pith. sign in

arxiv: 2606.00047 · v1 · pith:E63YJWLYnew · submitted 2026-05-01 · 💻 cs.CY · cs.AI

Comprehensive AI governance requires addressing non-model gains

Arthur Goemans , Dan Altman , Noemi Dreksler , Jonas Freund , Milan Gandhi , Zhengdong Wang , Sarah Cogan , Sebastien Krier
show 3 more authors
Demetra Brady Lewis Ho Allan Dafoe
This is my paper

Pith reviewed 2026-07-01 08:09 UTC · model grok-4.3

classification 💻 cs.CY cs.AI
keywords AI governancenon-model gainsmodel-level governanceinference gainsystems gainasset gainAI risk management
0
0 comments X

The pith

Model-level AI governance loses effectiveness as non-model gains increasingly drive capability progress.

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

The paper tries to establish that the model-level governance paradigm, which focuses on base models trained with certain compute and data, will become less effective as capability progress shifts to non-model gains. It formalizes non-model gains and taxonomizes them into inference gain, systems gain, and asset gain, showing how they can bypass pre-deployment controls. As a result, governance needs to incorporate system, entity, agent, and cloud approaches along with building societal resilience. A reader would care because this changes the focus from controlling models at deployment to managing a wider set of post-training and deployment factors that affect AI capabilities.

Core claim

Model-level governance becomes less effective when capability progress is increasingly driven by non-model gains, formalized as a taxonomy of inference gain, systems gain, and asset gain, which may undermine pre-deployment evaluation and mitigation, thus requiring system, entity, agent, and cloud governance plus societal resilience.

What carries the argument

Non-model gains, with the three-vector taxonomy of inference gain (scaling compute at test-time), systems gain (post-training enhancements such as scaffolds), and asset gain (enhancing a model with restricted assets).

Load-bearing premise

Non-model gains will increasingly dominate over model advances in driving capability progress.

What would settle it

Empirical measurements showing that base model training still accounts for the large majority of capability improvements despite advances in inference compute, scaffolds, and restricted assets.

read the original abstract

Frontier AI governance often centres on the model-level governance paradigm, which assumes that a model's capability profile is primarily a function of the compute and data used during training. This position paper argues that model-level governance becomes less effective when capability progress is increasingly driven by "non-model gains"--improvements that are independent from advances in the base model. We formalise the concept of non-model gains and provide a taxonomy of three distinct vectors of capability gain: inference gain (scaling compute at test-time), systems gain (post-training enhancements such as scaffolds), and asset gain (enhancing a model with restricted assets). We demonstrate how these vectors--alongside potential future impacts from embodiment, continual learning, and AI diffusion--may undermine risk management strategies that hinge mostly on pre-deployment evaluation and mitigation. We provide an overview of governance approaches that go beyond the model level: system, entity, agent, and cloud governance. Finally, we emphasise the importance of societal resilience as a complement to these governance layers.

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

0 major / 2 minor

Summary. The paper claims that frontier AI governance centered on the model-level paradigm—which assumes a model's capabilities are primarily determined by training compute and data—becomes less effective when progress is increasingly driven by non-model gains independent of the base model. It formalizes non-model gains via a taxonomy of three vectors (inference gain via test-time compute scaling, systems gain via post-training scaffolds, and asset gain via restricted assets), argues these (plus embodiment, continual learning, and diffusion) may undermine pre-deployment evaluation and mitigation strategies, and outlines complementary governance layers at the system, entity, agent, and cloud levels alongside societal resilience.

Significance. If the conditional argument holds, the work is significant for broadening AI governance discussions beyond model-centric approaches by supplying a clear taxonomy of post-training capability vectors and an overview of multi-layered alternatives. This provides a useful conceptual scaffold for policy analysis in a domain where deployment and integration factors are gaining prominence, without overclaiming empirical dominance.

minor comments (2)
  1. The demonstration of how the three vectors undermine pre-deployment strategies (mentioned in the abstract) would benefit from one or two concrete, referenced examples per vector to make the conditional claim more actionable for readers.
  2. The manuscript would be strengthened by explicit citations to prior work on inference scaling, scaffolding, or entity-level governance in the sections introducing the taxonomy and alternative approaches.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive assessment of the manuscript's significance and for recommending minor revision. The referee's summary accurately captures the core argument regarding the limitations of model-level governance in the presence of non-model gains. No specific major comments were listed in the report, so we interpret the minor revision request as pertaining to editorial or presentational improvements.

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

This is a position paper that defines non-model gains via a taxonomy of three vectors (inference, systems, asset) and argues conditionally that model-level governance may lose effectiveness if those vectors increasingly drive progress. No equations, fitted parameters, quantitative predictions, or derivations appear. The central claims rest on external observations of AI trends and a conditional 'may' framing rather than any self-referential definitions or self-citation load-bearing steps. The paper is self-contained against external benchmarks with no reduction of outputs to inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on a domain assumption about the trajectory of capability progress and introduces a new conceptual category without independent empirical support in the abstract.

axioms (1)
  • domain assumption Capability progress can increasingly be driven by factors independent of base model advances.
    This premise is required for the claim that model-level governance will become less effective.
invented entities (1)
  • non-model gains no independent evidence
    purpose: To categorize and highlight post-training sources of capability improvement.
    New framing device introduced to diagnose limitations in existing governance approaches.

pith-pipeline@v0.9.1-grok · 5728 in / 1142 out tokens · 32555 ms · 2026-07-01T08:09:30.931916+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

95 extracted references · 35 canonical work pages · 8 internal anchors

  1. [1]

    and others , title =

    Addlesee, A. and others , title =. Proceedings of the 2024 ACM/IEEE International Conference on Human-Robot Interaction (HRI '24) , year =

    2024

  2. [2]

    2024 , note =

    Anduril Partners with. 2024 , note =

    2024

  3. [3]

    2023 , address =

    Anthropic's Responsible Scaling Policy , institution =. 2023 , address =

    2023

  4. [4]

    2026 , address =

    Disrupting the first reported. 2026 , address =

    2026

  5. [5]

    2025 , note =

    Claude Gov Models for. 2025 , note =

    2025

  6. [6]

    2025 , note =

    Anthropic and the. 2025 , note =

    2025

  7. [7]

    2025 , note =

    Strengthening our safeguards through collaboration with. 2025 , note =

    2025

  8. [8]

    and others , title =

    Bowman, S. and others , title =. 2025 , note =

    2025

  9. [9]

    arXiv preprint arXiv:2102.06701 , year=

    Bahri, Y. and others , title =. arXiv preprint arXiv:2102.06701 , year =

    work page arXiv

  10. [10]

    and others , title =

    Baker, M. and others , title =. arXiv preprint arXiv:2507.15916 , year =

    work page arXiv

  11. [11]

    and others , title =

    Behrouz, A. and others , title =. 2025 , note =

    2025

  12. [12]

    and others , title =

    Bengio, Y. and others , title =. arXiv preprint arXiv:2506.20702 , year =

    work page arXiv

  13. [13]

    and others , title =

    Bernardi, J. and others , title =. arXiv preprint arXiv:2405.10295 , year =

    work page arXiv

  14. [14]

    and others , title =

    Bian, S. and others , title =. arXiv preprint arXiv:2501.18107 , year =

    work page arXiv

  15. [15]

    and others , title =

    Bommasani, R. and others , title =. Advances in Neural Information Processing Systems 35 (NeurIPS 2022) , year =

    2022

  16. [16]

    Robots That Can Chat , year =

  17. [17]

    and others , title =

    Brundage, M. and others , title =. arXiv preprint arXiv:2601.11699 , year =

    work page arXiv

  18. [18]

    and Ramakrishnan, K

    Ball, D. and Ramakrishnan, K. , title =. 2025 , address =

    2025

  19. [19]

    and others , title =

    Chan, A. and others , title =. arXiv preprint arXiv:2501.10114 , year =

    work page arXiv

  20. [20]

    Proceedings of the 33rd International Joint Conference on Artificial Intelligence (IJCAI 2024) , year =

    Cohen, Vanya and others , title =. Proceedings of the 33rd International Joint Conference on Artificial Intelligence (IJCAI 2024) , year =

    2024

  21. [21]

    A Technical Analysis of Confidential Computing (v1.3) , institution =

  22. [22]

    Enterprise Risk Management Framework , year =

  23. [23]

    and others , title =

    Davidson, T. and others , title =. arXiv preprint arXiv:2312.07413 , year =

    work page arXiv

  24. [24]

    2025 , address =

    Frontier Safety Framework 3.0 , institution =. 2025 , address =

    2025

  25. [25]

    2025 , note =

    Gemini Robotics 1.5 brings. 2025 , note =

    2025

  26. [26]

    2025 , note =

    Taking a responsible path to. 2025 , note =

    2025

  27. [27]

    arXiv preprint arXiv:2501.12948 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  28. [28]

    arXiv preprint arXiv:2512.02556 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  29. [29]

    and others , title =

    Driess, D. and others , title =. Proceedings of the 40th International Conference on Machine Learning (ICML 2023) , year =

    2023

  30. [30]

    and others , title =

    Du, X. and others , title =. arXiv preprint arXiv:2406.11147 , year =

    work page arXiv

  31. [31]

    , title =

    Erdil, E. , title =. 2024 , note =

    2024

  32. [32]

    and others , title =

    Somala, V. and others , title =. 2025 , note =

    2025

  33. [33]

    and Atkinson, D

    Villalobos, P. and Atkinson, D. , title =. 2023 , note =

    2023

  34. [34]

    and Denain, JS , title =

    Brand, F. and Denain, JS , title =. 2025 , note =

    2025

  35. [35]

    Regulation (EU) 2024/1689 of the European Parliament and of the Council (Artificial Intelligence Act) , year =

    2024

  36. [36]

    2025 , note =

    Code of Practice for General-Purpose. 2025 , note =

    2025

  37. [37]

    2025 , note =

    Frontier Mitigations , institution =. 2025 , note =

    2025

  38. [38]

    and others , title =

    Gandhi, M. and others , title =. arXiv preprint arXiv:2509.22742 , year =

    work page arXiv

  39. [39]

    Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion Transfer , journal =

  40. [40]

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities , journal =

  41. [41]

    and others , title =

    Goemans, A. and others , title =. arXiv preprint arXiv:2411.08088 , year =

    work page arXiv

  42. [42]

    and others , title =

    Heim, L. and others , title =. 2024 , address =

    2024

  43. [43]

    , title =

    Ord, T. , title =. 2025 , note =

    2025

  44. [44]

    and others , title =

    Bluemke, E. and others , title =. 2024 , note =

    2024

  45. [45]

    and others , title =

    Hammond, L. and others , title =. arXiv preprint arXiv:2502.14143 , year =

    work page arXiv

  46. [46]

    Scaling Laws for Autoregressive Generative Modeling

    Henighan, T. and others , title =. arXiv preprint arXiv:2010.14701 , year =

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  47. [47]

    Scaling Laws for Transfer

    Hernandez, D. and others , title =. arXiv preprint arXiv:2102.01293 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  48. [48]

    Deep Learning Scaling is Predictable, Empirically

    Hestness, J. and others , title =. arXiv preprint arXiv:1712.00409 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  49. [49]

    and others , title =

    Hoffman, J. and others , title =. Advances in Neural Information Processing Systems 35 (NeurIPS 2022) , year =

    2022

  50. [50]

    2024 , address =

    Comments to Department of Commerce Regarding Significant Malicious Cyber-Enabled Activities , institution =. 2024 , address =

    2024

  51. [51]

    AI Alignment: A Comprehensive Survey

    Ji, J. and others , title =. arXiv preprint arXiv:2310.19852 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  52. [52]

    and others , title =

    Jones, E. and others , title =. arXiv preprint arXiv:2406.14595 , year =

    work page arXiv

  53. [53]

    and others , title =

    Jones, E. and others , title =. arXiv preprint arXiv:2502.16797 , year =

    work page arXiv

  54. [54]

    Scaling Laws for Neural Language Models

    Kaplan, J. and others , title =. arXiv preprint arXiv:2001.08361 , year =

    work page internal anchor Pith review Pith/arXiv arXiv 2001

  55. [55]

    and others , title =

    Kawaharazuka, K. and others , title =. arXiv preprint arXiv:2510.07077 , year =

    work page arXiv

  56. [56]

    , title =

    Kleinberg, J and Raghavan, M. , title =. Proceedings of the National Academy of Sciences , volume =

  57. [57]

    and others , title =

    Kumar, K. and others , title =. arXiv preprint arXiv:2502.21321 , year =

    work page arXiv

  58. [58]

    and others , title =

    Lai, H. and others , title =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025) , year =

    2025

  59. [59]

    and others , title =

    Li, J. and others , title =. Proceedings of the 40th International Conference on Machine Learning (ICML 2023) , year =

    2023

  60. [60]

    and Linkov, I , title =

    Trump, B. and Linkov, I , title =. Environment Systems and Decisions , volume =

  61. [61]

    2023 , note =

    Responsible Scaling Policies (. 2023 , note =

    2023

  62. [62]

    AI in 2030) , year =

    2030

  63. [63]

    Reuters , year =

    Kanishka Singh , title =. Reuters , year =

  64. [64]

    Guidelines for capability elicitation , year =

  65. [65]

    2025 , note =

    Assembly Bill. 2025 , note =

    2025

  66. [66]

    2024 , note =

    Project. 2024 , note =

    2024

  67. [67]

    2025 , address =

    Preparedness Framework v2 , institution =. 2025 , address =

    2025

  68. [68]

    2025 , address =

    Deep Research System Card , institution =. 2025 , address =

    2025

  69. [69]

    Strengthening safety with external testing , year =

  70. [70]

    2025 , note =

    Preparing for future. 2025 , note =

    2025

  71. [71]

    and others , title =

    Wu, Q. and others , title =. arXiv preprint arXiv:2510.14036 , year =

    work page arXiv

  72. [72]

    2026 , note =

    Fazl Berez , title =. 2026 , note =

    2026

  73. [73]

    Park, J. S. and others , title =. Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST '23) , year =

  74. [74]

    and others , title =

    Parsons, M. and others , title =. International Journal of Disaster Risk Reduction , volume =

  75. [75]

    Pearson, L. J. and Pearson, C. J. , title =. Proceedings of the National Academy of Sciences , volume =

  76. [76]

    arXiv preprint arXiv:2403.13793 , year=

    Phuong, M. and others , title =. arXiv preprint arXiv:2403.13793 , year =

    work page arXiv

  77. [77]

    and others , title =

    Nevo, S. and others , title =. 2024 , number =

    2024

  78. [78]

    and others , title =

    Robey, A. and others , title =. arXiv preprint arXiv:2410.13691 , year =

    work page arXiv

  79. [79]

    and others , title =

    Rosenqvist, P. and others , title =. Critical Infrastructure Protection , publisher =. 2018 , pages =

    2018

  80. [80]

    and others , title =

    Ruan, Y. and others , title =. arXiv preprint arXiv:2405.10938 , year =

    work page arXiv

Showing first 80 references.