pith. sign in

arxiv: 2606.15386 · v2 · pith:ENSDZL6Onew · submitted 2026-06-13 · 💻 cs.LG

A Compositional Framework for Open-ended Intelligence

Ida Momennejad , Roberta Raileanu This is my paper

Pith reviewed 2026-06-27 03:59 UTC · model grok-4.3

classification 💻 cs.LG
keywords open-ended intelligencecompositional closureprimitivescompositional grammarnext primitive predictionadaptive responseslifelong learning
0
0 comments X

The pith

Open-ended intelligence arises from the compositional closure of a finite set of primitives and composition operators.

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

The paper argues that open-ended intelligence, defined as adapting to novel problems and environments unlike those in training, requires two pillars: a minimal set of representational primitives like states and actions plus algorithmic ones like nearest neighbor, and an acquired compositional grammar using selection, recursion, and branching. These together induce a closure L(P,C) whose properties enable unbounded generation of sequences and motifs across task families. A sympathetic reader would care because the framework supplies a mathematical basis for systems that generate new solutions through recombination instead of scaling or memorization. It also introduces next primitive prediction as a training objective to acquire reusable primitives and grammar, supported by curriculum learning and self-play for lifelong expansion of the closure.

Core claim

The paper claims that the compositional closure induced by a finite primitive set P and a set of composition operators C yields properties that support unbounded compositional generation across families of tasks and worlds. The closure of the two pillars produces infinite adaptive responses across a wide range of settings. The mathematics grounds complementary agendas such as explanation metrics and native compositional architectures, with next primitive prediction proposed as an objective that encourages reusable primitives so new solutions arise via recombination.

What carries the argument

The compositional closure L(P,C) induced by a finite primitive set P and composition operators C, which generates sequences of operations and recurring motifs from selection, recursion, and branching.

If this is right

  • Evaluation metrics for explanation and interpretability can be built directly from the properties of the induced closure.
  • Architectures can be designed so that compositional generalization is a built-in consequence of the grammar rather than an after-effect of training.
  • Curriculum learning and self-play become mechanisms to expand the closure by discovering new reusable primitives and transition motifs.
  • Next primitive prediction acts as a training objective that shifts focus from end-to-end prediction to acquisition of compositional building blocks.

Where Pith is reading between the lines

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

  • The framework could connect to program synthesis by treating open-ended behavior as grammar induction over a shared primitive base.
  • It suggests that biological open-endedness might arise from evolutionary mechanisms that enlarge the primitive set itself across generations.
  • Empirical tests could measure closure size growth under self-play in environments with controlled structural novelty.

Load-bearing premise

A finite set of primitives and composition operators can induce a closure whose properties support unbounded generation of adaptive responses to substantially novel problems and environments.

What would settle it

A concrete family of tasks and worlds where increasing novelty in structure cannot be solved by any finite P and C, even after exhaustive recombination of the induced closure.

Figures

Figures reproduced from arXiv: 2606.15386 by Ida Momennejad, Roberta Raileanu.

Figure 1
Figure 1. Figure 1: Compositional Open-Ended Intelligence. A minimal basis of primitives and operators feeding into a compositional grammar, which leads to an unbounded closure of increasingly complex solutions across infinite worlds. comparison (Lippl et al., 2026). Composition refers to how the system pieces the primitives together. It can be captured in terms of transition motifs among primitives, including sequential tran… view at source ↗
read the original abstract

Open-ended intelligence is the capacity to adapt to novel problems and environments that are substantially different from those in training. A mathematics of open-ended intelligence requires two pillars: first, a minimal set of representational primitives (e.g., states, actions) and algorithmic primitives (e.g., nearest neighbor); and second, an acquired compositional grammar for selection, recursion, and branching that produces sequences of operations and recurring motifs. We formalize open-ended intelligence in terms of the compositional closure induced by a finite primitive set $P$ and a set of composition operators $C$. We characterize properties of the induced closure $\mathcal{L}(P,C)$ that support unbounded compositional generation across families of tasks and worlds. The closure of the two pillars yields infinite adaptive responses across a wide range of settings. The mathematics supports complementary research agendas, including evaluation metrics for explanation and interpretability, and novel architectures where compositional generalization is native. We propose next primitive prediction (NPP) as a novel architectural objective, where training encourages the acquisition of reusable algorithmic primitives and their compositional grammar, such that new solutions are generated through recombination. Given such an objective, curriculum learning and self-play can enable lifelong learning, expanding the closure by discovering reusable primitives and transition motifs across settings. We ground the framework through case studies in physics, evolution, and neuroscience.

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 / 0 minor

Summary. The paper claims that open-ended intelligence—adaptation to novel problems and environments substantially different from training—can be formalized via two pillars: a finite set of representational/algorithmic primitives P and a compositional grammar (selection, recursion, branching) via operators C. This induces a closure L(P,C) whose properties support unbounded compositional generation across task and world families. The closure yields infinite adaptive responses; the paper introduces Next Primitive Prediction (NPP) as a training objective to acquire reusable primitives and grammar, enabling lifelong learning via curriculum and self-play, and grounds the ideas in case studies from physics, evolution, and neuroscience.

Significance. If the characterization of L(P,C) is made rigorous with explicit transfer conditions, the framework could supply a mathematical basis for compositional open-ended systems and motivate architectures where generalization is native rather than emergent. The NPP objective is a concrete, falsifiable proposal that could be tested against existing compositional generalization benchmarks. The interdisciplinary case studies are a positive feature that may help connect AI theory to other domains.

major comments (2)
  1. [Abstract] Abstract: the claim that L(P,C) 'yields infinite adaptive responses across a wide range of settings' is load-bearing for the central thesis yet rests on an unelaborated characterization; no theorem, lemma, or derivation is referenced that shows elements of the closure remain effective when the underlying world or task distribution lies outside the span of the initial primitives P.
  2. [Abstract] Abstract: adaptation is defined relative to 'substantially different' environments, but the formalization appears to equate syntactic recombination under C with semantic effectiveness; no transfer conditions, semantics for the primitives, or explicit criteria for 'substantial difference' are supplied, leaving the adaptation claim without a correctness argument.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive critique, which highlights important gaps in the formal grounding of our claims. We agree that the abstract's assertions about adaptive responses require more careful qualification and supporting discussion. Below we respond point-by-point and indicate the revisions we will make.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that L(P,C) 'yields infinite adaptive responses across a wide range of settings' is load-bearing for the central thesis yet rests on an unelaborated characterization; no theorem, lemma, or derivation is referenced that shows elements of the closure remain effective when the underlying world or task distribution lies outside the span of the initial primitives P.

    Authors: The manuscript (Section 3) defines L(P,C) via inductive closure under the operators in C and states that this structure permits unbounded generation of composite expressions. However, the text does not supply a theorem establishing that any such expression remains effective outside the support of the base primitives P. The claim in the abstract therefore overreaches. We will revise the abstract to read 'supports the potential for unbounded compositional generation' and add a short subsection (new Section 3.4) that explicitly lists the assumptions required for effectiveness (e.g., that base primitives are semantically sound in their original domains and that the operators preserve relevant invariants). No correctness proof will be claimed; the subsection will instead delineate the conditions under which transfer could hold. revision: yes

  2. Referee: [Abstract] Abstract: adaptation is defined relative to 'substantially different' environments, but the formalization appears to equate syntactic recombination under C with semantic effectiveness; no transfer conditions, semantics for the primitives, or explicit criteria for 'substantial difference' are supplied, leaving the adaptation claim without a correctness argument.

    Authors: The current formalization treats adaptation as the capacity to produce novel expressions via C; it does not derive semantic effectiveness from syntax alone. The manuscript leaves the semantics of P and the metric for 'substantial difference' to be supplied by the target domain (as illustrated in the case studies). This is a genuine omission in the abstract and introductory framing. In revision we will (i) add an explicit definition of 'substantial difference' in terms of divergence between task/world distributions, (ii) state that transfer conditions are domain-dependent and not provided by the framework itself, and (iii) qualify the abstract sentence to 'the closure supplies a generative mechanism whose effectiveness depends on domain-specific semantics of P'. These changes will be reflected in both the abstract and Section 2. revision: yes

Circularity Check

0 steps flagged

No circularity: definitional framework with no reduction to inputs by construction

full rationale

The paper presents a conceptual formalization of open-ended intelligence as the compositional closure L(P,C) induced by finite primitives P and operators C, then states that this closure yields unbounded adaptive responses. No equations, fitted parameters, or derivations are shown that equate any claimed output (such as adaptation or infinite responses) back to the inputs by construction. No self-citations appear in the provided text, and the characterization of L(P,C) properties is offered as part of the framework definition itself rather than a prediction derived from prior results. The derivation chain is therefore self-contained as a proposal of new primitives and operators without the enumerated circular patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Review based on abstract only; ledger reflects concepts stated without supporting derivations or evidence.

axioms (1)
  • domain assumption A finite primitive set P and composition operators C induce a closure L(P,C) with properties supporting unbounded compositional generation across families of tasks and worlds.
    This is the central formalization presented in the abstract.
invented entities (1)
  • Next Primitive Prediction (NPP) no independent evidence
    purpose: Architectural objective that encourages acquisition of reusable algorithmic primitives and their compositional grammar.
    Introduced as a novel training target but without evidence or formal definition in the abstract.

pith-pipeline@v0.9.1-grok · 5758 in / 1222 out tokens · 50411 ms · 2026-06-27T03:59:02.869736+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

192 extracted references · 14 canonical work pages · 8 internal anchors

  1. [1]

    Position: We Need An Algorithmic Understanding of Generative

    Eberle, Oliver and McGee, Thomas and Giaffar, Hamza and Webb, Taylor and Momennejad, Ida , booktitle=. Position: We Need An Algorithmic Understanding of Generative. 2025 , note=

    2025

  2. [2]

    International Conference on Machine Learning (ICML) , year=

    Algorithmic Primitives and Compositional Geometry of Reasoning in Language Models , author=. International Conference on Machine Learning (ICML) , year=

  3. [3]

    A brain-inspired agentic architecture to improve planning with

    Webb, Taylor and Mondal, Shanka Subra and Momennejad, Ida , journal=. A brain-inspired agentic architecture to improve planning with

  4. [4]

    Evaluating cognitive maps and planning in Large Language Models with

    Momennejad, Ida and Hasanbeig, Hosein and Frujeri, Felipe Vieira and Sharma, Hiteshi and Ness, Robert and Jojic, Nebojsa and Palangi, Hamid and Larson, Jonathan , booktitle=. Evaluating cognitive maps and planning in Large Language Models with

  5. [5]

    arXiv preprint arXiv:2310.00194 , year=

    A prefrontal cortex-inspired architecture for planning in Large Language Models , author=. arXiv preprint arXiv:2310.00194 , year=

    work page arXiv

  6. [6]

    A Rubric for Human-like Agents and

    Momennejad, Ida , journal=. A Rubric for Human-like Agents and

  7. [7]

    Space, Time, and Memory, Oxford University Press , year=

    Memory and Planning in Brains and Machines: Multiscale Predictive Representations , author=. Space, Time, and Memory, Oxford University Press , year=

  8. [8]

    Proceedings of the 2024 Artificial Life Conference , year=

    Collective Innovation in Groups of Large Language Models , author=. Proceedings of the 2024 Artificial Life Conference , year=

    2024

  9. [9]

    Philosophical Transactions of the Royal Society B , volume=

    Collective Minds: Social Network Topology Shapes Collective Cognition , author=. Philosophical Transactions of the Royal Society B , volume=

  10. [10]

    Current Opinion in Behavioral Sciences , year=

    Learning Structures: Predictive Representations, Replay, and Generalization , author=. Current Opinion in Behavioral Sciences , year=

  11. [11]

    PLoS Computational Biology , year=

    Predictive representations can link model-based reinforcement learning to model-free mechanisms , author=. PLoS Computational Biology , year=

  12. [12]

    Nature Human Behaviour , volume=

    The successor representation in human reinforcement learning , author=. Nature Human Behaviour , volume=

  13. [13]

    eLife , volume=

    Offline replay supports planning in human reinforcement learning , author=. eLife , volume=

  14. [14]

    Nature Communications , year=

    Bridge ties bind collective memories , author=. Nature Communications , year=

  15. [15]

    Proceedings of the National Academy of Sciences , year=

    Mnemonic convergence in social networks: The emergent properties of cognition at a collective level , author=. Proceedings of the National Academy of Sciences , year=

  16. [16]

    Social Network Structure Shapes Innovation: Experience-sharing in

    Nisioti, Eleni and Mahaut, Mateo and Oudeyer, Pierre-Yves and Momennejad, Ida and Moulin-Frier, Cl. Social Network Structure Shapes Innovation: Experience-sharing in. arXiv preprint arXiv:2206.05060 , year=

    work page arXiv

  17. [17]

    2017 , howpublished=

    Open-endedness: The last grand challenge you've never heard of , author=. 2017 , howpublished=

    2017

  18. [18]

    Evolutionary Computation , volume=

    Abandoning objectives: Evolution through the search for novelty alone , author=. Evolutionary Computation , volume=. 2011 , publisher=

    2011

  19. [19]

    Clune, Jeff , journal=

  20. [20]

    2019 , note=

    Wang, Rui and Lehman, Joel and Clune, Jeff and Stanley, Kenneth O , booktitle=. 2019 , note=

    2019

  21. [21]

    Enhanced

    Wang, Rui and Lehman, Joel and Rawal, Aditya and Zhi, Jiale and Li, Yulun and Clune, Jeff and Stanley, Kenneth O , booktitle=. Enhanced. 2020 , organization=

    2020

  22. [22]

    OpenReview , year=

    A path towards autonomous machine intelligence , author=. OpenReview , year=

  23. [23]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Self-supervised learning from images with a joint-embedding predictive architecture , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  24. [24]

    Advances in Neural Information Processing Systems , volume=

    Attention is all you need , author=. Advances in Neural Information Processing Systems , volume=

  25. [25]

    Artificial Intelligence , volume=

    Reward is enough , author=. Artificial Intelligence , volume=. 2021 , publisher=

    2021

  26. [26]

    2018 , publisher=

    Reinforcement Learning: An Introduction , author=. 2018 , publisher=

    2018

  27. [27]

    Illuminating search spaces by mapping elites

    Illuminating search spaces by mapping elites , author=. arXiv preprint arXiv:1504.04909 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  28. [28]

    Frontiers in Robotics and AI , volume=

    Quality diversity: A new frontier for evolutionary computation , author=. Frontiers in Robotics and AI , volume=

  29. [29]

    International Conference on Machine Learning , pages=

    Curiosity-driven exploration by self-supervised prediction , author=. International Conference on Machine Learning , pages=. 2017 , organization=

    2017

  30. [30]

    International Conference on Learning Representations , year=

    Exploration by random network distillation , author=. International Conference on Learning Representations , year=

  31. [31]

    Sutton, Richard S and Precup, Doina and Singh, Satinder , journal=. Between. 1999 , publisher=

    1999

  32. [32]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    The option-critic architecture , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  33. [33]

    Machado, Marlos C and Bellemare, Marc G and Bowling, Michael , booktitle=. A. 2017 , organization=

    2017

  34. [34]

    World Models

    World models , author=. arXiv preprint arXiv:1803.10122 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  35. [35]

    International Conference on Learning Representations , year=

    Dream to control: Learning behaviors by latent imagination , author=. International Conference on Learning Representations , year=

  36. [36]

    Behavioral and Brain Sciences , volume=

    Building machines that learn and think like people , author=. Behavioral and Brain Sciences , volume=. 2017 , publisher=

    2017

  37. [37]

    Cognition , volume=

    Connectionism and cognitive architecture: A critical analysis , author=. Cognition , volume=. 1988 , publisher=

    1988

  38. [38]

    On the Measure of Intelligence

    On the measure of intelligence , author=. arXiv preprint arXiv:1911.01547 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 1911

  39. [39]

    Psychological Review , volume=

    The logical primitives of thought: Empirical foundations for compositional cognitive models , author=. Psychological Review , volume=. 2016 , publisher=

    2016

  40. [40]

    2003 , publisher=

    The Algebraic Mind: Integrating Connectionism and Cognitive Science , author=. 2003 , publisher=

    2003

  41. [41]

    International Conference on Learning Representations , year=

    Function vectors in large language models , author=. International Conference on Learning Representations , year=

  42. [42]

    Advances in Neural Information Processing Systems , year=

    Emergent Complexity and Zero-shot Transfer via Unsupervised Environment Design , author=. Advances in Neural Information Processing Systems , year=

  43. [43]

    Advances in Neural Information Processing Systems , year=

    Replay-Guided Adversarial Environment Design , author=. Advances in Neural Information Processing Systems , year=

  44. [44]

    International Conference on Machine Learning , year=

    Evolving Curricula with Regret-Based Environment Design , author=. International Conference on Machine Learning , year=

  45. [45]

    International Conference on Learning Representations , year=

    Samvelyan, Mikayel and Khan, Akbir and Dennis, Michael and Jiang, Minqi and Parker-Holder, Jack and Foerster, Jakob and Raileanu, Roberta and Rockt. International Conference on Learning Representations , year=

  46. [46]

    Learning with

    Campero, Andres and Raileanu, Roberta and K. Learning with. International Conference on Learning Representations , year=

  47. [47]

    International Conference on Machine Learning , pages=

    Decoupling Value and Policy for Generalization in Reinforcement Learning , author=. International Conference on Machine Learning , pages=. 2021 , organization=

    2021

  48. [48]

    Advances in Neural Information Processing Systems , year=

    Automatic Data Augmentation for Generalization in Deep Reinforcement Learning , author=. Advances in Neural Information Processing Systems , year=

  49. [49]

    International Conference on Machine Learning , year=

    Fast Adaptation to New Environments via Policy-Dynamics Value Functions , author=. International Conference on Machine Learning , year=

  50. [50]

    International Conference on Machine Learning , year=

    Modeling Others using Oneself in Multi-Agent Reinforcement Learning , author=. International Conference on Machine Learning , year=

  51. [51]

    International Conference on Learning Representations , year=

    Raileanu, Roberta and Rockt. International Conference on Learning Representations , year=

  52. [52]

    Proceedings of the National Academy of Sciences , volume=

    Inferring the structure and dynamics of interactions in schooling fish , author=. Proceedings of the National Academy of Sciences , volume=. 2011 , doi=

    2011

  53. [53]

    Nature , volume=

    Effective leadership and decision-making in animal groups on the move , author=. Nature , volume=. 2005 , doi=

    2005

  54. [54]

    Trends in Cognitive Sciences , volume=

    Collective cognition in animal groups , author=. Trends in Cognitive Sciences , volume=. 2009 , doi=

    2009

  55. [55]

    Journal of Theoretical Biology , volume=

    Collective memory and spatial sorting in animal groups , author=. Journal of Theoretical Biology , volume=

  56. [56]

    Advances in the Study of Behavior , volume=

    Self-organization and collective behavior in vertebrates , author=. Advances in the Study of Behavior , volume=

  57. [57]

    Science , volume=

    Uninformed individuals promote democratic consensus in animal groups , author=. Science , volume=. 2011 , doi=

    2011

  58. [58]

    Science , volume=

    Shared decision-making drives collective movement in wild baboons , author=. Science , volume=

  59. [59]

    Proceedings of the National Academy of Sciences , volume=

    Revealing the hidden networks of interaction in mobile animal groups allows prediction of complex behavioral contagion , author=. Proceedings of the National Academy of Sciences , volume=

  60. [60]

    2025 , eprint=

    Next-Latent Prediction Transformers Learn Compact World Models , author=. 2025 , eprint=

    2025

  61. [61]

    Proceedings of the 41st International Conference on Machine Learning , series=

    Open-Endedness is Essential for Artificial Superhuman Intelligence , author=. Proceedings of the 41st International Conference on Machine Learning , series=. 2024 , publisher=

    2024

  62. [62]

    and Gopnik, Alison , title =

    Repacholi, Betty M. and Gopnik, Alison , title =. Developmental Psychology , year =

  63. [63]

    and Schulz, Laura E

    Gopnik, Alison and Glymour, Clark and Sobel, David M. and Schulz, Laura E. and Kushnir, Tamar and Danks, David , title =. Psychological Review , year =

  64. [64]

    and Kinzler, Katherine D

    Spelke, Elizabeth S. and Kinzler, Katherine D. , title =. Developmental Science , year =

  65. [65]

    and Aslin, Richard N

    Kidd, Celeste and Piantadosi, Steven T. and Aslin, Richard N. , title =. PLoS ONE , year =

  66. [66]

    and Tenenbaum, Joshua B

    Piantadosi, Steven T. and Tenenbaum, Joshua B. and Goodman, Noah D. , title =. Psychological Review , year =

  67. [67]

    and Ghazanfar, Asif A

    Krakauer, John W. and Ghazanfar, Asif A. and Gomez-Marin, Alex and MacIver, Malcolm A. and Poeppel, David , title =. Neuron , year =

  68. [68]

    2021 , eprint =

    Mitchell, Melanie , title =. 2021 , eprint =

    2021

  69. [69]

    International Conference on Learning Representations , year =

    Burda, Yuri and Edwards, Harrison and Storkey, Amos and Klimov, Oleg , title =. International Conference on Learning Representations , year =

  70. [70]

    , title =

    Burda, Yuri and Edwards, Harrison and Pathak, Deepak and Storkey, Amos and Darrell, Trevor and Efros, Alexei A. , title =. International Conference on Learning Representations , year =

  71. [71]

    and Sharma, Pratyusha and Islam, Riashat and Lamb, Alex and Langford, John , title =

    Teoh, Jayden and Tomar, Manan and Ahn, Kwangjun and Hu, Edward S. and Sharma, Pratyusha and Islam, Riashat and Lamb, Alex and Langford, John , title =. 2025 , eprint =

    2025

  72. [72]

    , title =

    Oudeyer, Pierre-Yves and Smith, Linda B. , title =. Topics in Cognitive Science , year =

  73. [73]

    2022 , eprint=

    Evaluating Understanding on Conceptual Abstraction Benchmarks , author=. 2022 , eprint=

    2022

  74. [74]

    Cell , volume=

    The geometry of abstraction in the hippocampus and prefrontal cortex , author=. Cell , volume=. 2020 , publisher=

    2020

  75. [75]

    Nature , volume=

    The importance of mixed selectivity in complex cognitive tasks , author=. Nature , volume=. 2013 , publisher=

    2013

  76. [76]

    Current Opinion in Neurobiology , volume=

    Why neurons mix: high dimensionality for higher cognition , author=. Current Opinion in Neurobiology , volume=. 2016 , publisher=

    2016

  77. [77]

    Neuron , volume=

    Tuned geometries of hippocampal representations meet the computational demands of social memory , author=. Neuron , volume=. 2024 , publisher=

    2024

  78. [78]

    Nature , volume=

    Geometry of abstract learned knowledge in the hippocampus , author=. Nature , volume=. 2021 , publisher=

    2021

  79. [79]

    Nature Communications , volume=

    Abstract representations emerge naturally in neural networks trained to perform multiple tasks , author=. Nature Communications , volume=. 2023 , publisher=

    2023

  80. [80]

    Advances in Neural Information Processing Systems , volume=

    Compositional generalization through abstract representations in human and artificial neural networks , author=. Advances in Neural Information Processing Systems , volume=

Showing first 80 references.