5G multicast is the part of the mobile standard nobody markets, and on 9 September 2026 it got a rare turn in the spotlight. Waseda University put out a press release under the headline “Bridging the information gap: AI-driven quality control for 5G multicast broadcasting”, and the trade press picked it up the same day. The claim travelled well: a lightweight AI model picked an error-free transmission setting for roughly 87% of video segments, against 32% for a conventional speed-oriented approach. Two numbers, one obvious winner.
The underlying paper is on arXiv, it is free to read, and it contains a table with five rows. The 87% and the 32% are two of them. The other three are not in the press release, and one of them is a plain 40-slot moving average — no machine learning of any kind — that scored 94.55% on the same metric, in the same simulation, on the same data. That is 7.66 points above the AI model the release was written to promote.
This is not a case of a paper hiding something. The paper prints the moving average’s score in bold type, explains exactly why it wins, and explains why winning that particular metric is not the point. The compression happened between the paper and the press release. What follows is what the paper actually measured, why the reliability score rewards the least intelligent method in the table, and what is genuinely new in the work — because there is something genuinely new in it, and it is not the 87%.
Table of contents
- What 5G Multicast Broadcasting Is Trying To Fix
- The 5G Multicast Table The Press Release Did Not Print
- What The 5G Multicast Paper Actually Demonstrates
- The Measurement Setup Was Not A 5G Multicast Network
- The 5G Multicast Speed Claim Holds Up Better Than The Accuracy Claim
- What Is Missing From The 5G Multicast Claim
- What A Reader Should Take From This
- References and Further Reading
What 5G Multicast Broadcasting Is Trying To Fix
The problem is a building. Many older apartment blocks and condominiums were built before fibre, and retrofitting fibre-to-the-home through them is either too expensive or physically impossible. Cable television still carries national broadcasts and emergency information into those homes, and the paper describes CATV as “reliability-first infrastructure” that internet-based services do not fully replace.
5G multicast broadcasting — 3GPP’s Multicast-Broadcast Services, standardised in Release 17 — offers a way out. One transmission serves every household in range instead of one stream per subscriber, so a single dedicated cell can deliver ultra-high-definition television without a cable to each flat. The spectrum efficiency is the entire argument. Japan’s Ministry of Internal Affairs and Communications funded this work under a project explicitly aimed at efficient frequency use for wireless IP multicasting.
The One-Way Street Problem
Ordinary 5G is a conversation. Your phone decodes a block, and if the check fails it says so; the base station sends it again. That mechanism is HARQ, running over the Acknowledged Mode of Radio Link Control, and it forgives a lot. Networks deliberately run at a 10% block error rate because retransmission cleans up the mess.
5G multicast broadcast mode has no such conversation. It runs in RLC Unacknowledged Mode. There is no per-user feedback and no retransmission, so a failed block is simply gone. In an MPEG-DASH stream that corrupts a group of pictures, and the viewer sees a freeze. The paper puts the consequence plainly: apply a conventional unicast link-adaptation algorithm to 5G multicast and you inherit its 10% block error rate as a 10% permanent packet loss rate.
Two Ways To Lose The Picture
There are exactly two failure modes, and they pull in opposite directions. Pick a modulation and coding scheme too aggressive for the channel and the receiver cannot decode; pick one too conservative and the data rate falls below the incoming video bitrate, the base station buffer saturates, and packets are dropped before they are ever transmitted. Because a 5G multicast deployment of this kind uses a fixed MIMO rank of 1, the data rate is strictly proportional to the MCS index, so that second failure is arithmetic rather than bad luck.
The escape from both is prediction. Forecast the 5G multicast channel far enough ahead and you can tell the video encoder to lower its bitrate before the link degrades, rather than after. That is a familiar shape of problem in 5G and 6G trends generally, and it is the shape this paper attacks.
The 5G Multicast Table The Press Release Did Not Print
The paper evaluates five 5G multicast methods in a trace-driven simulation that steps through the held-out 15% of the dataset one 500 ms video segment at a time. Reliability Score is defined as the percentage of segments where the selected MCS was less than or equal to the ground truth — in other words, the percentage of segments that would not have lost packets.
| Method | RMSE | Reliability Score | Average Bias | MAE | In the release? |
|---|---|---|---|---|---|
| Proposed (Transformer + Asymmetric Safety Loss) | 3.1693 | 86.89% | -1.8254 | 2.2978 | Yes, as “approximately 87%” |
| MAW (40-slot moving average) | 4.7178 | 94.55% | -3.4086 | 3.6002 | No |
| MSE-T (same Transformer, standard MSE loss) | 3.0110 | 83.81% | -1.5312 | 2.1478 | No |
| LRA (last observed MCS) | 5.1401 | 83.69% | -2.9341 | 3.5130 | No |
| Deterministic AI (MSE loss, 50% threshold) | 2.8641 | 31.65% | +1.4555 | 2.1438 | Yes, as “conventional speed-oriented approach” |
The release picked the highest and the lowest rows and left the middle three out. The method it calls “a conventional speed-oriented approach” is the Deterministic AI baseline — which is not a deployed convention at all. It is a model the authors built themselves, using their own architecture and a standard loss function, specifically to demonstrate what happens when you optimise for tracking accuracy in a no-retransmission environment. It has the best RMSE in the table, 2.8641, and it is the only method with a positive bias, meaning it consistently overestimates the channel.
Why The Dumbest Method Wins The Headline Metric
The reliability score has a structural bias, and the paper says so directly. It counts only whether the chosen MCS was at or below the ground truth. It does not penalise choosing an MCS far below the ground truth. A method that always picks a very low index therefore scores extremely well on it while delivering a badly degraded picture.
That is exactly what the moving average does. Its average bias is -3.4086 MCS steps, the largest conservative offset of any method, and its MAE of 3.6002 is the worst in the table bar none. The paper’s own verdict is that “while averaging is safe, it is highly inefficient and slow to adapt to improving channel conditions, leading to significant underutilization of channel capacity”. In a 5G multicast deployment where the data rate is proportional to the MCS index, three and a half steps of permanent conservatism is three and a half steps of video quality thrown away on every segment.
So 94.55% is real, and it is also not a claim to a better 5G multicast system. The proposed model’s -1.8254 bias buys its reliability at roughly half the capacity cost. That is the trade, and it is a defensible one. It is simply not the trade the press release described.
What The 5G Multicast Paper Actually Demonstrates
Strip out the headline comparison and a cleaner result remains, and it is the one the authors clearly care about. The MSE-T baseline is the same Transformer, the same 12 input features, the same 90% decision threshold — everything identical except the loss function. It scores 83.81%. The proposed model, trained with an Asymmetric Safety Loss that weights overestimation errors by a factor of 1.4, scores 86.89%.
Three points is a modest number and a clean one. It is attributable to a single named change, it comes with a matched control, and it is the sort of 5G multicast result that survives replication. The 55-point gap the release leads with is attributable instead to a baseline the authors constructed to fail.
The Genuinely Unusual Ingredient
The dataset is the part of this work that is hard to reproduce. The authors logged a commercial 5G network at 0.5 ms slot-level granularity and ended up with 25.9 million data points after preprocessing, split 70/15/15 into training, validation and test sets. Most published link-adaptation work aggregates telemetry at one-second intervals or uses simulated channels; at that resolution the millisecond-scale fast fading that decides whether a block decodes is averaged away entirely.
Getting that 5G multicast data took real effort. Parameters arrive at different rates — MCS, resource blocks, CRC state and CQI every 0.5 ms, SINR every 20 ms, RSRP every 160 ms — so everything was carried forward onto a single 0.5 ms timeline. The model itself has 3,469 trainable parameters, which is small enough to be startling: two encoder layers, a latent dimension of 8, two attention heads and a feed-forward dimension of 32.
What The Model Is Asked To Predict
The framework does not output one number. It outputs a 1×28 vector giving the success probability of every MCS index across a future 500 ms video segment, from a 25 ms observation window of 40 slots by 12 features, with a deliberate 100 ms gap for signal propagation and encoder processing. A decision policy then picks the highest index clearing a 90% success threshold.
That threshold is a dial, not a discovery. It was set at 90% because 10% block error rate is the standard target for enhanced mobile broadband, and the paper notes it can be moved to match an ultra-reliable low-latency target of 10⁻⁵ instead. Every reliability figure quoted anywhere in the coverage is conditional on that dial being where the authors left it. The model is a supervised predictor, not a reinforcement learning agent, so the threshold sits outside the trained system entirely.
The Measurement Setup Was Not A 5G Multicast Network
This is the disclosure that most deserves to survive the trip into a press release, and it did not. The paper is explicit and the reasoning is sound, but the gap between what was measured and what is claimed is wide.
| Dimension | Target deployment | What was measured |
|---|---|---|
| RLC mode | Unacknowledged (no HARQ) | Acknowledged, with HARQ active |
| Transmission | Broadcast to many receivers | Unicast to one test handset |
| Resources | Dedicated, no unicast contention | Shared commercial cell, off-peak walk |
| MIMO rank | Fixed rank 1 | Forced to 1Rx by rewriting modem NV items |
| Location | Japanese multi-dwelling units | An indoor shopping mall in Bangkok |
| Duration | Continuous service | ~5 hours across two mornings |
| Mobility | Fixed wireless access, stationary | Walking pace only |
Because no 5G multicast network was available, the ground truth had to come from somewhere else. The team ran a saturated download on the Advanced Info Service commercial network in Thailand, on band n41 with 100 MHz of bandwidth, at an indoor mall on 16–17 September 2025 during pre-opening hours. HARQ’s CRC pass/fail states became the labels. That is a reasonable proxy — a block that failed its check would have failed in broadcast mode too — but it is a proxy, and every reliability figure is a simulation over it rather than a measurement of a working 5G multicast service.
The Handset Was Modified
To emulate the single-antenna configuration a 5G multicast deployment would use, the team rewrote non-volatile items in a Galaxy S22 Ultra’s modem firmware to power down three of its four receive paths at the hardware level. They then quantified the cost across six test spots: an average CSI-SINR of 10.22 dB with four antennas against 7.58 dB with one, a penalty of 2.64 dB. The walk path covered RSRP from -140 dBm to -44 dBm, which is close to the full working range.
The Site Was A Substitute Too
The mall was chosen because its 4T4R passive panel antennas resemble a 5G multicast deployment more closely than the 64T64R massive-MIMO units used outdoors. The paper records, via an operator engineer, that this 4T4R setup was itself an interim installation rushed in for the mall’s opening while a permanent solution was still being designed. The first attempt at data collection was abandoned because opening-week crowds congested the network.
None of this is misconduct. It is what field research on unavailable infrastructure looks like, and the paper documents every step of it. But “tests on a real-world commercial 5G network” reads as validation of the product, and what happened was a carefully argued substitution at six separate points.
The 5G Multicast Speed Claim Holds Up Better Than The Accuracy Claim
The inference timing is the most solid part of the work, because it was measured directly on hardware rather than simulated. The 5G multicast model was converted to TensorFlow Lite and benchmarked over 1,000 iterations per device.
| Device | SoC | Year | Inference | % of a 0.5 ms slot |
|---|---|---|---|---|
| Google Pixel 3 | Snapdragon 845 | 2018 | 1.32 ms | 264% |
| Huawei P40 Pro | Kirin 990 5G | 2020 | 0.07 ms | 14% |
| Sony Xperia 1 III | Snapdragon 888 | 2021 | 0.04 ms | 8% |
| Samsung Galaxy S22 Ultra | Snapdragon 8 Gen 1 | 2022 | 0.04 ms | 8% |
| Samsung Galaxy Z Flip5 | Snapdragon 8 Gen 2 | 2023 | 0.04 ms | 8% |
| Google Pixel 10 Pro XL | Tensor G5 | 2025 | 0.03 ms | 6% |
The 0.07 Figure Is The Floor, Not The Typical Case
Coverage of this work says the model runs in “less than 0.07 ms on smartphones from 2020 onward”. Both halves of that sentence are doing work. The 0.07 ms belongs to the single slowest qualifying device, the 2020 Huawei P40 Pro running through Android’s neural networks API; every Snapdragon device in the pool, using Qualcomm’s Hexagon backend, ran at 0.04 ms or better.
The “2020 onward” qualifier is load-bearing too, because it excludes the one device that fails. The 2018 Pixel 3 needs 1.32 ms — more than two and a half 5G multicast transmission intervals — because it has no dedicated neural accelerator. Quoting the worst passing result rather than the best is the conservative choice, and it is to the authors’ credit.
What Is Missing From The 5G Multicast Claim
The paper is candid about scope, and its own future-work section names most of the gaps. It is worth listing them, because the press framing of a solved information gap does not survive contact with them.
No End-To-End System Was Built
The 5G multicast architecture the paper describes has a receiver predicting, an edge server aggregating those predictions and setting a global MCS, a base station applying it, and a video encoder adjusting its bitrate to match. Only the first of those exists. The paper says evaluating the complete system “exceeds the scope of this paper” and confines itself to the receiver-side prediction framework. There is no measurement anywhere of video quality, stall rate, or viewer experience — the quantities the headline is really about.
One Device, One Site, One Mobility Profile
Every figure comes from a single handset model on a single band at a single indoor location, walking. The authors list the expansions they intend: more handset models and modem chipsets, outdoor environments, stationary use, and vehicular scenarios on buses and trains with the Doppler shifts those bring. Until then, generalisation is an open question rather than a demonstrated property.
The Unicast Fallback Is Still Theoretical
The design assumes an adaptive 5G multicast scheme where receivers whose predicted channel quality falls below the global MCS get a supplementary unicast stream. That fallback is what would rescue the households in the worst reception — the ones in the hardest-to-reach flats, who are the stated reason for the whole exercise. It is described as future work. Longer-range thinking about where this leads sits in the agentic 7G roadmap for AI-run wireless networks.
What A Reader Should Take From This
The honest summary is smaller than the headline and more durable. On a 0.5 ms-resolution trace from a commercial network, a 3,469-parameter Transformer trained with a loss function that penalises overestimation 1.4 times as hard as underestimation picked a non-lossy transmission setting for 86.89% of simulated segments, against 83.81% for the identical model trained conventionally — and it did so fast enough to run on a phone with room to spare.
That is a competent, well-instrumented, honestly reported piece of engineering research, presented at a conference in early September 2026. It is not a demonstration that 5G multicast broadcasting now works reliably in the field, because no 5G multicast field deployment was tested. And the 87%-versus-32% framing inverts the table it came from: the 32% belongs to a straw baseline the authors built to be beaten, while the best score in the table belongs to a moving average that the same table shows would cost viewers three and a half MCS steps of picture quality.
How To Read The Next Release Like This One
Three checks would have caught this in about ten minutes, and they generalise well beyond 5G multicast. Find the paper — this one is on arXiv, free, with a full HTML version. Read the results table rather than the abstract, and count the rows: if the release names two of five, ask what the other three did. Then read the metric’s definition, because a metric that only punishes errors in one direction will always crown the most timid method in the comparison.
The last one is the general lesson. Reliability Score here is a one-sided measure, and one-sided measures reward doing nothing. The paper knew that, which is why it reports RMSE, bias and MAE alongside. A press release that quotes one column of a four-column table is not describing the result; it is describing one column.
References and Further Reading
Transformer-Based MCS Prediction for 5G Multicast-Broadcast Services (MBS) — arXiv abstract
Transformer-Based MCS Prediction for 5G Multicast-Broadcast Services (MBS) — full text
Bridging the information gap: AI-driven quality control for 5G multicast broadcasting — Tech Xplore
3GPP — Broadcast and multicast technologies
Hybrid automatic repeat request — Wikipedia
Radio Link Control — Wikipedia