A future wireless surface may do more than bounce a signal towards a phone. By changing how its elements reflect energy over time, it can preserve a clean constellation in one direction and scramble the same transmission elsewhere. The problem is choosing the settings. In the system studied by researchers at Rutgers and Princeton, one representative parameter space contained about 8^72 possibilities—roughly 10^65 configurations.[1]
Exhaustive search is not merely slow at that scale; it is meaningless as an engineering plan. The researchers used a generative flow network, or GFlowNet, to learn a probability distribution that favoured high-reward configurations. Instead of trying to identify one permanent optimum, the system learned to sample several strong designs, which can be useful when changing the pattern itself makes interception harder.
The surface is solving three jobs at once
One configuration, three objectives
| Objective | Desired behaviour | Failure to avoid |
|---|---|---|
| Communication | Keep the intended users’ QPSK constellation readable | Scrambling the legitimate signal or reducing its rateSource 1 |
| Sensing | Direct useful beam-pattern gain towards the target area | Losing the radar-like information the system needsSource 1 |
| Physical-layer security | Scramble symbols in unintended directions | Allowing an eavesdropper to reconstruct the transmitted constellationSource 1 |
These objectives can conflict. Steering energy towards a target can also make that direction informative to an eavesdropper. A setting that protects one user may perform poorly when several users occupy different angles. The reward function therefore combined communication sum rate with sensing and secrecy constraints, turning each complete surface configuration into a terminal state with a measurable score.[1]
The simulated system
| Measure | Value | Context |
|---|---|---|
| Base-station antennas | 8 | A uniform linear array feeding the reflecting surface.Source 1 |
| Reflecting elements | 6 × 6, or 36 | The time and phase controls created 72 discrete parameter choices in the highlighted training example.Source 1 |
| Waveform | 16 subcarriers and 1,024 OFDM symbols | QPSK modulation at a 24 GHz centre frequency.Source 1 |
| Training trajectories | 1,000,000 | Fewer than 0.000001% of the roughly 10^65 configurations.Source 1 |
| Training hardware | NVIDIA A100 and Apple M3 Max | Training was offline; online use requires much less computation.Source 1 |
Why a distribution can be more useful than one winner
A conventional optimiser often tries to converge on the single setting with the highest expected reward. A GFlowNet instead assigns more sampling probability to higher-reward terminal states. The distinction matters when many good configurations exist. The network can switch among them, reducing the value of an adversary learning one static pattern and offering alternatives when users or channel conditions change.[1]
In the simulations, the intended directions retained low symbol error while the target region received scrambled constellations. The approach also produced stronger low-signal performance than a rule-based comparison and higher worst-case secrecy rates than simulated annealing or uniform random search in the tested multi-user settings. The authors additionally modelled phase errors and other impairments, including levels larger than typical hardware errors.[1]
The number 10^65 is not a deployment result
The dramatic search-space figure describes how many discrete combinations the example formulation permits. It does not mean the model examined or certified them all. Generalisation from one million trajectories is the claim under test: the neural policy uses patterns learned from visited states to place probability on promising unvisited configurations.
The contribution is nevertheless concrete. It shows how generative modelling can be useful when engineering needs a portfolio of good answers rather than one generated object or one deterministic optimum. In that setting, “generation” means learning where valuable configurations live inside a combinatorial space—and being able to draw another one when the radio environment changes.
Sources & further reading
- Time-modulated intelligent reflecting surfaces for integrated sensing, communication and security with generative artificial intelligence design npj Wireless Technology · 16 Oct 2026 · Accessed 16 Oct 2026