A memory-augmented neural decoder scored higher than a Transformer when the two models were trained with limited data on three public neural benchmarks, according to an arXiv version 1 preprint dated 25 Aug 2026. The clearest separation came on MC RTT, where VN-SST reached a peak behavioral-decoding R2 of 0.351, compared with 0.207 for the Transformer.
The study focused on sample-efficient neural decoding: producing an accurate read-out of behavior from limited data and small models. Its main measure was validation behavioral-decoding R2, a score of how closely predicted finger or hand velocity matched recorded behavior. The joint neural codec also predicted the next population spikes from the same hidden state.
A decoder with a programmable inner layer
VN-SST's central change is a shared, low-rank instruction bank that synthesizes the model's feed-forward projections separately for each token. It also combines local causal attention with selective state-space and fast-weight memory pathways.
For each neural codec, the comparison fixed both architectures at four layers and varied their width across a target range of about 64,000 to 270,000 non-embedding parameters. The parameter sweep used roughly 25% of the full training budget, while the data and context sweeps used the largest available budget. Neural results were averaged over three random seeds.
The gap narrowed as more data arrived
Under limited data, VN-SST had the higher peak R2 on all three codecs. It scored 0.351 versus 0.207 on MC RTT, 0.716 versus 0.655 on MC Maze, and 0.696 versus 0.632 on Area2 Bump. A memoryless linear reference on MC RTT scored -0.236.
The benchmarks represent different recordings and movement targets. MC RTT had 130 units from primary motor cortex and tracked finger velocity. MC Maze had 182 units from primary motor and premotor cortex and tracked hand velocity. Area2 Bump had 65 units from somatosensory area 2 and also tracked hand velocity.
The largest separation appeared at the smallest tested data budgets. On MC RTT, VN-SST rose from 0.295 to 0.607 across the sweep, while the Transformer rose from 0.192 to 0.522. On MC Maze, the corresponding ranges were 0.672 to 0.854 and 0.655 to 0.847. On Area2 Bump, they were 0.678 to 0.795 and 0.599 to 0.778. At the full budget, the two models were closer on MC Maze and Area2 Bump, though VN-SST remained at least as accurate across the data sweep.
The codec was trained to predict both population spikes and hand or finger velocity from one hidden state. Spikes were binned in 50-millisecond intervals, smoothed across three bins, z-scored and passed through a 128-bin prefix. The behavioral loss was weighted by 10, while neural next-step RMSE stayed near an approximate 1.0 level.
Sequence length showed different patterns across models
A separate experiment varied the truncated backpropagation-through-time window, meaning the recent sequence length used during training, across 16, 32, 64 and 128 tokens. The model and full training set were held fixed. Epochs were scaled with window length to equalize optimizer steps, and the results were averaged over three seeds.
On scarce MC RTT data, the best Transformer score was 0.612 with a 16-token context, while VN-SST's best was 0.680 with 128 tokens. Across the tested lengths, the Transformer score declined as the context grew, while the VN-SST score rose. The data-rich codecs were mostly flat: on MC Maze, the best values were 0.879 at 32 tokens for the Transformer and 0.878 at 16 tokens for VN-SST; on Area2 Bump, they were 0.766 at 32 and 0.779 at 64. These results show an association in this benchmark, not that context length caused the difference.
The instruction bank was used selectively
The researchers also examined the model's instruction codes, an internal diagnostic of how its token-specific operators were used. With a bank of 32 instructions, code bits and participation ratio were 2.55 and 3.4 on MC RTT, 2.89 and 5.1 on MC Maze, and 3.20 and 6.5 on Area2 Bump. The pattern indicates concentrated use of a relatively small set of effective operators.
Decode R2 stayed essentially flat as the instruction-bank capacity increased. Across bank sizes, R2 ranged from 0.635 to 0.664 on MC RTT, 0.864 to 0.872 on MC Maze, and 0.779 to 0.787 on Area2 Bump. The diagnostic does not show that the learned codes are discrete biological programs.
A small text test produced a similar signal
The authors added an auxiliary language check using byte-level BPE versions of tiny-Shakespeare and WikiText-2. The same VN-SST and Transformer were compared with two layers, a parameter ladder of about 0.13 million to 1.7 million non-embedding parameters, three random seeds and data budgets equal to one or three times the baseline.
At the largest budget, VN-SST recorded lower validation perplexity, a score used for next-piece prediction, on both corpora: 49.4 versus 60.6 on tiny-Shakespeare and 45.1 versus 52.1 on WikiText-2. It reached the Transformer's best result with roughly two to three times fewer parameters. The comparison was an auxiliary check on two small text corpora.
What the comparison leaves unresolved
The comparison does not isolate the instruction bank's contribution. VN-SST changes both its memory pathways and its feed-forward operator, so the separate effects of instruction synthesis, selective state-space memory, fast-weight memory and local attention remain unresolved. Parameter counts were not exactly matched, and the evidence comes from three neural recordings and two small text corpora.
The outcomes are model-level decoding and language-modeling measures, not clinical endpoints or direct evidence about cortical computation. The reported neural and text results are means over three seeds, and the supplied analysis identifies no confidence intervals, formal significance tests or cross-seed dispersion.
The neural recordings and text corpora are publicly available. Experiment code is available from the authors on request, subject to internal review. The acknowledgments state that the institution supported the study.
Paper data and sources
Original title: The Von-Neumann State-Space Transformer for neural decoding
Authors: Morteza Sarafyazd
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text