its container.

Each neuron in an LNN is governed by an ODE of the form:

dx/dt = -tau(x(t), u(t)) * x(t) + B * u(t)

where:

• x(t) is the neuron's hidden state
• u(t) is the input signal
• tau is the liquid time constant — a learned, input-dependent function that controls the neuron's reaction speed
• B represents input sensitivity coefficients

The time constant tau varies dynamically based on both the current state and input, enabling each neuron to react at different speeds to different stimuli.³⁾ The system is solved using numerical ODE integration, producing stable, bounded outputs.

Coefficients A (internal coupling), B (input sensitivity), and C (output mapping) form a graphical ODE model that supports feedback loops and causal reasoning — unlike static architectures or Bayesian networks.

• Ramin Hasani — postdoctoral associate at MIT CSAIL (now Principal AI Scientist at Vanguard and CSAIL affiliate), lead author of the foundational LTC paper⁴⁾
• Daniela Rus — Director of MIT CSAIL, co-author and advocate for LNNs in safety-critical robotics⁵⁾
• Mathias Lechner — co-author, contributed to the Closed-form Continuous-time (CfC) extension
• Alexander Amini — co-author, MIT CSAIL researcher

A landmark demonstration showed a 19-neuron LNN successfully navigating a car through varying conditions including rain and noise, outperforming networks with orders of magnitude more parameters.⁶⁾

LNNs applied to drone flight navigation via imitation learning demonstrated robust handling of out-of-distribution conditions, selectively attending to relevant visual features while ignoring irrelevant ones.⁷⁾

Time-series analysis for brain activity monitoring, cardiac signal processing, and patient state prediction — domains where adaptability and interpretability are critical for safety.

Stock price prediction and risk assessment, leveraging the network's ability to handle non-stationary time series with shifting dynamics.

In 2022, the MIT CSAIL team introduced CfC models, which approximate the neuron ODEs without requiring numerical solvers during inference. This makes LNNs significantly faster while retaining their adaptive and causal properties.⁸⁾ CfC models are particularly suited for safety-critical applications like brain and heart monitoring, where trustworthiness and computational efficiency are equally important.

• Neural ODEs
• State Space Models (SSMs)
• Continual Learning Frameworks