BanglaMamba, a Mamba-based model trained from scratch, used markedly less GPU memory and recorded lower latency than the two BERT-based models in a three-way comparison of Bangla fake-news classifiers. But pretrained BanglaBERT had the highest Macro-F1 on both the main dataset and an external test set. BanglaMamba was close to from-scratch CustomBERT on the main data, while the external results placed CustomBERT between BanglaBERT and BanglaMamba. The comparison points to a trade-off between classification results and computing demands within the tested setup.
The work is arXiv:2608.25190v1, dated 25 Aug 2026, and remains a preprint. No journal or peer-review status is reported. It compared pretrained BanglaBERT with CustomBERT, a comparable Transformer trained from scratch, and BanglaMamba, a Mamba-based model also trained from scratch.
How the test was built
The analysis used processed Bangla news articles labeled Fake or Real. The primary BanFakeNews-2.0 corpus contained 58,001 samples, with 16.5% labeled Fake and 83.5% Real. The separate external test set contained 3,979 samples and was nearly evenly divided, with 49.9% Fake and 50.1% Real.
Before testing, the pipeline joined each headline to its body with a [SEP] marker. It removed HTML, URLs, duplicate and empty samples, normalized Unicode and whitespace, and excluded articles shorter than 20 words or longer than 2,000 words.
After cleaning, the data were split by stratified sampling into 80% training, 10% validation and 10% testing. Inverse-frequency class weights calculated only from the training set were used in weighted cross-entropy.
All three systems used the BanglaBERT SentencePiece tokenizer, with a 32,000-token vocabulary and a 512-token maximum input length. The study's short-versus-long comparison was made under that same cap.
Training used weighted cross-entropy and AdamW for five epochs, with BF16 on one NVIDIA A100 GPU with 40 GB of memory. The checkpoint with the highest validation Macro-F1 was evaluated, and reported results were averaged across random seeds 22, 33 and 42.
Cleaning and exact deduplication preceded partitioning. The pretrained tokenizer was not adapted to the study data, and test data were excluded from checkpoint selection and parameter updates.
BanglaBERT led on the main score
On the main dataset, BanglaBERT's Macro-F1 was 0.9260, with a standard deviation of 0.0008. BanglaMamba scored 0.9029 plus or minus 0.0034, and CustomBERT 0.9057 plus or minus 0.0078. BanglaBERT was highest, while the two from-scratch systems produced similar lower scores.
By class, BanglaBERT also led. Its Fake F1 was 0.8753 plus or minus 0.0011 and its Real F1 was 0.9766 plus or minus 0.0005. BanglaMamba recorded 0.8361 plus or minus 0.0057 for Fake and 0.9697 plus or minus 0.0017 for Real, while CustomBERT recorded 0.8408 plus or minus 0.0136 and 0.9706 plus or minus 0.0021. Every model's Fake F1 was lower than its Real F1.
The imbalance in the primary corpus provides context for those class scores: 16.5% of samples were Fake and 83.5% were Real. The lower Fake F1 values were observed alongside that uneven class distribution.
Speed and memory favored BanglaMamba
BanglaMamba's clearest advantage appeared in the efficiency benchmark. Its P50 latency, the midpoint response-time measure, was 14.8433 plus or minus 2.2382 milliseconds, compared with 37.0767 plus or minus 0.2701 milliseconds for BanglaBERT and 37.0333 plus or minus 0.2845 milliseconds for CustomBERT. At P95, a slower-end response-time measure, BanglaMamba was at 17.4267 plus or minus 3.1683 milliseconds, versus 39.03 plus or minus 3.0967 and 39.05 plus or minus 3.12 milliseconds. It also had higher throughput than both BERT-based models under the reported benchmark.
Peak GPU memory use showed the same contrast. BanglaMamba required approximately 49% less peak VRAM during inference and nearly 78% less during training than the BERT-based models.
The short-versus-long comparison showed a more mixed picture. Under the 512-token maximum, BanglaBERT's Macro-F1 was 0.9309 plus or minus 0.0047 for short articles and 0.9054 plus or minus 0.0175 for long articles. BanglaMamba's figures were 0.9030 plus or minus 0.0012 and 0.9012 plus or minus 0.0217, while CustomBERT's were 0.9049 plus or minus 0.0063 and 0.9080 plus or minus 0.0149. The observed changes were generally small: BanglaBERT showed the larger decrease, while CustomBERT slightly improved on long articles. No formal statistical test was reported for this comparison.
The ranking widened on an external set
On the external dataset, BanglaBERT's Macro-F1 was 0.7914 plus or minus 0.0348. CustomBERT scored 0.7227 plus or minus 0.0892 and BanglaMamba 0.5753 plus or minus 0.0703. The ranking was the same for Fake F1: 0.7763 plus or minus 0.0365 for BanglaBERT, 0.6824 plus or minus 0.1081 for CustomBERT and 0.4564 plus or minus 0.1253 for BanglaMamba.
The authors interpret the results as a trade-off. BanglaMamba was close to CustomBERT on the main dataset and had lower latency, higher throughput and lower memory use, while BanglaBERT was stronger on the main classification measure and the external test. They suggest BanglaBERT's pretraining may partly explain its stronger cross-dataset performance, while acknowledging that architecture and optimization differences also varied between the systems.
An additional check compared BanglaBERT's original [CLS] approach with masked mean pooling. Mean pooling recorded a Macro-F1 of 0.9144 plus or minus 0.0081, while the authors described the result as broadly comparable to the original approach, which performed slightly better. No formal significance test was reported for this ablation.
What the comparison leaves open
The evidence is limited to computational experiments on the two datasets and the model configurations described in the preprint. BanglaMamba was trained from scratch on the downstream fake-news data, without large-scale pretraining, so the comparison does not answer how it would perform after such pretraining. It also does not establish broad generalization across Bangla news domains or real-world deployment reliability.
All inputs were capped at 512 tokens, leaving open how the models would perform on longer articles without truncation. Cross-dataset evidence came from one external dataset, and the results were averaged across three random seeds with standard deviations reported. Whether the efficiency advantage would persist on other hardware or deployment configurations remains unresolved.
No funding source is identified in the supplied text. The authors say AI tools assisted with writing and code development, while the experimental data were generated through their own experiments and pipelines.
Paper data and sources
Original title: BanglaMamba: Exploring State Space Models for Bangla Fake News Detection
Authors: M. K. Khalidi Siam
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text