pith. sign in

arxiv: 2605.17903 · v1 · pith:PGVVBZEBnew · submitted 2026-05-18 · 💻 cs.AI · cs.CL· cs.HC· cs.IR

Agentic Chunking and Bayesian De-chunking of AI Generated Fuzzy Cognitive Maps: A Model of the Thucydides Trap

Akash Kumar Panda , Olaoluwa Adigun , Bart Kosko This is my paper

Pith reviewed 2026-05-20 10:16 UTC · model grok-4.3

classification 💻 cs.AI cs.CLcs.HCcs.IR
keywords fuzzy cognitive mapsThucydides TrapLLM agentsagentic chunkingBayesian de-chunkingcausal knowledge graphsconflict predictiondynamical systems
0
0 comments X

The pith

LLM agents chunk text into overlapping parts whose fuzzy cognitive maps mix to predict war when a rising power's ambition activates.

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

The paper shows how to use large-language-model agents to split an essay into overlapping text chunks, build a fuzzy cognitive map from each chunk, and convexly mix the resulting sparse causal matrices into one representative cyclic knowledge graph. Applied to Graham Allison's description of the Thucydides Trap between a dominant power and a rising power, the mixed maps are then run as dynamical systems. When the node for the rising power's ambition and entitlement is held on, seven of the eight resulting graphs equilibrate to attractors that indicate some form of war. A sympathetic reader would care because the method supplies a scalable, largely automatic route from unstructured political text to quantitative causal simulation without requiring hand-crafted models.

Core claim

Convex mixing of chunk FCMs generated by Gemini 3.1 agents from overlapping segments of Allison's Thucydides Trap essay produces representative causal knowledge graphs. These graphs, when stimulated by keeping the rising-power ambition node on, equilibrate to fixed-point or limit-cycle attractors that encode a type of war in seven out of eight cases. The same mixing structure supports an operator-level Bayesian update that yields de-chunked, posterior-like FCMs available for further iterations.

What carries the argument

Convex mixing of sparse causal chunk matrices from overlapping text chunks, together with the operator-level Bayesian inference that produces de-chunked posterior FCMs.

If this is right

  • The mixing operation remains computationally light even for long texts because it operates on sparse matrices.
  • De-chunked FCMs can be fed back into the same pipeline for successive Bayesian updates.
  • Dynamical simulation of the final FCM directly yields concrete outcome predictions such as war or peace.
  • The method applies to any essay-length text that contains identifiable causal claims.

Where Pith is reading between the lines

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

  • The same chunk-and-mix procedure could be run on other historical or policy documents to test whether consistent conflict predictions emerge across different source texts.
  • If future LLMs reduce causal-extraction error, the reliability of the resulting war predictions would increase without changing the mixing mathematics.
  • The de-chunking step offers a route to combine multiple independent text sources into a single updated FCM while preserving interpretability.
  • One could compare the attractor reached under different chunk-overlap levels to measure how robust the war prediction is to the choice of chunk size.

Load-bearing premise

The LLM agents correctly extract and encode the causal relations present in each text chunk without introducing hallucinations or systematic distortions that would change the mixed map's dynamical behavior.

What would settle it

Stimulate the mixed FCM by holding the rising-power ambition node on and observe that the system equilibrates to a stable non-war attractor rather than to any war-type attractor.

Figures

Figures reproduced from arXiv: 2605.17903 by Akash Kumar Panda, Bart Kosko, Olaoluwa Adigun.

Figure 1
Figure 1. Figure 1: AI-automated chunking and de-chunking of sampled text to create chunk-based Fuzzy Cognitive Maps (FCMs) and the Bayesian-like posterior FCMs. The chunking experiment uses Graham Allison’s article in The Atlantic titled “The Thucydides Trap: Are the U.S. and China headed for war?” as input. Allison explains: “When a rising power is threatening to displace a ruling power, standard crises that would otherwise… view at source ↗
Figure 2
Figure 2. Figure 2: Likelihood FCM Mixture. The 3 likelihood FCMs F1, F2, and F3 extracted from chunks of the “Thucydides Trap” article are on the top. The 11-node mixture FCM FLikelihood mixes the 3 likelihood FCMs with equal mixing weights 1 3 . 2 Fuzzy Cognitive Maps FCMs model causal dynamical systems as directed weighted graphs [3,5–7,9–15]. The concept nodes describe the causal variables in the dynamical system and the … view at source ↗
Figure 3
Figure 3. Figure 3: FCM De-chunking and Bayesian-like posterior FCMs. Equation (3) converts the mixed likelihood FCM edge matrix E l into 3 sparse posterior-like edge matrices E p 1 , E p 2 , and E p 3 . These posterior matrices then define the posterior FCMs P1, P2, and P3. causal relationship between Ci and Cj . Low magnitude of eij describes a weak causal relationship between Ci and Cj . An n × n matrix E describes all the… view at source ↗
Figure 4
Figure 4. Figure 4: Binary recursive text splitting: The binary recursive text method breaks down an input text X to small, contiguous, and non-overlapping text chunks. This method uses a binary tree structure and the text chunks are the leaf nodes at depth dmax. The maximum depth dmax = 3 so the number of text chunks is 2 3 = 8. 3.1 Text Chunking Text chunking is the process of breaking a text down into small, non-overlappin… view at source ↗
Figure 5
Figure 5. Figure 5: Overlapping text chunking: This illustrates the overlapping chunking method (Algorithm 2) with dmax = 3. This breaks down an input text X to a sequence of overlapping chunks. The first step applies the binary recursive text split to generate the non-overlapping chunks: C(3, 1), ..., C(3, 8). The second step extracts proportional overlapping chunks Ok(α) for the non-overlapping chunks. Each text chunk break… view at source ↗
Figure 6
Figure 6. Figure 6: Equilibrium answers to the clamping question: What if the rising power’s ambi￾tions remain unchecked? The time steps are along the x-axis and the nodes are along the y-axis. The purple nodes are inactive and the yellow nodes are active. The clamped￾on nodes are in red. (a), (c), and (e) show the equilibria from the 3 likelihood FCMs F1, F2, and F3 and (g) shows the equilibrium from their equal weight mixtu… view at source ↗
read the original abstract

We automatically generate feedback causal fuzzy cognitive maps (FCMs) from text by teaching large-language-model agents to break the text into overlapping chunks of text. Convex mixing of these chunk FCMs gives a representative cyclic FCM knowledge graph. The text chunks can have different levels of overlap. The chunk FCMs still mix to form a new FCM causal knowledge graph. The mixing technique scales because it uses light computation with sparse causal chunk matrices. The mixing structure allows an operator-level type of Bayesian inference that produces "de-chunked" or posterior-like FCMs from the mixed FCM. These de-chunked FCMs are useful in their own right and allow further iterations of Bayesian updating. We demonstrate these mixing techniques on the essay text of Allison's "Thucydides Trap" model of conflict between a dominant power such as the United States and a rising power such as China. The FCM dynamical systems predict outcomes as they equilibrate to fixed-point or limit-cycle attractors. Seven out of 8 FCM knowledge graphs predicted a type of war when we stimulated them by turning on and keeping on the concept node that stands for the rising power's ambition and entitlement. Gemini 3.1 LLMs served as the chunking AI agents.

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

3 major / 2 minor

Summary. The manuscript introduces an agentic chunking approach in which LLM agents (Gemini 3.1) divide source text into overlapping segments, extract causal relations to produce sparse Fuzzy Cognitive Maps (FCMs), and then apply convex mixing followed by Bayesian de-chunking to obtain representative cyclic FCMs. The method is demonstrated on Graham Allison’s essay on the Thucydides Trap; dynamical simulations of the resulting maps show that seven of eight FCMs reach war-like attractors when the rising-power ambition/entitlement node is clamped on.

Significance. If the LLM-extracted causal edges can be shown to faithfully encode the source text, the combination of sparse-matrix convex mixing and operator-level Bayesian de-chunking supplies a computationally light, iteratively refinable pipeline for turning narrative sources into simulatable causal models. The approach is novel in its explicit handling of chunk overlap and posterior-like de-chunking, and the demonstration on a well-known geopolitical text illustrates a potential use case in automated international-relations modeling.

major comments (3)
  1. [Abstract and Results] Abstract and Results section: The headline finding that 7/8 FCMs predict war after clamping the rising-power ambition node is presented as a demonstration, yet no precision, recall, or inter-annotator agreement figures are supplied that compare the LLM-extracted causal links and weights against human-coded ground truth for any chunk. Without such validation, it is impossible to determine whether the war predictions reflect relations stated in Allison’s text or artifacts introduced by the LLM agents.
  2. [Methods] Methods (chunking and mixing): Chunk overlap levels are explicitly listed as free parameters, yet the manuscript reports no sensitivity analysis or error bars showing how changes in overlap or mixing weights alter the final attractor outcomes. Because the 7/8 result is load-bearing for the central claim, the robustness of the dynamical predictions to these choices must be quantified.
  3. [Results] Dynamical simulation: The FCMs are described as equilibrating to fixed-point or limit-cycle attractors, but the text supplies neither simulation parameters (e.g., update rule, convergence tolerance) nor any measure of variability across runs or initial conditions. This omission prevents assessment of whether the reported war predictions are stable properties of the extracted graphs.
minor comments (2)
  1. [Methods] The distinction between “de-chunked” and “posterior-like” FCMs would benefit from an explicit equation or pseudocode definition in the methods section.
  2. [Figures] Figure captions and axis labels for the attractor plots should indicate the exact node clamped and the simulation length used.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive comments, which have helped us identify areas for improvement in the manuscript. We address each major comment below and outline the revisions we plan to make.

read point-by-point responses

  1. Referee: Abstract and Results section: The headline finding that 7/8 FCMs predict war after clamping the rising-power ambition node is presented as a demonstration, yet no precision, recall, or inter-annotator agreement figures are supplied that compare the LLM-extracted causal links and weights against human-coded ground truth for any chunk. Without such validation, it is impossible to determine whether the war predictions reflect relations stated in Allison’s text or artifacts introduced by the LLM agents.

    Authors: We agree that validation against human ground truth is important to ensure the extracted causal relations faithfully represent the source text. The manuscript presents this as an initial demonstration of the agentic chunking and de-chunking pipeline. To address this, we will conduct a limited human annotation study on selected chunks in the revised version, computing precision, recall, and inter-annotator agreement where applicable. This will provide evidence on the accuracy of the LLM extractions. revision: yes

  2. Referee: Methods (chunking and mixing): Chunk overlap levels are explicitly listed as free parameters, yet the manuscript reports no sensitivity analysis or error bars showing how changes in overlap or mixing weights alter the final attractor outcomes. Because the 7/8 result is load-bearing for the central claim, the robustness of the dynamical predictions to these choices must be quantified.

    Authors: We acknowledge the need for robustness checks given that overlap and mixing weights are parameters. In the revision, we will include a sensitivity analysis by varying these parameters within reasonable ranges and reporting the resulting variations in the attractor outcomes, including any error bars or confidence intervals for the proportion of maps predicting war. revision: yes

  3. Referee: Dynamical simulation: The FCMs are described as equilibrating to fixed-point or limit-cycle attractors, but the text supplies neither simulation parameters (e.g., update rule, convergence tolerance) nor any measure of variability across runs or initial conditions. This omission prevents assessment of whether the reported war predictions are stable properties of the extracted graphs.

    Authors: We appreciate this observation regarding reproducibility. We will expand the Results section to include the specific simulation parameters, such as the update rule (e.g., the iterative inference method used), convergence tolerance, and number of iterations. Furthermore, we will report results from multiple simulations with different initial conditions to demonstrate the stability of the war-like attractors. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation starts from external text and produces independent dynamical predictions

full rationale

The paper extracts causal structure from Allison's external essay via LLM chunking agents, forms FCMs through convex mixing of sparse matrices, applies Bayesian de-chunking, and then simulates the resulting dynamical system by clamping the rising-power node. The reported 7/8 war predictions are attractor outcomes of these constructed graphs rather than quantities defined by fitted parameters, self-referential equations, or load-bearing self-citations. No step matches the enumerated circularity patterns; the chain remains open to external checks against the source text and human-coded relations.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The paper relies on the domain assumption that convex mixing of chunk FCMs produces a representative graph and that LLM chunking faithfully captures causal structure; no new physical entities are introduced and free parameters appear limited to choices of overlap level.

free parameters (1)
  • chunk overlap levels
    Different levels of overlap in text chunks are chosen to generate varied FCMs that still mix into one graph.
axioms (1)
  • domain assumption Convex mixing of chunk FCMs gives a representative cyclic FCM knowledge graph.
    This is invoked as the basis for combining sparse causal chunk matrices into a single FCM.

pith-pipeline@v0.9.0 · 5773 in / 1449 out tokens · 68004 ms · 2026-05-20T10:16:06.897495+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

15 extracted references · 15 canonical work pages · 1 internal anchor

  1. [1]

    and china headed for war? The Atlantic (Sep 2015), published September 24, 2015

    Allison, G.: The thucydides trap: Are the u.s. and china headed for war? The Atlantic (Sep 2015), published September 24, 2015

    work page 2015

  2. [2]

    Artificial intelligence 42(2-3), 393–405 (1990)

    Cooper, G.F.: The computational complexity of probabilistic inference using bayesian belief networks. Artificial intelligence 42(2-3), 393–405 (1990)

    work page 1990

  3. [3]

    Glykas, M.: Fuzzy cognitive maps: Advances in theory, methodologies, tools and applications, vol. 247. Springer (2010)

    work page 2010

  4. [4]

    https://deepmind.google/models/gemini/ pro/ (March 2026), preview model

    Google DeepMind: Gemini 3.1 pro. https://deepmind.google/models/gemini/ pro/ (March 2026), preview model. Google DeepMind

    work page 2026

  5. [5]

    In: AIP Conference proceedings

    Kosko, B.: Differential Hebbian learning. In: AIP Conference proceedings. vol. 151, pp. 277–282. American Institute of Physics (1986)

    work page 1986

  6. [6]

    International journal of man-machine studies 24(1), 65–75 (1986)

    Kosko, B.: Fuzzy cognitive maps. International journal of man-machine studies 24(1), 65–75 (1986)

    work page 1986

  7. [7]

    Inter- national Journal of Approximate Reasoning 2(4), 377–393 (1988)

    Kosko, B.: Hidden patterns in combined and adaptive knowledge networks. Inter- national Journal of Approximate Reasoning 2(4), 377–393 (1988)

    work page 1988

  8. [8]

    Beyond DAGs: Modeling Causal Feedback with Fuzzy Cognitive Maps

    Osoba, O., Kosko, B.: Beyond dags: modeling causal feedback with fuzzy cognitive maps. arXiv preprint arXiv:1906.11247 (2019)

    work page internal anchor Pith review Pith/arXiv arXiv 1906

  9. [9]

    The Journal of Defense Modeling and Simulation 14(1), 17–32 (2017)

    Osoba, O.A., Kosko, B.: Fuzzy cognitive maps of public support for insurgency and terrorism. The Journal of Defense Modeling and Simulation 14(1), 17–32 (2017)

    work page 2017

  10. [10]

    arXiv preprint arXiv:2601.00097 (2025)

    Panda, A.K., Adigun, O., Kosko, B.: The agentic leash: Extracting causal feedback fuzzy cognitive maps with llms. arXiv preprint arXiv:2601.00097 (2025)

    work page arXiv 2025

  11. [11]

    In: 24th IEEE International Con- ference on Machine Learning and Applications (ICMLA)

    Panda, A.K., Adigun, O., Kosko, B.: Causal autoencoder-like generation of feed- back fuzzy cognitive maps with an llm agent. In: 24th IEEE International Con- ference on Machine Learning and Applications (ICMLA). pp. 1234–1241. IEEE (2025)

    work page 2025

  12. [12]

    Papageorgiou, E.I.: Fuzzy cognitive maps for applied sciences and engineering: from fundamentals to extensions and learning algorithms, vol. 54. Springer Science & Business Media (2013)

    work page 2013

  13. [13]

    Fuzzy Sets and Systems 161(19), 2515–2532 (2010)

    Stach, W., Kurgan, L., Pedrycz, W.: A divide and conquer method for learning large fuzzy cognitive maps. Fuzzy Sets and Systems 161(19), 2515–2532 (2010)

    work page 2010

  14. [14]

    International Journal of Intelligent Systems 22(2), 181–202 (2007)

    Taber, R., Yager, R.R., Helgason, C.M.: Quantization effects on the equilibrium behavior of combined fuzzy cognitive maps. International Journal of Intelligent Systems 22(2), 181–202 (2007)

    work page 2007

  15. [15]

    Applied Energy 210, 487–498 (2018)

    Ziv, G., Watson, E., Young, D., Howard, D.C., Larcom, S.T., Tanentzap, A.J.: The potential impact of brexit on the energy, water and food nexus in the uk: A fuzzy cognitive mapping approach. Applied Energy 210, 487–498 (2018)

    work page 2018