Anonymous submission · nss
One program. One scene.
Channel, network, and radar, all coherent.
nss is an integrated sensing and communication (ISAC) simulator. An LLM writes a scene program over a closed vocabulary of zones, roles and itineraries; a compiler supplies every quantity from data and builds the venue. From that one scene, the same bodies drive one ray-traced channel, a 5G network, and a radar that samples on the network's own transmissions.
- 7venue types, from homes to an airport
- 441people in the generated airport, each with a phone
- 0sizes, speeds or durations written by the LLM
- 7–67×faster fluid network model that ranks controllers like ns-3
- 14 minfrom prompt to network KPIs for a 60 s mall scene
01Blockage, scrubbed
See the column cut the link, and the radar see it coming.
One traveler (UE2) walks through the airport. Around 14.2 s a column cuts her line of sight to AP9, while AP1 stays clear. One slider drives the video, the sensed track, the forecast, and the SNR of both APs.
- true track
- sensed track
- forecast
- SNR AP9 AP1
- AP link no LOS
02Scene gallery
Every venue is a program you can read.
Hover a card to play the scene; click it to flip and read the program the LLM wrote, with what the compiler made of it. The programs ship with the release, so every scene rebuilds without the LLM.
03Get started
From a prompt to network KPIs.
Install, compile a shipped program, and score a controller. Shipped programs replay without an LLM.
- Simulatorscene compiler, body model, channel, network and sensing stages behind typed interfaceslink TBD
- Programsthe venue programs and household screenplays from the paper, each from a single LLM calllink TBD
- Scenescompiled scenes with people, phones, traffic and traced channelslink TBD
- Control harness100 ms controller loop, tuned baselines, genie bound and counterfactual feedslink TBD
@inproceedings{anonymous2026nss,
title = {nss: Symbolic Programs for Integrated Sensing and Communication Simulation},
author = {Anonymous Author(s)},
booktitle = {Under review},
year = {2026}
}