NeurIPS 2026 spotlight
Shivang Rawat*, Mirko Morello, Flaviano Morone, David J. Heeger
TelepathNew York University
*shivang@telepath.com
Each layer is a leaky integrator driven by the layer above it.
The look-ahead undoes the leaky integration.
Each layer, as before
With look-ahead on each layer’s input
Prospective firing: cortex, retina, cerebellum, muscle spindles
Energy of one unit
A unit balances the bottom-up input against an internal prediction. γ sets the trade-off.
Descend the energy
A leaky integrator relaxing toward a blend of input and prediction.
Heeger (2017), Theory of cortical function, PNAS
Simplest model of an oscillation
Exact rotation over one time constant
Look-ahead: the first-order version of that rotation
Internal-model principle (Francis and Wonham, 1976)
One constraint: each channel’s input gain is tied to its own decay rate, α = γ/τ.
The RQF
Make the input drive prospective
One discretized step, instantaneous input
With prospective input
# one step of an RQF or any diagonal SSM y = A_bar * y + B_bar * z[t] y = A_bar * y + B_plus * z[t] + B_minus * z_prev z_prev = z[t] # B_plus = B_bar + γ·A_bar, B_minus = −γ·A_bar
Raw-audio Speech Commands with 31.9k parameters. No residuals, gating, or normalization.
Path-X, 16,384 steps with a minimal six-layer stack. 86.3% with a selective RQF.
Selective Copying with a selective RQF. Look-ahead: 5 of 5 seeds; without it, 2 of 5.