The central result is an efficiency gain in a computer simulation. A system built around a shared ontology reported 83.4% less transmitted data on average than complete scene-graph transmission, while recording the highest multi-view question-answering score at 64.8%. The comparison also put it 4.5%, 5.6% and 16.0% above three named baselines. Here, multi-view visual question answering, or VQA, means answering questions about a scene using information from several views. The work is an arXiv v1 preprint, so the figures describe the tested model and benchmark rather than a demonstrated real-world deployment.
A shared language between machines
The proposed design is called ontology-driven semantic communication, or ODSC. An ontology here is a shared set of terms and rules that gives machines a common way to represent facts. ODSC puts that knowledge base at the transmitters and the receiver, using synonym mappings, inference rules and consistency constraints. Transmitters remove scene-graph information that can be inferred; the receiver aligns different expressions, recovers the removed information and checks for conflicts between views.
That removal is described by the authors as semantically lossless within the shared ontology. In other words, the selected and original scene graphs are treated as carrying the same facts under the system's rules, even though the transmitted version contains fewer entries.
What the simulation compared
The simulation used seven transmitters, one receiver and seven heterogeneous commercial vision-language models, or VLMs, to extract scene descriptions. Its benchmark was the GQA validation set: 10,696 scenes with ground-truth scene graphs and 132,062 questions. ODSC was compared with four algorithmic alternatives: sending complete scene graphs, task-adaptive semantic filtering, relation filtering and string-matching verification of semantic consistency.
The data savings were substantial
On the ground-truth scene graphs, ODSC removed 60.2% of the subject-relation-object facts known as triples, while all answers remained identical to the complete-graph baseline. In the transmitter simulation, each sender's data size was 77.4% to 87.1% lower than baseline a, with an average reduction of 83.4%. The reported reduction was 38.5% larger than task-adaptive filtering and 53.9% larger than relation filtering.
Those savings did not come with a lower reported score in the evaluated comparisons. In the selected single-view analysis, which covered 8,067 questions supported by the complete scene graphs of the seven transmitters, ODSC's average accuracy was 11.7% higher than baseline a, 12.7% higher than baseline b and 22.9% higher than baseline c. For multi-view answering, its 64.8% was the highest reported result. The analysis also reported disambiguation gains of 4.9% to 29.3% for individual transmitters, while a random permutation of transmitter correspondences changed accuracy by less than 1%.
ODSC was also assessed on whether it could identify erroneous objects across views. Recall here means the share of identified erroneous objects. It rose from 37.5% with two transmitters to 81.3% with seven in the reported analysis. Complete-scene-graph transmission identified none, while string-matching verification stayed below 46% and stopped increasing beyond four transmitters. The measure is narrower than an all-purpose error rate and does not capture every extraction error.
Delay fell, but the test remained simulated
The efficiency difference also appeared in delay measurements. At a simulated signal-to-noise ratio of 10 dB, average transmission delay was 4.2 milliseconds with ODSC, compared with 26.4 milliseconds for baseline a, a reported single-transmitter reduction of up to 87.1%. For per-scene processing, ODSC took 5.9 ms and used 20,000 logical operations. Baseline b took 21.9 seconds and used 200 billion floating-point operations.
With one transmitter, ODSC's reported multi-view accuracy was 52.8%; with seven, it was 64.8%. The shared ontology listed 3,049 predefined synonyms, rules and constraints. The processing figures came from a single Linux server with one NVIDIA H100 GPU, so they are measurements of that hardware setup rather than a general performance guarantee.
The reported figures are point estimates. The evaluation provided no confidence intervals, p-values or other formal uncertainty estimates, and it excluded offline ontology construction and provisioning from communication overhead.
The evidence stops at the simulation
The caveats are substantial. Evidence was limited to the GQA validation set, the selected commercial VLM extractors, one shared ontology and the listed simulated conditions. The supplied report does not identify the seven VLMs by name or give extraction-quality measurements, and its ontology coverage and sensitivity to mapping errors were not quantified. Unmapped triples were excluded from question answering, and the study does not establish how the method would perform on other datasets, scene types, VLMs, wireless conditions or correlated extraction errors across views.
Within its computational setup, the preprint reports less transmitted information, shorter simulated delay and higher VQA scores. It does not establish benefits for human users or real-world deployment, nor does it show that all incorrect scene-graph information is corrected. Independent tests on other datasets and with other VLMs, along with full accounting for ontology provisioning, would be needed to clarify how widely the result applies. The document is an arXiv v1 preprint dated 26 Aug 2026.
Paper data and sources
Original title: Towards Faithful and Efficient Semantic Communication: An Ontological Approach
Authors: Yixiao Feng, Yueting Wang, Yining Wang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text