pith. sign in

arxiv: 2606.11066 · v1 · pith:GDPDQWMYnew · submitted 2026-06-09 · 💻 cs.LG · q-bio.NC

GRAFT: Gain-Recalibrated Adapters for Transformer-Based Neural Population Activity Modeling

Xiangsheng Ge , Yang Xie This is my paper

Pith reviewed 2026-06-27 14:11 UTC · model grok-4.3

classification 💻 cs.LG q-bio.NC
keywords neural population activityTransformer modeladapterscross-day recalibrationbrain-computer interfacesNLB'21 benchmarkgain mechanisms
0
0 comments X

The pith

GRAFT separates reusable temporal dynamics from neuron-specific interfaces in Transformers, achieving state-of-the-art neural population activity modeling and efficient cross-day recalibration.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents GRAFT, a Transformer model for neural population activity that decouples the shared temporal backbone from a neuron interface. This separation allows the model to reach 0.3866 co-bps on the MC Maze dataset under standard protocols, establishing a new state of the art. For cross-day adaptation, GRAFT updates only 9.21 percent of parameters to handle scaled versions of the dataset while maintaining strong performance scores of 0.3749, 0.3112, and 0.3152 co-bps. A sympathetic reader would care because this addresses the practical challenge of changing neuron recordings in long-term brain-computer interfaces without retraining the entire model.

Core claim

GRAFT introduces gain-recalibrated adapters that control neuron entry and exit from the shared Transformer backbone, supported by auxiliary gain and positional mechanisms, enabling both high-accuracy modeling on fixed datasets and data-efficient recalibration across days by updating only the interface parameters.

What carries the argument

Gain-recalibrated adapters separating the neuron interface from the reusable temporal Transformer backbone.

If this is right

  • The same architecture achieves new state-of-the-art co-bps on the primary NLB'21 metric.
  • Cross-day recalibration requires updating only 9.21% of parameters.
  • The interface-backbone separation supports both strong modeling and adaptation.
  • Performance holds with restricted target-day support sets on scaled MC Maze datasets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This separation approach might extend to other sequence modeling tasks where input dimensions change over time.
  • Testing on additional neural datasets beyond MC Maze could reveal the generality of the interface-backbone split.
  • If the backbone captures universal dynamics, similar adapters could apply to multi-subject neural recordings.

Load-bearing premise

The gain-recalibrated adapters successfully isolate reusable temporal dynamics from neuron-specific interfaces such that updating only the interface parameters preserves modeling accuracy across days without hidden degradation from the fixed backbone.

What would settle it

A controlled experiment showing that performance on target days drops substantially below the reported levels when the backbone remains fixed and only the interface is updated, or that the ensemble fails to exceed prior reported results on the standard protocol.

Figures

Figures reproduced from arXiv: 2606.11066 by Xiangsheng Ge, Yang Xie.

Figure 1
Figure 1. Figure 1: Overview of GRAFT. GRAFT uses a reusable Transformer dynamics backbone with a neuron interface for read [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
read the original abstract

Neural population activity models can recover rich temporal structure from binned spikes, but their read-in and readout layers often remain tied to a fixed set of recorded neurons. This coupling limits reuse in long-term brain-computer interfaces, where recorded neuron identities, counts, and response statistics can change across days. We introduce GRAFT, a Transformer-based neural population activity model that separates reusable temporal dynamics from a recalibratable neuron interface. The neuron interface controls how recorded neurons enter and leave the shared backbone, and auxiliary gain and positional mechanisms support neural activity modeling inside the Transformer. On MC Maze under the standard NLB'21 protocol, GRAFT reaches 0.3866 co-bps as an ensemble, setting a new state of the art on the primary co-bps metric among public and reported NLB'21 results. In a cross-day protocol constructed from the NLB'21 MC Maze dataset series, GRAFT recalibrates from MC Maze to the scaled MC Maze datasets (Large/Medium/Small) by updating only 9.21% of parameters, reaching 0.3749, 0.3112, and 0.3152 co-bps with restricted target-day support sets. These results show that the same interface-backbone separation supports both strong Transformer-based neural population activity modeling and data-efficient cross-day recalibration.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 3 minor

Summary. The paper introduces GRAFT, a Transformer-based architecture for neural population activity modeling that employs gain-recalibrated adapters to separate reusable temporal dynamics in a shared backbone from neuron-specific read-in/read-out interfaces. It reports an ensemble co-bps of 0.3866 on the MC Maze dataset under the NLB'21 protocol (new SOTA) and shows that cross-day recalibration to scaled MC Maze variants (Large/Medium/Small) can be performed by updating only 9.21% of parameters while reaching 0.3749/0.3112/0.3152 co-bps with restricted target-day support.

Significance. If the empirical results hold under standard evaluation protocols, the work provides a concrete demonstration that interface-backbone separation can support both high-accuracy single-session modeling and parameter-efficient cross-session reuse. This is directly relevant to long-term BCI applications where neuron identities and counts change across days. The use of a public benchmark (NLB'21) and explicit parameter-update percentages are strengths that aid reproducibility and comparison.

major comments (2)
  1. [§4.3, Table 3] §4.3 and Table 3: the cross-day protocol is described as 'constructed from the NLB'21 MC Maze dataset series' with 'scaled' Large/Medium/Small variants, but the precise scaling procedure (e.g., neuron subsampling ratios, spike-rate adjustments, or temporal alignment) is not stated; without this, it is impossible to verify that the reported 9.21% update figure isolates the claimed interface separation rather than benefiting from dataset-specific artifacts.
  2. [§5.1, Table 1] §5.1, Table 1: the ensemble co-bps of 0.3866 is presented as SOTA, yet the table does not report per-model standard deviations or the number of independent runs; given that the single-model numbers are close to prior baselines, the statistical significance of the improvement over the next-best reported method cannot be assessed from the provided data.
minor comments (3)
  1. [§3.2] §3.2: the auxiliary gain and positional mechanisms are introduced without an equation or diagram showing how the gain vector is computed from the adapter output and applied inside the Transformer blocks; a single equation would clarify the 'recalibration' step.
  2. [Figure 2] Figure 2: the schematic of the adapter insertion points is helpful, but the legend does not distinguish frozen vs. updated parameters during the cross-day phase, making it harder to map the 9.21% figure to the diagram.
  3. [§4.1] §4.1: the support-set sizes for the target-day recalibration experiments are stated only in the text; adding them to Table 3 would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive review and the recommendation for minor revision. We address each major comment below.

read point-by-point responses

  1. Referee: [§4.3, Table 3] §4.3 and Table 3: the cross-day protocol is described as 'constructed from the NLB'21 MC Maze dataset series' with 'scaled' Large/Medium/Small variants, but the precise scaling procedure (e.g., neuron subsampling ratios, spike-rate adjustments, or temporal alignment) is not stated; without this, it is impossible to verify that the reported 9.21% update figure isolates the claimed interface separation rather than benefiting from dataset-specific artifacts.

    Authors: We agree that the scaling procedure must be stated explicitly. The current manuscript describes the variants only at a high level. In the revision we will expand §4.3 with a complete specification of the construction steps, including the exact neuron subsampling ratios, any spike-rate normalization, and the temporal alignment method used to create the Large/Medium/Small target-day sets. This addition will allow direct verification that the 9.21 % parameter update isolates the interface recalibration. revision: yes

  2. Referee: [§5.1, Table 1] §5.1, Table 1: the ensemble co-bps of 0.3866 is presented as SOTA, yet the table does not report per-model standard deviations or the number of independent runs; given that the single-model numbers are close to prior baselines, the statistical significance of the improvement over the next-best reported method cannot be assessed from the provided data.

    Authors: We acknowledge that Table 1 currently reports only the ensemble co-bps without accompanying run statistics. The 0.3866 figure is the mean performance of an ensemble of independently trained models. In the revised manuscript we will augment Table 1 with the number of constituent models and the per-model standard deviation (or range) so that readers can evaluate the stability of the reported improvement relative to the next-best baseline. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper introduces an empirical architecture (GRAFT) and reports direct performance measurements (co-bps values) on the public NLB'21 MC Maze benchmark and a constructed cross-day protocol. No derivation chain, first-principles result, or prediction is claimed that reduces by construction to fitted parameters, self-definitions, or load-bearing self-citations. The central results are benchmark numbers obtained by training and evaluating the model; the separation hypothesis is tested via those measurements rather than derived from prior equations within the paper.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The SOTA and recalibration claims rest on the validity of the NLB'21 evaluation protocol and on the assumption that the constructed cross-day dataset series fairly tests interface recalibration; no additional free parameters or invented physical entities are introduced beyond the model itself.

axioms (1)
  • domain assumption The standard NLB'21 protocol and the constructed cross-day splits from the MC Maze series provide a fair and comparable test of co-bps performance.
    The paper invokes this protocol to claim both the absolute SOTA and the cross-day results.
invented entities (1)
  • Gain-recalibrated adapters no independent evidence
    purpose: Separate reusable temporal dynamics from a neuron-specific interface that can be updated with few parameters.
    New architectural component introduced to solve the cross-day reuse problem; no independent evidence outside the reported experiments is supplied.

pith-pipeline@v0.9.1-grok · 5773 in / 1418 out tokens · 36278 ms · 2026-06-27T14:11:03.145912+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 1 canonical work pages

  1. [1]

    and Wiesel, Torsten N

    Hubel, David H. and Wiesel, Torsten N. , title =. The Journal of Physiology , year =

  2. [2]

    and Wang, Xiao-Jing and Daw, Nathaniel D

    Rigotti, Mattia and Barak, Omri and Warden, Melissa R. and Wang, Xiao-Jing and Daw, Nathaniel D. and Miller, Earl K. and Fusi, Stefano , title =. Nature , year =

  3. [3]

    , title =

    Saxena, Shreya and Cunningham, John P. , title =. Current Opinion in Neurobiology , year =

  4. [4]

    and Perich, Matthew G

    Gallego, Juan A. and Perich, Matthew G. and Miller, Lee E. and Solla, Sara A. , title =. Neuron , year =

  5. [5]

    and Sahani, Maneesh and Churchland, Mark M

    Shenoy, Krishna V. and Sahani, Maneesh and Churchland, Mark M. , title =. Annual Review of Neuroscience , year =

  6. [6]

    and Sussillo, David and Shenoy, Krishna V

    Vyas, Saurabh and Golub, Matthew D. and Sussillo, David and Shenoy, Krishna V. , title =. Annual Review of Neuroscience , year =

  7. [7]

    and Naud, Richard and Paninski, Liam , title =

    Gerstner, Wulfram and Kistler, Werner M. and Naud, Richard and Paninski, Liam , title =

  8. [8]

    and Fellows, Matthew R

    Truccolo, Wilson and Eden, Uri T. and Fellows, Matthew R. and Donoghue, John P. and Brown, Emery N. , title =. Journal of Neurophysiology , year =

  9. [9]

    Neuron , year =

    Salinas, Emilio and Thier, Peter , title =. Neuron , year =

  10. [10]

    and Newsome, William T

    Mante, Valerio and Sussillo, David and Shenoy, Krishna V. and Newsome, William T. , title =. Nature , year =

  11. [11]

    Le, Trung and Fang, Hao and Li, Jingyuan and Nguyen, Tung and Mi, Lu and Orsborn, Amy L. and S. SPINT: Spatial Permutation-Invariant Neural Transformer for Consistent Intracortical Motor Decoding , booktitle =. 2025 , location =

    2025

  12. [12]

    and Hochberg, Leigh R

    Fan, Chaofei and Hahn, Nick and Kamdar, Foram and Avansino, Donald and Wilson, Guy H. and Hochberg, Leigh R. and Shenoy, Krishna V. and Henderson, Jaimie M. and Willett, Francis R. , title =. Advances in Neural Information Processing Systems , volume =. 2023 , location =

    2023

  13. [13]

    and Kapitonava, Anastasia and Kamdar, Foram and Henderson, Jaimie M

    Pun, Tsam Kiu and Khoshnevis, Mona and Hosman, Tommy and Wilson, Guy H. and Kapitonava, Anastasia and Kamdar, Foram and Henderson, Jaimie M. and Simeral, John D. and Vargas-Irwin, Carlos E. and Harrison, Matthew T. and Hochberg, Leigh R. , title =. Communications Biology , year =

  14. [14]

    and Ali, Yahia H

    Karpowicz, Brianna M. and Ali, Yahia H. and Wimalasena, Lahiru N. and Sedler, Andrew R. and Keshtkaran, Mohammad Reza and Bodkin, Kevin and Ma, Xuan and Rubin, Daniel B. and Williams, Ziv M. and Cash, Sydney S. and Hochberg, Leigh R. and Miller, Lee E. and Pandarinath, Chethan , title =. Nature Communications , year =

  15. [15]

    and Ye, Joel and Fan, Chaofei and Tostado-Marcos, Pablo and Rizzoglio, Fabio and Washington, Clay and Scodeler, Thiago and de Lucena, Diogo and Nason-Tomaszewski, Samuel R

    Karpowicz, Brianna M. and Ye, Joel and Fan, Chaofei and Tostado-Marcos, Pablo and Rizzoglio, Fabio and Washington, Clay and Scodeler, Thiago and de Lucena, Diogo and Nason-Tomaszewski, Samuel R. and Mender, Matthew J. and Ma, Xuan and Arneodo, Ezequiel Matias and Hochberg, Leigh R. and Chestek, Cynthia A. and Henderson, Jaimie M. and Gentner, Timothy Q. a...

    2024

  16. [16]

    and Amematsro, Elom A

    Perkins, Sean M. and Amematsro, Elom A. and Cunningham, John P. and Wang, Qi and Churchland, Mark M. , title =. eLife , year =

  17. [17]

    and Hennig, Matthias H

    Kudryashova, Nina and Hurwitz, Cole and Perich, Matthew G. and Hennig, Matthias H. , title =. bioRxiv , year =

  18. [18]

    and Moorman, Helene G

    Orsborn, Amy L. and Moorman, Helene G. and Overduin, Simon A. and Shanechi, Maryam M. and Dimitrov, Dimitar F. and Carmena, Jose M. , title =. Neuron , year =

  19. [19]

    , title =

    Bishop, Christopher M. , title =

  20. [20]

    2016 , userd =

    周志华 , title =. 2016 , userd =

    2016

  21. [21]

    and Cunningham, John P

    Yu, Byron M. and Cunningham, John P. and Santhanam, Gopal and Ryu, Stephen I. and Shenoy, Krishna V. and Sahani, Maneesh , title =. Journal of Neurophysiology , year =

  22. [22]

    , title =

    Ghahramani, Zoubin and Hinton, Geoffrey E. , title =

  23. [23]

    , title =

    Ghahramani, Zoubin and Hinton, Geoffrey E. , title =. Neural Computation , year =

  24. [24]

    and Cunningham, John P

    Churchland, Mark M. and Cunningham, John P. and Kaufman, Matthew T. and Foster, Justin D. and Nuyujukian, Paul and Ryu, Stephen I. and Shenoy, Krishna V. , title =. Nature , year =

  25. [25]

    and Abbaspourazad, Hamidreza and Wong, Yan T

    Sani, Omid G. and Abbaspourazad, Hamidreza and Wong, Yan T. and Pesaran, Bijan and Shanechi, Maryam M. , title =. Nature Neuroscience , year =

  26. [26]

    and Johnson, Matthew J

    Linderman, Scott W. and Johnson, Matthew J. and Miller, Andrew C. and Adams, Ryan P. and Blei, David M. and Paninski, Liam , title =. Proceedings of the 20th International Conference on Artificial Intelligence and Statistics , volume =. 2017 , pages =

    2017

  27. [27]

    and Welling, Max , title =

    Kingma, Diederik P. and Welling, Max , title =. International Conference on Learning Representations , location =

  28. [28]

    and Collins, Jasmine and Jozefowicz, Rafal and Stavisky, Sergey D

    Pandarinath, Chethan and O'Shea, Daniel J. and Collins, Jasmine and Jozefowicz, Rafal and Stavisky, Sergey D. and Kao, Jonathan C. and Trautmann, Eric M. and Kaufman, Matthew T. and Ryu, Stephen I. and Hochberg, Leigh R. and Henderson, Jaimie M. and Shenoy, Krishna V. and Abbott, L. F. and Sussillo, David , title =. Nature Methods , year =

  29. [29]

    and Pandarinath, Chethan , title =

    Sedler, Andrew R. and Pandarinath, Chethan , title =. arXiv preprint arXiv:2309.01230 , year =

    work page arXiv

  30. [30]

    and Chowdhury, Raeed H

    Keshtkaran, Mohammad Reza and Sedler, Andrew R. and Chowdhury, Raeed H. and Tandon, Raghav and Basrai, Diya and Nguyen, Sarah L. and Sohn, Hansem and Jazayeri, Mehrdad and Miller, Lee E. and Pandarinath, Chethan , title =. Nature Methods , year =

  31. [31]

    and Kaiser, Lukasz and Polosukhin, Illia , title =

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N. and Kaiser, Lukasz and Polosukhin, Illia , title =. Advances in Neural Information Processing Systems , volume =. 2017 , pages =

    2017

  32. [32]

    and Lewis, Mike , title =

    Press, Ofir and Smith, Noah A. and Lewis, Mike , title =. International Conference on Learning Representations , location =

  33. [33]

    Advances in Neural Information Processing Systems , volume =

    Qiu, Zihan and Wang, Zekun and Zheng, Bo and Huang, Zeyu and Wen, Kaiyue and Yang, Songlin and Men, Rui and Yu, Le and Huang, Fei and Huang, Suozhi and Liu, Dayiheng and Zhou, Jingren and Lin, Junyang , title =. Advances in Neural Information Processing Systems , volume =. 2025 , location =

    2025

  34. [34]

    Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , volume =

    Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina , title =. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , volume =

    2019

  35. [35]

    Neurons, Behavior, Data Analysis, and Theory , year =

    Ye, Joel and Pandarinath, Chethan , title =. Neurons, Behavior, Data Analysis, and Theory , year =

  36. [36]

    , title =

    Mifsud, Domenick M. , title =

  37. [37]

    and Sohn, Hansem and O'Doherty, Joseph E

    Pei, Felix and Ye, Joel and Zoltowski, David and Wu, Anqi and Chowdhury, Raeed H. and Sohn, Hansem and O'Doherty, Joseph E. and Shenoy, Krishna V. and Kaufman, Matthew T. and Churchland, Mark M. and Jazayeri, Mehrdad and Miller, Lee E. and Pillow, Jonathan and Park, Il Memming and Dyer, Eva L. and Sadtler, Patrick T. and Batista, Aaron P. and Cunningham, ...

  38. [38]

    Advances in Neural Information Processing Systems , volume =

    Le, Trung and Shlizerman, Eli , title =. Advances in Neural Information Processing Systems , volume =

  39. [39]

    and Stoica, Ion , title =

    Moritz, Philipp and Nishihara, Robert and Wang, Stephanie and Tumanov, Alexey and Liaw, Richard and Liang, Eric and Elibol, Melih and Yang, Zongheng and Paul, William and Jordan, Michael I. and Stoica, Ion , title =. Proceedings of the 13th USENIX Symposium on Operating Systems Design and Implementation , pages =

  40. [40]

    Proceedings of the 37th International Conference on Machine Learning , volume =

    Chen, Ting and Kornblith, Simon and Norouzi, Mohammad and Hinton, Geoffrey , title =. Proceedings of the 37th International Conference on Machine Learning , volume =

  41. [41]

    Proceedings of the 37th International Conference on Machine Learning , volume =

    Huang, Xiao Shi and Perez, Felipe and Ba, Jimmy and Volkovs, Maksims , title =. Proceedings of the 37th International Conference on Machine Learning , volume =

  42. [42]

    Efficiently Modeling Long Sequences with Structured State Spaces , booktitle =

    Gu, Albert and Goel, Karan and R. Efficiently Modeling Long Sequences with Structured State Spaces , booktitle =

  43. [43]

    Smith, Jimmy T. H. and Warrington, Andrew and Linderman, Scott W. , title =. International Conference on Learning Representations , location =

  44. [44]

    Lipman, Yaron and Chen, Ricky T. Q. and Ben-Hamu, Heli and Nickel, Maximilian and Le, Matthew , title =. International Conference on Learning Representations , location =

  45. [45]

    , title =

    Kapoor, Jaivardhan and Schulz, Auguste and Vetter, Julius and Pei, Felix and Gao, Richard and Macke, Jakob H. , title =. Advances in Neural Information Processing Systems , volume =

  46. [46]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Ge, Ningling and Dai, Sicheng and Zhu, Yu and Yu, Shan , title =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =

  47. [47]

    Proceedings of the 42nd International Conference on Machine Learning , volume =

    Wang, Puli and Qi, Yu and Wang, Yueming and Pan, Gang , title =. Proceedings of the 42nd International Conference on Machine Learning , volume =

  48. [48]

    2026 , url =

    Wang, Puli and Qi, Yu and Wang, Yueming and Pan, Gang , title =. 2026 , url =

    2026

  49. [49]

    NLB'21 Challenge Info and Local Evaluation Data Release , year =