In massive Internet-of-Things uplinks, devices transmit without first asking permission ("grant-free" access). Each time slot, the base station must work out from one noisy superposition of signals which of hundreds of devices spoke, what their radio channels look like, and what data they sent. This combined problem is called joint activity and data detection. The best receivers solve it with iterative message passing, a family of algorithms that repeatedly exchange probability estimates between variables. They spend the same effort every slot and on every device, even though traffic is highly repetitive. A device silent in the last slot is very likely silent now, and a device whose codeword already passes its parity checks needs no further work.
Video transformers have learned to exploit exactly this kind of repetition. They merge or prune tokens that did not change between frames, and the careful versions keep pruned tokens in a "recycle bin" so they can be restored when needed. We propose to carry that pattern into the receiver, where it acquires one advantage the language-model compute-allocation literature lacks: a free, label-less verifier, the error-correcting code's parity check. The receiver will keep each potential user in one of three states. A user can be frozen (confidently inactive and carried over from the last slot, then re-tested cheaply), live (inside the full message-passing loop), or decoded (parity checks pass, so its signal is subtracted and it leaves the loop). The claim to test is that shrinking the problem per user in this way saves substantially more compute than the obvious fix of warm-starting from the previous slot and stopping early. It should do so without raising the miss rate on newly waking devices.
Core Claim
Existing receivers for correlated grant-free traffic already use temporal memory, but only to sharpen priors. Cooperative Multi-Cell Massive Access With Temporally Correlated Activity builds a sliding-window message-passing algorithm (DCS-MMV-GAMP) with Markov activity transitions. It still runs a fixed number of iterations over every potential user in every slot, and it does not decode payloads. Turbo receivers such as Joint Activity Detection and Data Decoding in Massive Random Access via a Turbo Receiver and Joint Activity Detection, Channel Estimation, and Data Decoding for Grant-free Massive Random Access decode payloads jointly with detection, but they treat slots as independent and use fixed iteration schedules. On the adaptive-compute side, Learning How Hard to Think shows that predicting marginal gain and reallocating compute can halve cost at equal accuracy. Its stated weakness is that roughly 3,000 labeled queries with ground-truth rewards are needed to train the predictor.
Our claim has two parts. First, in a Bayesian turbo receiver with temporal activity priors, the dominant cost is carried by users whose state is already nearly certain. The cost of one message-passing iteration grows with pilot length times antenna count times the number of users in the loop. Removing frozen and decoded users therefore cuts cost multiplicatively across all remaining iterations, which a stopping rule alone cannot do. Second, a DyCoke-style re-test keeps this reversible, borrowing the recycle-bin idea from DyCoke. The re-test is a single correlation of the current residual against the frozen users' pilots, and a frozen user is restored to the live set if the correlation shows evidence of activity. This guards against the irreversible loss that DyCoke documents for one-shot pruning.
Research Program
The base receiver is a BiG-AMP turbo receiver with an LDPC belief-propagation decoder, following Joint Activity Detection and Data Decoding in Massive Random Access via a Turbo Receiver. We extend it with Markov activity priors in the manner of DCS-MMV-GAMP, so that the posterior activity probability from slot $t-1$ becomes the prior at slot $t$. The paper uses the previous work in three distinct ways. The turbo papers supply the method. DCS-MMV-GAMP supplies the temporal prior and serves as the main competitor. DyCoke and STTS motivate the design: STTS's fixed retention quota is the video analogue of fixed iteration counts, and DyCoke's recycle bin is the analogue of our re-test.
Three mechanisms are built on top of the base receiver.
The first mechanism freezes users at the start of each slot. A user enters the slot frozen if its predicted activity probability is below a threshold $\tau_f$. At slot start, one matched-filter pass computes a test statistic for every frozen user against the pilot residual. Users exceeding a restore threshold $\tau_r$ join the live set. We choose $\tau_r$ to hit a target miss rate on newly waking devices, estimated on held-out simulations, and this rate is reported directly.
The second mechanism removes users by parity check. After each turbo outer iteration, any user whose hard decisions satisfy all LDPC parity checks is marked decoded. Its reconstructed contribution is subtracted from the received signal, and it leaves the loop. This is the code-as-verifier replacement for the learned reward predictors of the language-model literature, and it requires no labels.
The third mechanism allocates effort across slots. Under a compute budget defined over a window of slots, a small controller decides how many further outer iterations to spend on the current slot. Its inputs are the live-set size, the mean posterior entropy of live users, and the fraction already decoded. We compare two versions of this controller, taking both from Learning How Hard to Think. One is online greedy allocation by predicted marginal gain. The other is offline binning of slots into easy, medium and hard, which that paper found more robust to predictor noise. The controller's training targets are the parity-check outcomes the receiver observes anyway, so it can be fit self-supervised from simulation or live logs.
Compute is counted in multiply-accumulate operations, including the re-test pass and the subtraction. We report both the mean and the 99th-percentile per-slot cost, because base-station latency is set by the tail.
Baselines
The naive reference point is the full turbo receiver run per slot with a fixed iteration schedule and no temporal prior, as in Joint Activity Detection and Data Decoding in Massive Random Access via a Turbo Receiver. That paper reports that at BLER $10^{-3}$ it supports about 75% more active users than a separate design and closely approaches the known-activity bound. The companion paper Joint Activity Detection, Channel Estimation, and Data Decoding for Grant-free Massive Random Access reports 40 versus 20 supported users at BLER $10^{-3}$. It also reports that its SI-aided receiver saves over 60% of execution time at a capacity cost. We include that SI-aided receiver as the "cheap fixed pipeline" reference.
The strongest existing temporal method is DCS-MMV-GAMP from Cooperative Multi-Cell Massive Access With Temporally Correlated Activity. The file gives no reported number for it on Lacuna beyond qualitative gains in error detection ratio and NMSE over single-frame benchmarks, so we reimplement it. To make it comparable on data decoding, we attach the same LDPC turbo loop.
The strongest simple competitor, and the one we consider the real threat, is the turbo receiver with the temporal prior, warm-started from the previous slot, stopped per slot as soon as all users' syndromes pass or a residual-change threshold is met. We call this warm-start plus syndrome-stop. It captures temporal redundancy and verifier-based halting without any per-user freezing.
We also include DU-JAD from Deep-Unfolded Joint Activity and Data Detection for Grant-Free Transmission in Cell-Free Systems as the fixed-small-budget learned alternative. It reports an average symbol error rate 0.129 times that of its iterative baselines at 60 access points, using 10 layers.
All methods share the same simulated channels, pilots, codes and seeds. Every tunable threshold or iteration cap, for baselines and for our method alike, gets the same grid-search budget on a shared validation set. Comparisons are made on the same compute-versus-capacity curves.
Ablations remove freezing, parity-check removal, the re-test, and the cross-slot controller one at a time. The ablation without the re-test isolates whether reversibility matters.
Our method beats a baseline only if both conditions hold at matched supported users under BLER $\le 10^{-3}$. It must reach that capacity at clearly lower mean and 99th-percentile compute. Its new-arrival miss rate must also be no worse than that baseline's.
Why not just…?
Why not just warm-start from the last slot and stop when all syndromes pass? This is the most plausible way the proposal could fail, and it is our primary comparison rather than a straw man. Its limitation is structural. It cuts the number of iterations, but every remaining iteration still touches all potential users, and the slowest user in the slot determines when it stops. Freezing and per-user removal attack the per-iteration dimension instead. If the cost-versus-capacity curves of the two coincide across activity-persistence levels, the three-state machinery is not worth it, and we will report that.
Why not just gate on the change in the received covariance between slots? Massive Connectivity With Massive MIMO - Part I shows that user activity is recoverable from covariance because fast fading averages out over many antennas. At moderate antenna counts, however, slot-to-slot covariance differences mix fading fluctuations with activity changes. A single global gate would also mask one weak newly active user under many stable ones. We run this gate as an additional baseline. The new-arrival miss rate, stratified by the new user's path loss, will show whether it misses weak arrivals.
Why not just use a fixed small unfolded network like DU-JAD? Its authors note that DU-JAD's activity detection varies with the number of access points. A fixed depth also cannot scale up when load bursts. We test this directly with bursty Markov traffic and report 99th-percentile error during bursts alongside mean cost.
Proposed Experiments
The simulation follows the setup of the turbo-receiver paper: 500 m cell, 200 potential users, 64 base-station antennas, QPSK with a rate-1/2 LDPC code, 23 dBm transmit power, and −169 dBm/Hz noise density. We add two-state Markov activity with persistence swept from memoryless up to 0.99, and channel aging swept by Doppler from quasi-static to high mobility. We implement it in Sionna, using the differentiable simulation stack described in PyJama. Metrics are user detection error rate, channel NMSE, BLER, supported active users at BLER $10^{-3}$, the new-arrival miss rate, and mean and 99th-percentile compute per slot.
Three checks are decisive. First, at persistence 0.9 or above, our receiver should match the capacity of warm-start plus syndrome-stop at substantially lower compute. The gap should grow with persistence and vanish for memoryless traffic; if it does not vanish there, the savings come from something other than temporal redundancy, and the measurement is suspect. Second, removing the re-test should raise the new-arrival miss rate measurably, which shows reversibility is load-bearing. Third, during injected activity bursts, the cross-slot controller should hold BLER while the fixed-budget receivers show error spikes. As an offline-versus-online controller comparison, we test whether the binning result of Learning How Hard to Think transfers to a setting where rewards come from parity checks rather than labels.
References
- Cooperative Multi-Cell Massive Access With Temporally Correlated Activity
- Joint Activity Detection and Data Decoding in Massive Random Access via a Turbo Receiver
- Joint Activity Detection, Channel Estimation, and Data Decoding for Grant-Free Massive Random Access
- Learning How Hard to Think: Input-Adaptive Allocation of LM Computation
- DyCoke: Dynamic Compression of Tokens for Fast Video Large Language Models
- Efficient Video Transformers with Spatial-Temporal Token Selection
- Deep-Unfolded Joint Activity and Data Detection for Grant-Free Transmission in Cell-Free Systems
- Massive Connectivity with Massive MIMO-Part I: Device Activity Detection and Channel Estimation
- PyJama: Differentiable Jamming and Anti-Jamming with NVIDIA Sionna