ReflectiChain: Epistemic Grounding in LLM-Driven World Models for Supply Chain Resilience

Jia Luo

arxiv: 2606.10359 · v1 · pith:W4E2WD5Fnew · submitted 2026-06-09 · 💻 cs.AI

ReflectiChain: Epistemic Grounding in LLM-Driven World Models for Supply Chain Resilience

Jia Luo This is my paper

Pith reviewed 2026-06-27 13:35 UTC · model grok-4.3

classification 💻 cs.AI

keywords supply chain resilienceLLM world modelsepistemic uncertaintyaleatoric uncertaintygenerative modelsrisk propagationpolicy adaptationreinforcement learning

0 comments

The pith

ReflectiChain encodes supply networks in a 6D latent space with conservation laws to ground LLM policies and separate uncertainty types.

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

The paper claims that LLMs can interpret supply chain policies but lack physical grounding while reinforcement learning optimizes flows without semantic awareness of constraints. ReflectiChain addresses this by building a Generative Supply Chain World Model that maps heterogeneous networks into a 6-dimensional graph-latent space respecting physical conservation, then applies Double-Loop Learning to isolate epistemic uncertainty through KL-trust-region adaptation from aleatoric uncertainty through stochastic rollouts. On a 10-node semiconductor simulation benchmark that includes SIR risk spread, six perturbation types, and ten constraint templates, the method raises rationale consistency by 33 percent while keeping 82 percent operability under shocks and gaining performance under moderate stress. A reader would care because supply-chain decisions that stay both semantically coherent and physically feasible could reduce costly misalignments between stated rules and actual material flows.

Core claim

ReflectiChain bridges the epistemic gap in LLM-driven supply chain agents by means of a Generative Supply Chain World Model that encodes networks into a 6-dim graph-latent space equipped with physical conservation laws together with Double-Loop Learning that isolates epistemic uncertainty via KL-trust-region-bounded policy adaptation from aleatoric uncertainty via stochastic latent rollouts, producing a 33 percent rise in Rationale Consistency Score, 82.3 percent operability under adversarial shocks, and anti-fragile gains on the Semi-Sim benchmark.

What carries the argument

The Generative Supply Chain World Model (SC-WM) that compresses heterogeneous supply networks into a 6-dimensional graph-latent space subject to physical conservation laws, paired with Double-Loop Learning for separating epistemic from aleatoric uncertainty.

If this is right

The method yields a 33 percent increase in Rationale Consistency Score on the Semi-Sim benchmark.
Operability remains at 82.3 percent under the six tested adversarial shocks.
Performance improves by 40.2 percent under moderate pressure, indicating anti-fragile dynamics.
Three epistemic mechanisms are isolated: uncertainty separation, knowledge-boundary detection, and empirical Bayesian policy updating.

Where Pith is reading between the lines

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

The same latent-space approach could be tested on non-semiconductor networks such as food or pharmaceutical logistics to check whether the conservation-law embedding generalizes.
If the KL-trust-region adaptation proves stable, the framework might support online policy updates when new regulatory text arrives without full retraining.
Anti-fragile gains under moderate pressure suggest experiments that deliberately increase constraint density to measure the point at which the separation mechanism breaks.
The five limitation categories mentioned could be turned into targeted ablation tests on the benchmark to quantify each category's contribution to the observed scores.

Load-bearing premise

The 6-dimensional graph-latent space with physical conservation laws accurately represents heterogeneous supply networks and allows reliable separation of epistemic uncertainty from aleatoric uncertainty.

What would settle it

Running the same 10-node semiconductor benchmark with a new perturbation type that violates the assumed conservation laws and observing that rationale consistency gains disappear or operability falls below the reported 82 percent would falsify the central claim.

Figures

Figures reproduced from arXiv: 2606.10359 by Jia Luo.

**Figure 1.** Figure 1: ReflectiChain architecture. (Left) SC-WM: Graph Encoder → Latent zt → Multi-step Rollout → Dual-Head Decoder. (Right) Double-Loop: Reflection-in-Action (Crule-bounded candidate scoring) + Reflection-on-Action (KLtrust-region LoRA updates). (2; 3) and optimization (4; 5). KG-augmented LLMs parse geopolitical risk (6; 7; 8) but remain static interpreters— classifying policies without simulating physical pro… view at source ↗

**Figure 2.** Figure 2: and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Ablation (5 seeds × 3 eps). SC-WM: CEE −49%. Retro RL: RCI −12.8pp. KL trust: variance +81%. Crule: RCI −15.8pp [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Figure 4: Reasoning traces. Constraint: “Do not access Node C.” Channel A blocked. ReAct: Exception Injection. ReflAct: Goal Drift. Ours: Semantic Anchoring via SCWM+Crule. safeguard but human oversight is essential. Scalability: MPNN scales linearly but LLM scoring grows quadratically. Societal impact: Automated agents could be misused; compliance-first design (α>β) and Crule prevent this. We do not advocate 100… view at source ↗

**Figure 6.** Figure 6: Semi-Sim specification. Left: Topology—10 nodes, 4 tiers, ∼30 edges. Certified (solid), uncertified (dashed, red). Right: Prompt structure (4 components) and constraint violation rates (3,000 trajectories). Full Prompt Templates (Single-Column, Color-Coded) ReflectiChain (Ours) — Full System Prompt You are a supply chain AI agent with a Generative World Model (SC-WM). Make sequential decisions under policy… view at source ↗

read the original abstract

AI agents in supply chains face a fundamental epistemic gap: large language models (LLMs) interpret policies but lack physical grounding, while reinforcement learning (RL) optimizes flows but is semantically blind to unstructured constraints. We introduce REFLECTICHAIN, bridging this gap through a Generative Supply Chain World Model (SC-WM) - encoding heterogeneous supply networks into a 6-dim graph-latent space with physical conservation - and Double-Loop Learning that separates epistemic uncertainty (KL-trust-region-bounded policy adaptation) from aleatoric uncertainty (stochastic latent rollouts). On Semi-Sim, a 10-node semiconductor benchmark with SIR risk propagation, 6 perturbation types, and 10 policy constraint templates, REFLECTICHAIN improves Rationale Consistency Score by 33.0% (p < 0.0001, d = 2.78), maintains 82.3% operability under adversarial shocks, and exhibits anti-fragile behavior (+40.2% gain under moderate pressure). We identify three operational epistemic mechanisms - uncertainty separation, knowledge-boundary detection, and empirical Bayesian policy updating - and discuss five limitation categories.

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.

Desk Editor's Note private letter to a colleague

ReflectiChain pairs a 6-dim graph-latent supply chain model with double-loop epistemic/aleatoric separation and reports benchmark gains, but the latent space choice and uncertainty split lack visible support.

read the letter

ReflectiChain tries to connect LLM policy reading with RL flow optimization in supply chains. It does this through a generative world model that squeezes networks into a fixed 6-dimensional graph-latent space while enforcing physical conservation, then uses double-loop learning to bound epistemic uncertainty with KL trust regions and handle aleatoric uncertainty with stochastic rollouts.

The concrete new element is the specific pairing on the Semi-Sim 10-node semiconductor benchmark that includes SIR risk spread, six perturbation types, and ten policy templates. The reported results are a 33% rise in rationale consistency score, 82.3% operability under shocks, and a +40.2% anti-fragile gain under moderate pressure. The paper also names three operational mechanisms: uncertainty separation, boundary detection, and empirical Bayesian updating.

Those numbers and the mechanism list are the parts that could interest people working on grounded agents for operations. The benchmark itself is a focused testbed that makes the claims checkable in principle.

The soft spot is the 6-dimensional latent space. The abstract supplies no derivation, ablation, or reconstruction error showing why six dimensions are enough for heterogeneous supply networks or how conservation laws are actually maintained inside the generative model. The stress-test worry that this setup may not cleanly isolate epistemic uncertainty from model misspecification therefore stands, because nothing in the provided text rules it out.

Without the full equations, dataset details, or code, the p-values and effect sizes cannot be treated as confirmed. This is a limitation of the current presentation rather than a fatal flaw in the idea.

The paper is aimed at researchers who build world models for supply-chain or logistics agents and want to add epistemic handling. A reader already working on uncertainty-aware RL would get the most from it. It deserves peer review so the derivations and experiments can be examined directly.

Referee Report

3 major / 1 minor

Summary. The paper introduces REFLECTICHAIN, which uses a Generative Supply Chain World Model (SC-WM) to encode heterogeneous supply networks into a fixed 6-dimensional graph-latent space enforcing physical conservation laws, combined with Double-Loop Learning to separate epistemic uncertainty (via KL-trust-region-bounded adaptation) from aleatoric uncertainty (via stochastic latent rollouts). On the Semi-Sim 10-node semiconductor benchmark with SIR risk propagation, 6 perturbation types, and 10 policy constraints, it reports a 33.0% improvement in Rationale Consistency Score (p < 0.0001, d = 2.78), 82.3% operability under adversarial shocks, and +40.2% anti-fragile gain under moderate pressure, while identifying three epistemic mechanisms and five limitation categories.

Significance. If the 6-dimensional encoding and uncertainty separation are shown to be valid and the quantitative gains reproducible, the work would offer a concrete bridge between LLM semantic reasoning and physically grounded RL-style optimization in supply-chain settings. The introduction of a controlled benchmark with explicit risk propagation and policy templates is a positive step toward falsifiable evaluation in this domain.

major comments (3)

[Abstract] Abstract: The headline performance numbers rest on the SC-WM's fixed 6-dimensional graph-latent space with physical conservation laws, yet no derivation, ablation, or reconstruction metric is supplied showing why dimension 6 suffices for heterogeneous 10-node networks or how conservation is enforced inside the generative model for SIR dynamics.
[Abstract] Abstract: The claimed separation of epistemic uncertainty (KL-trust-region-bounded policy adaptation) from aleatoric uncertainty (stochastic latent rollouts) is load-bearing for the epistemic-grounding claim, but the manuscript provides neither the explicit form of the KL bound nor any diagnostic showing that the bound isolates epistemic uncertainty rather than conflating it with model misspecification.
[Abstract] Abstract: The reported 33.0% RCS lift, 82.3% operability, and +40.2% anti-fragile gain are presented with p-values and effect sizes, but the absence of methods, dataset details, or code prevents verification that these quantities are supported by the underlying derivations rather than post-hoc choices.

minor comments (1)

[Abstract] The abstract introduces several new terms (SC-WM, Double-Loop Learning, Rationale Consistency Score) without forward references to their formal definitions in later sections.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback highlighting the need for greater transparency on the latent space design, uncertainty separation, and reproducibility. We address each major comment below and will incorporate the requested additions in a revised manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: The headline performance numbers rest on the SC-WM's fixed 6-dimensional graph-latent space with physical conservation laws, yet no derivation, ablation, or reconstruction metric is supplied showing why dimension 6 suffices for heterogeneous 10-node networks or how conservation is enforced inside the generative model for SIR dynamics.

Authors: We agree that the manuscript lacks an explicit derivation and supporting metrics for the fixed 6-dimensional latent space. In revision we will add a dedicated subsection deriving the dimension from the degrees of freedom in a 10-node flow network (state variables plus SIR compartments) and include an ablation table reporting reconstruction MSE and conservation violation error across latent dimensions 4–10 on the Semi-Sim benchmark. The conservation enforcement mechanism (a differentiable projection layer in the decoder) will be stated formally with the corresponding loss term. revision: yes
Referee: [Abstract] Abstract: The claimed separation of epistemic uncertainty (KL-trust-region-bounded policy adaptation) from aleatoric uncertainty (stochastic latent rollouts) is load-bearing for the epistemic-grounding claim, but the manuscript provides neither the explicit form of the KL bound nor any diagnostic showing that the bound isolates epistemic uncertainty rather than conflating it with model misspecification.

Authors: We concur that the explicit form of the KL bound and isolation diagnostics are missing. The revision will state the bound as KL(q_φ(·|s) || p(·|s)) ≤ ε_t where ε_t is adapted from policy-gradient variance, and will add diagnostic plots comparing the bound against held-out epistemic shifts versus aleatoric noise levels to demonstrate separation. revision: yes
Referee: [Abstract] Abstract: The reported 33.0% RCS lift, 82.3% operability, and +40.2% anti-fragile gain are presented with p-values and effect sizes, but the absence of methods, dataset details, or code prevents verification that these quantities are supported by the underlying derivations rather than post-hoc choices.

Authors: Sections 3–4 already contain the Semi-Sim generation procedure, SIR propagation rules, and evaluation protocol. To strengthen verifiability we will expand these sections with additional pseudocode, hyperparameter tables, and exact statistical test specifications. We will also release the implementation code upon acceptance. revision: partial

Circularity Check

0 steps flagged

No significant circularity; derivation chain self-contained against external benchmarks

full rationale

The abstract and description introduce REFLECTICHAIN via a Generative Supply Chain World Model (SC-WM) that encodes networks into a 6-dim graph-latent space with conservation laws and uses Double-Loop Learning for uncertainty separation. No equations, fitted parameters, or self-citations are supplied that reduce any claimed performance metric (RCS improvement, operability, anti-fragile gain) to a quantity defined by those inputs by construction. The reported results are presented as empirical outcomes on the Semi-Sim benchmark rather than predictions forced by the latent-space definition or trust-region bounds. This meets the default expectation of a non-circular paper; the central claims rest on the benchmark evaluation rather than internal redefinition.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

Review is based on abstract only; the paper introduces new modeling entities and domain assumptions whose independent support cannot be assessed without the full text.

axioms (1)

domain assumption Supply networks can be faithfully encoded into a 6-dimensional graph-latent space that enforces physical conservation
Central modeling choice stated in the abstract description of SC-WM.

invented entities (2)

Generative Supply Chain World Model (SC-WM) no independent evidence
purpose: Encode heterogeneous supply networks into 6-dim latent space with conservation
New component introduced to ground LLMs
Double-Loop Learning no independent evidence
purpose: Separate epistemic (KL-bounded) from aleatoric (stochastic rollout) uncertainty
New learning procedure proposed for policy adaptation

pith-pipeline@v0.9.1-grok · 5726 in / 1383 out tokens · 25303 ms · 2026-06-27T13:35:25.341372+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

20 extracted references · 11 canonical work pages · 5 internal anchors

[1]

Large language models in supply chain manage- ment: a systematic literature review and application framework.International Journal of Production Re- search, 0(0):1–41, 2026

Zhe Song, Ying Xie, Lichao Yang, and Yifan Zhao. Large language models in supply chain manage- ment: a systematic literature review and application framework.International Journal of Production Re- search, 0(0):1–41, 2026. doi: 10.1080/00207543. 2026.2641103. URLhttps://doi.org/10. 1080/00207543.2026.2641103

work page doi:10.1080/00207543 2026
[2]

Large language models are zero-shot time series forecasters, 2024

Nate Gruver, Marc Finzi, Shikai Qiu, and An- drew Gordon Wilson. Large language models are zero-shot time series forecasters, 2024. URL https://arxiv.org/abs/2310.07820

work page arXiv 2024
[3]

Shuning Jia, Baijun Song, Canming Ye, and Chun Yuan. M3time: Llm-enhanced multi-modal, multi- scale, and multi-frequency multivariate time se- ries forecasting.Proceedings of the AAAI Con- ference on Artificial Intelligence, 40(27):22265– 22273, Mar. 2026. doi: 10.1609/aaai.v40i27.39383. URLhttps://ojs.aaai.org/index.php/ AAAI/article/view/39383

work page doi:10.1609/aaai.v40i27.39383 2026
[4]

Or-llm-agent: Automating modeling and solving of operations research optimization problem with reasoning large language model.arXiv preprint arXiv:2503.10009, 2025

Bowen Zhang, Pengcheng Luo, Genke Yang, Boon- Hee Soong, and Chau Yuen. Or-llm-agent: Automat- ing modeling and solving of operations research op- timization problems with reasoning llm, 2025. URL https://arxiv.org/abs/2503.10009

work page arXiv 2025
[5]

Deepor: A deep reasoning foundation model for optimization model- ing

Ziyang Xiao, Yuan Jessica Wang, Xiongwei Han, Shisi Guan, Jingyan Zhu, Jingrong Xie, Lilin Xu, Han Wu, Wing Yin Yu, Zehua Liu, et al. Deepor: A deep reasoning foundation model for optimization model- ing. InProceedings of the AAAI Conference on Ar- 4 tificial Intelligence, volume 40, pages 34052–34060, 2026

2026
[6]

Supplygraph: A benchmark dataset for supply chain planning using graph neural networks.arXiv preprint arXiv:2401.15299, 2024

Azmine Toushik Wasi, MD Islam, and Adipto Raihan Akib. Supplygraph: A benchmark dataset for supply chain planning using graph neural networks.arXiv preprint arXiv:2401.15299, 2024

work page arXiv 2024
[7]

The ai-gpr in- dex: Measuring geopolitical risk using artificial intel- ligence

Matteo Iacoviello and Jonathan Tong. The ai-gpr in- dex: Measuring geopolitical risk using artificial intel- ligence. Working Paper, 2026

2026
[8]

Bank for International Settlements, Monetary and Economic Department, 2025

Byeungchun Kwon, Taejin Park, Phurichai Rungcharoenkitkul, and Frank Smets.Parsing the pulse: decomposing macroeconomic sentiment with LLMs. Bank for International Settlements, Monetary and Economic Department, 2025

2025
[9]

Mastering Diverse Domains through World Models

Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models, 2024. URLhttps:// arxiv.org/abs/2301.04104

work page internal anchor Pith review Pith/arXiv arXiv 2024
[10]

Lillicrap, and David Silver

Julian Schrittwieser, Ioannis Antonoglou, Thomas Hubert, Karen Simonyan, Laurent Sifre, Simon Schmitt, Arthur Guez, Edward Lockhart, Demis Has- sabis, Thore Graepel, Timothy Lillicrap, and David Silver. Mastering atari, go, chess and shogi by planning with a learned model.Nature, 588(7839): 604–609, December 2020. ISSN 1476-4687. doi: 10.1038/s41586-020-0...

work page internal anchor Pith review doi:10.1038/s41586-020-03051-4 2020
[11]

Contrastive learning of structured world models,

Thomas Kipf, Elise van der Pol, and Max Welling. Contrastive learning of structured world models,
[12]

URLhttps://arxiv.org/abs/1911. 12247

1911
[13]

Reasoning with Language Model is Planning with World Model

Shibo Hao, Yi Gu, Haodi Ma, Joshua Jiahua Hong, Zhen Wang, Daisy Zhe Wang, and Zhiting Hu. Rea- soning with language model is planning with world model, 2023. URLhttps://arxiv.org/abs/ 2305.14992

work page internal anchor Pith review Pith/arXiv arXiv 2023
[14]

Patel, Paul Pu Liang, Daniel Khashabi, Cheng Peng, Rama Chellappa, Tianmin Shu, Alan Yuille, Yilun Du, and Jieneng Chen

Jiahan Zhang, Muqing Jiang, Nanru Dai, Taiming Lu, Arda Uzunoglu, Shunchi Zhang, Yana Wei, Ji- ahao Wang, Vishal M. Patel, Paul Pu Liang, Daniel Khashabi, Cheng Peng, Rama Chellappa, Tianmin Shu, Alan Yuille, Yilun Du, and Jieneng Chen. World-in-world: World models in a closed-loop world, 2025. URLhttps://arxiv.org/abs/ 2510.18135

work page arXiv 2025
[15]

Reflexion: Language Agents with Verbal Reinforcement Learning

Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal rein- forcement learning, 2023. URLhttps://arxiv. org/abs/2303.11366

work page internal anchor Pith review Pith/arXiv arXiv 2023
[16]

React: Synergizing reasoning and acting in language models,

Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models,
[17]

URLhttps://arxiv.org/abs/2210. 03629
[18]

Learning to (learn at test time): Rnns with expressive hidden states,

Yu Sun, Xinhao Li, Karan Dalal, Jiarui Xu, Ar- jun Vikram, Genghan Zhang, Yann Dubois, Xin- lei Chen, Xiaolong Wang, Sanmi Koyejo, Tatsunori Hashimoto, and Carlos Guestrin. Learning to (learn at test time): Rnns with expressive hidden states,
[19]

URLhttps://arxiv.org/abs/2407. 04620
[20]

Learning from Trials and Errors: Reflective Test-Time Planning for Embodied LLMs

Yining Hong, Huang Huang, Manling Li, Li Fei-Fei, Jiajun Wu, and Yejin Choi. Learning from trials and errors: Reflective test-time planning for embodied llms, 2026. URLhttps://arxiv.org/abs/ 2602.21198. Appendix: Semi-Sim Specification Topology (Fig. 6, left):|V|=10(3S+2M+2D+3R),∼30 edges. Node: inventory, cash, compliance, risk, capac- ity, congestion, q...

work page internal anchor Pith review Pith/arXiv arXiv 2026

[1] [1]

Large language models in supply chain manage- ment: a systematic literature review and application framework.International Journal of Production Re- search, 0(0):1–41, 2026

Zhe Song, Ying Xie, Lichao Yang, and Yifan Zhao. Large language models in supply chain manage- ment: a systematic literature review and application framework.International Journal of Production Re- search, 0(0):1–41, 2026. doi: 10.1080/00207543. 2026.2641103. URLhttps://doi.org/10. 1080/00207543.2026.2641103

work page doi:10.1080/00207543 2026

[2] [2]

Large language models are zero-shot time series forecasters, 2024

Nate Gruver, Marc Finzi, Shikai Qiu, and An- drew Gordon Wilson. Large language models are zero-shot time series forecasters, 2024. URL https://arxiv.org/abs/2310.07820

work page arXiv 2024

[3] [3]

Shuning Jia, Baijun Song, Canming Ye, and Chun Yuan. M3time: Llm-enhanced multi-modal, multi- scale, and multi-frequency multivariate time se- ries forecasting.Proceedings of the AAAI Con- ference on Artificial Intelligence, 40(27):22265– 22273, Mar. 2026. doi: 10.1609/aaai.v40i27.39383. URLhttps://ojs.aaai.org/index.php/ AAAI/article/view/39383

work page doi:10.1609/aaai.v40i27.39383 2026

[4] [4]

Or-llm-agent: Automating modeling and solving of operations research optimization problem with reasoning large language model.arXiv preprint arXiv:2503.10009, 2025

Bowen Zhang, Pengcheng Luo, Genke Yang, Boon- Hee Soong, and Chau Yuen. Or-llm-agent: Automat- ing modeling and solving of operations research op- timization problems with reasoning llm, 2025. URL https://arxiv.org/abs/2503.10009

work page arXiv 2025

[5] [5]

Deepor: A deep reasoning foundation model for optimization model- ing

Ziyang Xiao, Yuan Jessica Wang, Xiongwei Han, Shisi Guan, Jingyan Zhu, Jingrong Xie, Lilin Xu, Han Wu, Wing Yin Yu, Zehua Liu, et al. Deepor: A deep reasoning foundation model for optimization model- ing. InProceedings of the AAAI Conference on Ar- 4 tificial Intelligence, volume 40, pages 34052–34060, 2026

2026

[6] [6]

Supplygraph: A benchmark dataset for supply chain planning using graph neural networks.arXiv preprint arXiv:2401.15299, 2024

Azmine Toushik Wasi, MD Islam, and Adipto Raihan Akib. Supplygraph: A benchmark dataset for supply chain planning using graph neural networks.arXiv preprint arXiv:2401.15299, 2024

work page arXiv 2024

[7] [7]

The ai-gpr in- dex: Measuring geopolitical risk using artificial intel- ligence

Matteo Iacoviello and Jonathan Tong. The ai-gpr in- dex: Measuring geopolitical risk using artificial intel- ligence. Working Paper, 2026

2026

[8] [8]

Bank for International Settlements, Monetary and Economic Department, 2025

Byeungchun Kwon, Taejin Park, Phurichai Rungcharoenkitkul, and Frank Smets.Parsing the pulse: decomposing macroeconomic sentiment with LLMs. Bank for International Settlements, Monetary and Economic Department, 2025

2025

[9] [9]

Mastering Diverse Domains through World Models

Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models, 2024. URLhttps:// arxiv.org/abs/2301.04104

work page internal anchor Pith review Pith/arXiv arXiv 2024

[10] [10]

Lillicrap, and David Silver

Julian Schrittwieser, Ioannis Antonoglou, Thomas Hubert, Karen Simonyan, Laurent Sifre, Simon Schmitt, Arthur Guez, Edward Lockhart, Demis Has- sabis, Thore Graepel, Timothy Lillicrap, and David Silver. Mastering atari, go, chess and shogi by planning with a learned model.Nature, 588(7839): 604–609, December 2020. ISSN 1476-4687. doi: 10.1038/s41586-020-0...

work page internal anchor Pith review doi:10.1038/s41586-020-03051-4 2020

[11] [11]

Contrastive learning of structured world models,

Thomas Kipf, Elise van der Pol, and Max Welling. Contrastive learning of structured world models,

[12] [12]

URLhttps://arxiv.org/abs/1911. 12247

1911

[13] [13]

Reasoning with Language Model is Planning with World Model

Shibo Hao, Yi Gu, Haodi Ma, Joshua Jiahua Hong, Zhen Wang, Daisy Zhe Wang, and Zhiting Hu. Rea- soning with language model is planning with world model, 2023. URLhttps://arxiv.org/abs/ 2305.14992

work page internal anchor Pith review Pith/arXiv arXiv 2023

[14] [14]

Patel, Paul Pu Liang, Daniel Khashabi, Cheng Peng, Rama Chellappa, Tianmin Shu, Alan Yuille, Yilun Du, and Jieneng Chen

Jiahan Zhang, Muqing Jiang, Nanru Dai, Taiming Lu, Arda Uzunoglu, Shunchi Zhang, Yana Wei, Ji- ahao Wang, Vishal M. Patel, Paul Pu Liang, Daniel Khashabi, Cheng Peng, Rama Chellappa, Tianmin Shu, Alan Yuille, Yilun Du, and Jieneng Chen. World-in-world: World models in a closed-loop world, 2025. URLhttps://arxiv.org/abs/ 2510.18135

work page arXiv 2025

[15] [15]

Reflexion: Language Agents with Verbal Reinforcement Learning

Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal rein- forcement learning, 2023. URLhttps://arxiv. org/abs/2303.11366

work page internal anchor Pith review Pith/arXiv arXiv 2023

[16] [16]

React: Synergizing reasoning and acting in language models,

Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models,

[17] [17]

URLhttps://arxiv.org/abs/2210. 03629

[18] [18]

Learning to (learn at test time): Rnns with expressive hidden states,

Yu Sun, Xinhao Li, Karan Dalal, Jiarui Xu, Ar- jun Vikram, Genghan Zhang, Yann Dubois, Xin- lei Chen, Xiaolong Wang, Sanmi Koyejo, Tatsunori Hashimoto, and Carlos Guestrin. Learning to (learn at test time): Rnns with expressive hidden states,

[19] [19]

URLhttps://arxiv.org/abs/2407. 04620

[20] [20]

Learning from Trials and Errors: Reflective Test-Time Planning for Embodied LLMs

Yining Hong, Huang Huang, Manling Li, Li Fei-Fei, Jiajun Wu, and Yejin Choi. Learning from trials and errors: Reflective test-time planning for embodied llms, 2026. URLhttps://arxiv.org/abs/ 2602.21198. Appendix: Semi-Sim Specification Topology (Fig. 6, left):|V|=10(3S+2M+2D+3R),∼30 edges. Node: inventory, cash, compliance, risk, capac- ity, congestion, q...

work page internal anchor Pith review Pith/arXiv arXiv 2026