pith. machine review for the scientific record. sign in

arxiv: 2605.04194 · v1 · submitted 2026-05-05 · 💻 cs.CY

Recognition: 3 theorem links

· Lean Theorem

Coupled-NeuralHP: Directional Temporal Coupling Between AI Innovation Exposure and Public Response

Amir Rafe , Subasish Das
Authors on Pith no claims yet

Pith reviewed 2026-05-08 17:50 UTC · model grok-4.3

classification 💻 cs.CY
keywords AI patentsGoogle Trendstemporal couplinghybrid neural modelpublic responseinnovation forecastingUSPTO data
0
0 comments X

The pith

A hybrid neural model finds one-way coupling from AI patent streams to public response, with superior held-out forecasts but no support for the reverse direction.

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

The paper introduces Coupled-NeuralHP to link irregular USPTO AI patent publication events with monthly Google Trends response data through a hybrid event-state structure. It compares model variants under a cleaned response protocol and finds that the validation-selected one-way version, where innovation exposure influences response, delivers the strongest predictions for future innovation counts. This outperforms several baselines on innovation forecasting metrics while matching the best performer on response error, and semi-synthetic tests confirm better recovery of true directional links than vector autoregression. Ablations isolate the predictive contribution to the structured forecast head, and a placebo split-date check finds no clear 2022 milestone break.

Core claim

Under the cleaned response protocol, the validation-selected one-way real-data variant of Coupled-NeuralHP gives the best held-out innovation count forecasts in the registered comparison set (pseudo-log-likelihood -30.4 versus -34.7; RMSE 471 versus 532) while matching the stronger multi-lag factor-family baseline on response RMSE (0.295); the reverse response-to-innovation block receives no support on held-out count prediction, and the broader coupled family recovers known innovation-to-response links better than VARX on semi-synthetic data.

What carries the argument

Coupled-NeuralHP, a hybrid event-plus-state model that couples patent event streams to a response index through neural components for directional temporal forecasting.

If this is right

  • Innovation count forecasts can be improved by incorporating one-way response signals from cleaned trend data.
  • The structured forecast head carries the main contribution from real response data to innovation prediction.
  • No evidence supports a reverse response-to-innovation predictive link on held-out tests.
  • The coupling structure shows no robust regime break at the 2022 split date in placebo analysis.

Where Pith is reading between the lines

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

  • Similar directional models could be tested on other technology domains to check whether innovation exposure consistently precedes public interest signals.
  • Adding alternative response proxies such as social media volume might reduce reliance on a single search index.
  • If the one-way pattern holds, timing of patent announcements could be adjusted to anticipate measurable public interest spikes.

Load-bearing premise

The cleaned Google Trends index serves as a reliable, unbiased proxy for public response to AI innovation exposure.

What would settle it

New held-out data in which a bidirectional or reverse-only variant of the model outperforms the one-way real-data version on innovation count prediction would falsify the directional claim.

Figures

Figures reproduced from arXiv: 2605.04194 by Amir Rafe, Subasish Das.

Figure 1
Figure 1. Figure 1: Coupled-NeuralHP architecture overview. Innovation event streams (USPTO AIPD, eight AI technology components) feed a multivariate Hawkes process block, while monthly public response (Google Trends PCA index) evolves through a latent state-space block. Sparse hard-concrete gates control directional coupling: innovation-to-response (B, gated) and response-to-innovation (γ, gated). A structured response forec… view at source ↗
Figure 2
Figure 2. Figure 2: Held-out 2023 model comparison. Left: count pseudo-log-likelihood (higher is better). Right: response RMSE (lower is better). The selected coupled model leads the held-out benchmark on the count side and matches the stronger multi-lag factor-family baseline on response. boundary by six months), the directional asymmetry is maintained. Within the forward block, natural language processing (NLP), speech, and… view at source ↗
Figure 3
Figure 3. Figure 3: Regime analysis: actual milestones vs. placebo months. Diamonds indicate milestones; circles indicate placebos. No milestone consistently outranks the placebo distribution in likelihood gain (left), supporting a gradualist interpretation. 7 view at source ↗
Figure 4
Figure 4. Figure 4: Semi-synthetic directional recovery (60 replications). Coupled-NeuralHP (orange) vs. VARX (blue) on I→R F1. months of coupled innovation/response data from a known data-generating process, fits both Coupled￾NeuralHP and a VARX baseline, and evaluates directional recovery as the F1 score between recovered and true non-zero coupling entries (generation protocol in Appendix F). Coupled-NeuralHP recovers innov… view at source ↗
Figure 5
Figure 5. Figure 5: Ablation forest plot. Point estimates and 95% bootstrap confidence intervals for response RMSE changes under each ablation. Removing the response head produces the largest degradation. rather than unique superiority. The clearest mechanism result is that the structured response forecast head matters for held-out response prediction, while forcing the reverse response-to-innovation block harms count predict… view at source ↗
Figure 6
Figure 6. Figure 6: Rolling-window coupling stability. Innovation-to-response and response-to-innovation gate densities across four rolling temporal windows, each shifting the train/test boundary by six months. The one-directional coupling structure (I→R = 0.5, R→I = 0.0) is preserved across all windows. E Ablation delta intervals view at source ↗
Figure 7
Figure 7. Figure 7: Metric regret analysis. Held-out metric regret (gap to best model on each metric) across all model variants. The tuned coupled model achieves zero or near-zero regret on both count and response metrics simultaneously. F Semi-synthetic recovery details Each of the 60 semi-synthetic replications generates 120 months of coupled innovation/response data. The data-generating process specifies ground-truth innov… view at source ↗
Figure 8
Figure 8. Figure 8: Placebo analysis summary. Distribution of joint gain scores for placebo months compared to actual milestone months. The overlap between distributions confirms the absence of detectable structural breaks at milestone events. 16 view at source ↗
read the original abstract

Artificial intelligence innovation exposure and public response co-evolve, but innovation arrives as irregular event streams while response is observed monthly. We introduce Coupled-NeuralHP, a hybrid event-plus-state model linking eight-domain USPTO AI patent publication streams to a train-only Google Trends response index. Under the cleaned response protocol, the validation-selected one-way real-data variant gives the best held-out innovation count forecasts in the registered comparison set (pseudo-log-likelihood -30.4 vs. -34.7; root mean squared error (RMSE) 471 vs. 532) while matching the stronger multi-lag factor-family baseline on response RMSE (0.295). Ablations show that the real-data response signal is carried mainly by the structured forecast head, whereas the reverse response-to-innovation block is not supported on held-out count prediction. Across 60 semi-synthetic replications with known structure, the broader coupled family recovers innovation-to-response links much better than vector autoregression with exogenous inputs (VARX) (F1 = 0.734 vs. 0.386). A placebo-controlled 2022 split-date analysis finds no robust milestone-specific regime break.

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 / 2 minor

Summary. The manuscript introduces Coupled-NeuralHP, a hybrid event-plus-state model linking irregular eight-domain USPTO AI patent publication streams to a monthly Google Trends response index under a train-only protocol. It reports that the validation-selected one-way real-data variant yields the best held-out innovation-count forecasts among registered comparators (pseudo-log-likelihood -30.4 vs. -34.7; RMSE 471 vs. 532) while matching the stronger multi-lag factor-family baseline on response RMSE (0.295). Ablations indicate the real-data signal is carried primarily by the structured forecast head; the reverse response-to-innovation block is unsupported on held-out counts. Across 60 semi-synthetic replications with known structure the coupled family recovers innovation-to-response links better than VARX (F1 0.734 vs. 0.386). A placebo 2022 split-date analysis finds no robust milestone-specific regime break.

Significance. If the results hold, the work supplies a novel hybrid architecture for directional temporal coupling between irregular innovation events and continuous public-response signals, together with concrete held-out metrics, ablations, and 60 semi-synthetic replications that furnish falsifiable grounding for the one-way claim. These elements strengthen the empirical case that AI patent exposure drives public attention without detectable reverse causality.

major comments (2)
  1. [§2] §2 (Data and cleaning protocol): The cleaned Google Trends index is load-bearing for the headline directional result, yet the manuscript provides no explicit demonstration that the cleaning isolates patent-driven response from external-event confounding (media cycles, unrelated news spikes). The reported placebo 2022 split and semi-synthetic recovery (F1 0.734) do not directly test whether residual correlation from non-patent shocks drives the held-out gains (-30.4 pseudo-log-likelihood).
  2. [§3.3] §3.3 (Validation selection and one-way architecture): The validation-selected one-way variant is claimed to isolate true directional coupling, but the combination of train-only fitting and post-hoc validation selection risks parameter correlation with the response signal. Without additional diagnostics showing that this selection does not inflate apparent innovation-to-response performance relative to the reverse block, the claim that the reverse direction is unsupported remains vulnerable.
minor comments (2)
  1. [§3.1] The notation for the hybrid event-plus-state blocks would be clearer with an explicit equation or diagram distinguishing the one-way forecast head from the coupled variant.
  2. [Table 2] Table 2 (semi-synthetic results): confirm that all VARX baselines use identical lag structure and exogenous inputs as the Coupled-NeuralHP variants for direct comparability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major point below with clarifications and indicate where the manuscript will be revised to incorporate additional evidence and diagnostics.

read point-by-point responses

  1. Referee: [§2] §2 (Data and cleaning protocol): The cleaned Google Trends index is load-bearing for the headline directional result, yet the manuscript provides no explicit demonstration that the cleaning isolates patent-driven response from external-event confounding (media cycles, unrelated news spikes). The reported placebo 2022 split and semi-synthetic recovery (F1 0.734) do not directly test whether residual correlation from non-patent shocks drives the held-out gains (-30.4 pseudo-log-likelihood).

    Authors: We agree that an explicit demonstration of isolation from non-patent confounding is necessary to support the directional claim. The cleaning protocol, as described in §2, removes documented spikes attributable to non-AI events using a predefined rule set based on external event logs. However, to directly address residual correlation concerns, we will add a new appendix containing: (i) side-by-side held-out performance metrics on raw versus cleaned response series, and (ii) an extended robustness check that augments the model with binary indicators for major media events unrelated to AI. The semi-synthetic replications test recovery of known directional structure, while the 2022 placebo tests for regime breaks; these do not substitute for the requested confounding test. We will therefore revise the manuscript to include the additional controls and comparisons. revision: yes

  2. Referee: [§3.3] §3.3 (Validation selection and one-way architecture): The validation-selected one-way variant is claimed to isolate true directional coupling, but the combination of train-only fitting and post-hoc validation selection risks parameter correlation with the response signal. Without additional diagnostics showing that this selection does not inflate apparent innovation-to-response performance relative to the reverse block, the claim that the reverse direction is unsupported remains vulnerable.

    Authors: The protocol fits all parameters exclusively on the training partition and uses a disjoint validation set solely for architecture selection (one-way versus coupled). This design prevents direct leakage into test evaluation. Nevertheless, we acknowledge the referee's point that post-hoc selection could still favor the innovation-to-response direction. In the revision we will add explicit diagnostics: (i) held-out metrics for the reverse block under identical validation selection, (ii) a comparison of one-way performance when selection is replaced by a fixed a-priori choice, and (iii) a brief sensitivity table across multiple validation splits. These additions will allow readers to assess whether selection inflates the reported asymmetry. We maintain that the current train-only protocol already limits correlation, but the requested diagnostics will be incorporated. revision: yes

Circularity Check

0 steps flagged

No significant circularity; held-out and semi-synthetic checks are independent of fitting inputs

full rationale

The reported performance (pseudo-log-likelihood -30.4, RMSE 471 on held-out innovation counts; F1 0.734 on 60 semi-synthetic replications) is obtained after a train-only protocol with separate validation selection and placebo split-date analysis. These evaluations use data partitions and generated data with known structure that are not equivalent to the fitted parameters or the cleaned Google Trends index by construction. No self-definitional equations, fitted inputs renamed as predictions, or load-bearing self-citations appear in the abstract or described protocol; the one-way architecture and cleaning step are design choices whose value is assessed via external recovery metrics rather than reducing tautologically to the inputs.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 1 invented entities

The central claim rests on standard neural network training assumptions plus domain-specific choices about data cleaning and proxy validity; no new physical entities are postulated.

free parameters (1)
  • neural network weights and hyperparameters
    Fitted during training on the response index under the train-only protocol; exact count and values not stated in abstract.
axioms (2)
  • domain assumption Google Trends index after cleaning accurately reflects public response to AI innovation exposure.
    Invoked to justify linking patent streams to the response variable.
  • domain assumption The validation selection procedure identifies the true directional structure without overfitting to noise.
    Used to choose the one-way real-data variant as best.
invented entities (1)
  • Coupled-NeuralHP hybrid architecture no independent evidence
    purpose: To jointly model irregular event streams and monthly state observations with directional blocks.
    Newly proposed model family; no independent evidence outside the paper's own tests.

pith-pipeline@v0.9.0 · 5499 in / 1703 out tokens · 73705 ms · 2026-05-08T17:50:09.832634+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

51 extracted references · 5 canonical work pages · 1 internal anchor

  1. [1]

    Spectra of some self-exciting and mutually exciting point processes,

    Hawkes, Alan G. , year =. Spectra of Some Self-Exciting and Mutually Exciting Point Processes , volume =. Biometrika , publisher =. doi:10.1093/biomet/58.1.83 , number =

    work page doi:10.1093/biomet/58.1.83

  2. [2]

    , booktitle =

    Mei, Hongyuan and Eisner, Jason M. , booktitle =. The Neural Hawkes Process: A Neurally Self-Modulating Multivariate Point Process , url =

  3. [3]

    Recurrent Marked Temporal Point Processes: Embedding Event History to Vector , url =

    Du, Nan and Dai, Hanjun and Trivedi, Rakshit and Upadhyay, Utkarsh and Gomez-Rodriguez, Manuel and Song, Le , year =. Recurrent Marked Temporal Point Processes: Embedding Event History to Vector , url =. doi:10.1145/2939672.2939875 , booktitle =

    work page doi:10.1145/2939672.2939875

  4. [4]

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

    Transformer Hawkes Process , author =. Proceedings of the 37th International Conference on Machine Learning , series =. 2020 , publisher =

    2020

  5. [5]

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

    Self-Attentive Hawkes Process , author =. Proceedings of the 37th International Conference on Machine Learning , series =. 2020 , publisher =

    2020

  6. [6]

    , booktitle =

    Louizos, Christos and Welling, Max and Kingma, Diederik P. , booktitle =. Learning Sparse Neural Networks through. 2018 , url =

    2018

  7. [7]

    Uncovering Causality from Multivariate

    Achab, Massil and Bacry, Emmanuel and Ga. Uncovering Causality from Multivariate. Proceedings of the 34th International Conference on Machine Learning , pages =. 2017 , publisher =

    2017

  8. [8]

    Learning Granger Causality for

    Xu, Hongteng and Farajtabar, Mehrdad and Zha, Hongyuan , booktitle =. Learning Granger Causality for. 2016 , publisher =

    2016

  9. [9]

    2026 , howpublished =

    Artificial Intelligence Patent Dataset , author =. 2026 , howpublished =

    2026

  10. [10]

    Predicting the Present with

    Choi, Hyunyoung and Varian, Hal , journal =. Predicting the Present with. 2012 , doi =

    2012

  11. [11]

    Ten Years of Research Change Using

    Jun, Seung-Pyo and Yoo, Hyoung Sun and Choi, San , journal =. Ten Years of Research Change Using. 2018 , doi =

    2018

  12. [12]

    2023 , howpublished =

    Growing Public Concern About the Role of Artificial Intelligence in Daily Life , author =. 2023 , howpublished =

    2023

  13. [13]

    Artificial Intelligence:

    Zhang, Baobao and Dafoe, Allan , institution =. Artificial Intelligence:. 2019 , url =

    2019

  14. [14]

    Management Science , volume =

    A New Product Growth for Model Consumer Durables , author =. Management Science , volume =. 1969 , doi =

    1969

  15. [15]

    2012 , publisher =

    Time Series Analysis by State Space Methods , author =. 2012 , publisher =

    2012

  16. [16]

    , year =

    Harvey, Andrew C. , year =. Forecasting, Structural Time Series Models and the

  17. [17]

    Statistical Science , volume =

    A Review of Self-Exciting Spatio-Temporal Point Processes and Their Applications , author =. Statistical Science , volume =. 2018 , doi =

    2018

  18. [18]

    arXiv preprint arXiv:1806.00221 , year =

    Temporal Point Processes and the Conditional Intensity Function , author =. arXiv preprint arXiv:1806.00221 , year =

    work page arXiv

  19. [19]

    and Kaiser,

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N. and Kaiser,. Attention is All you Need , url =. Advances in Neural Information Processing Systems , editor =

  20. [20]

    and Welling, Max , journal =

    Kingma, Diederik P. and Welling, Max , journal =. Auto-Encoding Variational

  21. [21]

    Nature Machine Intelligence , volume =

    Hopes and Fears for Intelligent Machines in Fiction and Reality , author =. Nature Machine Intelligence , volume =. 2019 , doi =

    2019

  22. [22]

    Graphical Modeling for Multivariate

    Eichler, Michael and Dahlhaus, Rainer and Dueck, Johannes , journal =. Graphical Modeling for Multivariate. 2017 , doi =

    2017

  23. [23]

    International Conference on Learning Representations (ICLR) , year =

    Intensity-Free Learning of Temporal Point Processes , author =. International Conference on Learning Representations (ICLR) , year =

  24. [24]

    International Joint Conference on Artificial Intelligence (IJCAI) , year =

    Neural Temporal Point Processes: A Review , author =. International Joint Conference on Artificial Intelligence (IJCAI) , year =

  25. [25]

    Advances in Neural Information Processing Systems (NeurIPS) , volume =

    Fully Neural Network based Model for General Temporal Point Processes , author =. Advances in Neural Information Processing Systems (NeurIPS) , volume =

  26. [26]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year =

    Neural Granger Causality , author =. IEEE Transactions on Pattern Analysis and Machine Intelligence , year =

  27. [27]

    Deep Kalman Filters

    Deep Kalman Filters , author =. arXiv preprint arXiv:1511.05121 , year =

    work page Pith review arXiv

  28. [28]

    Advances in Neural Information Processing Systems (NeurIPS) , volume =

    Deep State Space Models for Time Series Forecasting , author =. Advances in Neural Information Processing Systems (NeurIPS) , volume =

  29. [29]

    De Brouwer, Edward and Simm, Jaak and Arany, Adam and Moreau, Yves , booktitle =

  30. [30]

    Diffusion of Innovations , author =

  31. [31]

    and Jaffe, Adam B

    Hall, Bronwyn H. and Jaffe, Adam B. and Trajtenberg, Manuel , institution =. The. 2001 , url =

    2001

  32. [32]

    Economics of Innovation and New Technology , volume =

    International Knowledge Flows: Evidence from Patent Citations , author =. Economics of Innovation and New Technology , volume =. 1999 , doi =

    1999

  33. [33]

    2019 , doi =

    Mavragani, Amaryllis and Ochoa, Gabriela , journal =. 2019 , doi =

    2019

  34. [34]

    International Conference on Machine Learning (ICML) , year =

    Discovering Latent Network Structure in Point Process Data , author =. International Conference on Machine Learning (ICML) , year =

  35. [35]

    Farajtabar, Mehrdad and Wang, Yichen and Gomez Rodriguez, Manuel and Li, Shuang and Zha, Hongyuan and Song, Le , booktitle =

  36. [36]

    and He, Hera Y

    Zhao, Qingyuan and Erdogdu, Murat A. and He, Hera Y. and Rajaraman, Anand and Leskovec, Jure , booktitle =. 2015 , doi =

    2015

  37. [37]

    Market Microstructure and Liquidity , volume =

    Hawkes Processes in Finance , author =. Market Microstructure and Liquidity , volume =. 2015 , doi =

    2015

  38. [38]

    Journal of the American Statistical Association , volume =

    Statistical Models for Earthquake Occurrences and Residual Analysis for Point Processes , author =. Journal of the American Statistical Association , volume =. 1988 , doi =

    1988

  39. [39]

    Daley, D. J. and Vere-Jones, D. , edition =. An Introduction to the Theory of Point Processes: Volume. 2003 , publisher =

    2003

  40. [40]

    Advances in Neural Information Processing Systems , volume =

    Neural Ordinary Differential Equations , author =. Advances in Neural Information Processing Systems , volume =

  41. [41]

    Advances in Neural Information Processing Systems , volume =

    Latent Ordinary Differential Equations for Irregularly-Sampled Time Series , author =. Advances in Neural Information Processing Systems , volume =

  42. [42]

    Econometrica , volume =

    Investigating Causal Relations by Econometric Models and Cross-spectral Methods , author =. Econometrica , volume =. 1969 , doi =

    1969

  43. [43]

    Learning Social Infectivity in Sparse Low-rank Networks Using Multi-dimensional

    Zhou, Ke and Zha, Hongyuan and Song, Le , booktitle =. Learning Social Infectivity in Sparse Low-rank Networks Using Multi-dimensional. 2013 , publisher =

    2013

  44. [44]

    The Economics of Artificial Intelligence: An Agenda , pages =

    The Impact of Artificial Intelligence on Innovation: An Exploratory Analysis , author =. The Economics of Artificial Intelligence: An Agenda , pages =. 2018 , publisher =

    2018

  45. [45]

    , journal =

    Brynjolfsson, Erik and Li, Danielle and Raymond, Lindsey R. , journal =. Generative. 2023 , note =

    2023

  46. [46]

    2023 , type =

    Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence , author =. 2023 , type =

    2023

  47. [47]

    and Wen, Qingsong and Zhou, Jun and Mei, Hongyuan , booktitle =

    Xue, Siqiao and Shi, Xiaoming and Chu, Zhixuan and Wang, Yan and Hao, Hongyan and Zhou, Fan and Jiang, Caigao and Pan, Chen and Zhang, James Y. and Wen, Qingsong and Zhou, Jun and Mei, Hongyuan , booktitle =

  48. [48]

    Mamba: Linear-Time Sequence Modeling with Selective State Spaces

    Mamba: Linear-Time Sequence Modeling with Selective State Spaces , author =. arXiv preprint arXiv:2312.00752 , year =

    work page internal anchor Pith review arXiv

  49. [49]

    Transformers are

    Dao, Tri and Gu, Albert , booktitle =. Transformers are

  50. [50]

    Kennedy, Brian and Yam, Eileen and Kikuchi, Emma and Pula, Isabelle and Fuentes, Javier , year =. How

  51. [51]

    2025 , note =

    Artificial Intelligence Index Report 2025 , author =. 2025 , note =

    2025