Pith. sign in

REVIEW 5 major objections 7 minor 2 cited by

Reinforcement Learning in hyperbolic space for multi-step reasoning

T0 review · 5 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that replacing the Euclidean transformer inside a GRPO-trained policy with a hyperbolic Poincaré-ball transformer reduces error by 32–44% on hard mathematics problems and 43–45% on nonlinear optimal-control tasks, while…

desk verdict The claimed 32-44% accuracy gains on FrontierMath are actually relative MSE reductions on 11 micro-transformer problems, not task-level accuracy; the hyperbolic-RL combination is worth a careful re-run but the central claim collapses as stated. read the letter →

arxiv 2507.16864 v1 pith:4WKIGLHL submitted 2025-07-21 cs.LG cs.AI

Tao Xu , Dung-Yang Lee , Momiao Xiong This is my paper
classification cs.LGcs.AI
keywords reinforcementlearninghyperbolicgeometryPoincaréballtransformermulti-stepreasoningGroupRelativePolicyOptimizationoptimalcontrolrepresentation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that a reinforcement-learning agent doing multi-step reasoning can be improved by embedding its transformer policy in hyperbolic space rather than Euclidean space. The proposed recipe keeps the same attention and mixture-of-experts backbone and the same group-relative policy optimization training rule, but maps representations into the Poincaré ball and replaces Euclidean operations with curved-space analogues. On the paper's small-scale experiments, this swap is reported to reduce mean squared error by 32–44% on a set of hard mathematics problems, by 43–45% on two nonlinear optimal-control problems, and to cut wall-clock time by 16–32%. A sympathetic reader would care because, if true, it means a cheap geometric change to an RL policy network can substantially improve both error and speed for tasks with hierarchical structure.

What carries the argument

The object that carries the argument is the Poincaré-ball hyperbolic transformer, a transformer whose token embeddings and hidden states live in the unit ball $\mathbb{D}^d = \{x \in \mathbb{R}^d : \|x\| < 1\}$ rather than in Euclidean space. Inputs are pushed into the ball by the exponential map $\exp_0$; attention, layer norm, and feed-forward layers are computed by mapping to the tangent space with $\log_0$, doing the Euclidean operation, and mapping back with $\exp_0$; residual connections use Möbius addition, the curved-space analogue of vector addition on the ball; and this hyperbolic network is trained as the policy inside group-relative policy optimization. The paper also converts multi-head latent attention and mixture-of-experts layers into hyperbolic counterparts.

What would settle it

Run the same GRPO comparison on a held-out suite of multi-step reasoning tasks using exact-answer scoring (does the final output match the known answer) with per-seed error bars and matched wall-clock budgets; if the hyperbolic transformer's relative advantage over the vanilla transformer is not statistically significant, or vanishes under exact-answer scoring, the central claim is not supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that swapping the Euclidean transformer backbone for a hyperbolic one—with the same multi-head latent attention, mixture-of-experts feed-forward, and GRPO training—improves multi-step reasoning. On eleven released mathematics problems it reports relative MSE reductions of 32–44%, which it calls increased accuracy; on two nonlinear optimal-control problems it reports 43–45% better final cost; and wall-clock time drops by 16–32% on the mathematics problems and 16–17% on the control problems. The same pattern appears in a scalar root-finding benchmark, where hyperbolic RL reaches the same error in about 35% fewer updates.

Load-bearing premise

The reported gains rest on the assumption that the very small transformer setup (width 32, one block, batch 1024) and the paper's 'increased accuracy' measure, defined as relative reduction in mean squared error, actually capture multi-step reasoning ability.

Editorial extensions

If this is right

  • If the central claim is correct, hyperbolic GRPO reaches the same error threshold in roughly 33–36% fewer gradient updates than the vanilla transformer on the reported benchmarks.
  • Wall-clock time drops by 16–32% on the mathematics problems and 16–17% on the control problems under the reported CPU settings.
  • Relative MSE reductions of 32–44% on mathematics and 43–45% on control follow from changing the geometry alone, with the same attention, mixture-of-experts, and GRPO infrastructure.
  • The framework supplies a recipe for converting attention, layer norm, feed-forward networks, residual connections, multi-head latent attention, and mixture-of-experts into hyperbolic counterparts for RL policies.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The headline percentages are relative MSE reductions, not exact-solution rates; a reader should not infer that 32–44% more problems were actually solved.
  • A direct test of the mechanism would be a curvature sweep: keeping the architecture fixed and changing the ball's curvature would separate the claim that hyperbolic geometry helps from the claim that this particular choice of curvature $-1$ helps.
  • If the $\log_0$-based residual connection is corrected to a fully tangent-space operation, a clean re-run would tell whether the reported gains come from the hyperbolic representation or from an artifact of the mixed-space addition.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 7 minor

Summary. The paper proposes integrating hyperbolic (Poincaré ball) transformers into GRPO reinforcement learning for multi-step reasoning. It describes component-by-component conversions of the transformer (embedding, layer norm, attention, feed-forward, residual connections, MLA, and DeepSeekMoE) into hyperbolic operations, and reports experiments on a scalar root-finding task, 11 FrontierMath problems, a prime-field problem, and two optimal-control benchmarks. The central claim is that hyperbolic RL improves task accuracy by 32%–44% on FrontierMath and 43%–45% on optimal control while reducing wall-clock time by 16%–32%.

Significance. If the central claim were correct, the paper would make a valuable contribution by showing that hyperbolic geometry in policy transformers improves multi-step reasoning in RL. The systematic mapping of Euclidean transformer components into Poincaré ball operations is a useful reference, and the micro-benchmark tables report raw numbers that allow scrutiny. However, the headline accuracy claims rest on a mislabeled metric, the main experiments lack exact-answer verification and error bars, and the hyperbolic residual operation in §3.2.8 is not well defined as written. The evidence as presented does not support the paper's core quantitative claims, so the contribution remains unvalidated.

major comments (5)
  1. [Results, Figure 6] The abstract's claim of 'improves accuracy by 32%~44%' on FrontierMath is not supported because 'increased accuracy' is defined in Figure 6 as (MSE_Vanilla−T − MSE_Hyper−T)/MSE_Vanilla−T, which is a relative reduction in mean squared error, not task-level accuracy. The FrontierMath problems in Appendix A require exact answers, and the paper reports no exact-match verification, pass rates, or confidence intervals for the 11 problems. A model can halve MSE while still producing zero correct final outputs, so the reported 32–44% figure cannot be interpreted as an accuracy improvement.
  2. [§3.2.8] The hyperbolic residual formula HMlog(z̃) = log0(HMultiHeadAtten(z̃)) + z̃ adds a tangent vector log0(HMultiHeadAtten(z̃)) to a point z̃ in the Poincaré ball, which is a type error. The surrounding text says the operations are done 'in the Euclidean space,' but z̃ is a hyperbolic point and is not mapped to the tangent space before the addition. The residual should be either Möbius addition z̃ ⊕ exp0(log0(HMultiHeadAtten(z̃))) or exp0(log0(z̃) + log0(HMultiHeadAtten(z̃))). As written, the forward pass is not a well-defined hyperbolic operation, so the gradients used in training do not match the stated architecture.
  3. [Tables 3 and 4] The reported 'final MAE ×10^-6' values are mean absolute errors on quantities of order 10^-6, and the text's claims of 43% and 45% 'improvement in final cost' are not supported by any reported policy evaluation on the actual optimal-control objective. For the unicycle problem the reference optimal cost is J* = 0, so any small nonzero MAE produces a large relative change that has no meaning for controller quality. The experiments should report the closed-loop cost J(π) with variance, not a relative reduction of a near-zero regression metric.
  4. [Prime field continuous extensions] This experiment is a single problem, and Table 2 reports miss-prediction rates without specifying the number of rollouts, seeds, or confidence intervals. The statement that Hyper-T 'increases accuracy from 54% to 69%' is based on one problem and cannot support the paper's general accuracy claims. The 'Updates to hit' numbers also have large standard deviations (±800 and ±600), so the reported 33.7% step reduction is not shown to be statistically significant.
  5. [All experiments] Every experiment fixes the Poincaré curvature at −1 and uses a single micro-transformer configuration (width 32, one block, batch 1024). No sensitivity analysis is provided for curvature, depth, width, number of experts, or GRPO batch size. Without such ablations, the reported improvements cannot be distinguished from artifacts of the specific small model setting, and the paper offers no evidence that the gains would persist at scale.
minor comments (7)
  1. [Figure 5 and text] The text says '11 FrontierMath problems' but Figure 5 shows performance on 10 problems; the relationship between Figures 5, 6, and 7 should be clarified.
  2. [Figure 6] The label 'increased accuracy' should be renamed 'relative MSE reduction' to avoid conflating regression loss with task accuracy.
  3. [Table 3] The header contains the typo 'Vamilla-T' instead of 'Vanilla-T'.
  4. [Equations (41)-(43)] Equation (41) is used twice: once for the group-relative advantage and once for the probability ratio ρ1; the equations should be renumbered to avoid ambiguity.
  5. [Wall-clock definition] The definition contains the typo 'Well − clockmodel'; it should be 'Wall-clock'.
  6. [Appendix A, problem 3] The problem statement contains a placeholder 'Type equation here.' that should be removed or filled in.
  7. [References] References 27 and 28 are incomplete; they lack full author and title information and cite non-archival web sources.

Circularity Check

1 steps flagged · score 2.0 of 10

No significant derivation-level circularity; one definitional relabeling of relative MSE reduction as 'accuracy' keeps the score low.

  1. renaming known result [Results, Section 2 (FrontierMath problems), Figure 6 paragraph]
    "Figure 6 shows the increased accuracy of Hyper-T over the Vanilla-T where increased accuracy is defined as MSEVanilla−T−MSEHyper−T / MSEVanilla−T. We observe from Figure 6 that the Hyper-T significantly improves accuracy (32%~44%) compared to Vanilla-T."

    The reported 'increased accuracy' is introduced by definition as the relative MSE reduction between the two models, and the headline claim 'improves accuracy by 32%~44%' is then a restatement of that same ratio rather than an independently measured task-level accuracy. The quantity equals the relative MSE decrease by construction, so the accuracy claim reduces to the metric definition. The underlying MSE decrease itself is an empirical measurement and is not fitted or circular, which is why this is a weak, partial form of circularity.

full rationale

The paper's contribution is empirical: it trains micro-transformers under the same GRPO/MLA/MoE infrastructure and compares final MAE/MSE, gradient steps, and wall-clock. These comparisons are self-contained head-to-head measurements and do not reduce to any fitted parameter, self-citation chain, or imported uniqueness theorem. The GRPO equations are standard group-relative policy optimization, and the hyperbolic-transformer components are presented as an architecture choice with external citations rather than as a result derived from the paper's own prior claims. The main construction-like step is the FrontierMath 'accuracy' metric: the paper defines increased accuracy as relative MSE reduction and then reports that ratio as an accuracy improvement, which is a definitional relabeling rather than independent evidence of solved-problem accuracy. The hyperbolic residual type issue in Section 3.2.8 and the absence of error bars are correctness/validity concerns, not circularity. Overall, the empirical MSE improvements are not forced by construction, so the circularity score is low.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The empirical claim relies on several hand-chosen hyperparameters (width 32, one block, latent length 32, 4 experts, batch 1024) and a fixed curvature of -1, none of which receive sensitivity analysis. The architecture assumes without proof that hyperbolic geometry is beneficial for the tested reasoning tasks. The GRPO objective also depends on a standard-math assumption that is misstated in the paper.

free parameters (6)
  • Poincaré ball curvature c = -1 (assumed)
    Set to -1 'for simplicity' in Sec. 3.2.2; no fitting or sensitivity analysis, and all maps depend on it.
  • model width = 32
    Used for all micro-transformers in the experiments; chosen by hand, no sweep reported.
  • number of transformer blocks = 1
    Single block in every experiment; representative scale is not justified.
  • MLA latent length = 32
    KV/query compression dimension used in experiments; no ablation.
  • number of MoE experts = 4 (top-1 routing)
    MoE configuration in experiments; no ablation.
  • GRPO batch size = 1024 roll-outs per update
    Used for all benchmarks; no sensitivity analysis.
assumptions (4)
  • domain assumption Poincaré ball with curvature -1 is a valid model of hyperbolic geometry for the transformer
    Invoked throughout Sec. 3.2; no justification that curvature -1 suits reasoning data.
  • domain assumption Reasoning data have tree-like hierarchical structure that hyperbolic space represents better
    Stated in the Introduction and Motivation; no empirical test on the benchmarks.
  • standard math GRPO objective and KL penalty are correct as written
    Used in Eq. (44); the KL estimator in Eq. (43) is inconsistent with standard forms, so the objective is suspect.
  • domain assumption The MDP formulation of multi-step reasoning is valid
    Assumed in Section 3; the policy states are transformer representations, so the RL loop is constructed, not derived from data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reinforcement Learning in hyperbolic space for multi-step reasoning." pith.science (2026). https://pith.science/paper/4WKIGLHL

@misc{pith2026250716864,
  author       = {Pith},
  title        = {Pith review of: Reinforcement Learning in hyperbolic space for multi-step reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4WKIGLHL}},
  note         = {Machine review of arXiv:2507.16864}
}
read the original abstract

Multi-step reasoning is a fundamental challenge in artificial intelligence, with applications ranging from mathematical problem-solving to decision-making in dynamic environments. Reinforcement Learning (RL) has shown promise in enabling agents to perform multi-step reasoning by optimizing long-term rewards. However, conventional RL methods struggle with complex reasoning tasks due to issues such as credit assignment, high-dimensional state representations, and stability concerns. Recent advancements in Transformer architectures and hyperbolic geometry have provided novel solutions to these challenges. This paper introduces a new framework that integrates hyperbolic Transformers into RL for multi-step reasoning. The proposed approach leverages hyperbolic embeddings to model hierarchical structures effectively. We present theoretical insights, algorithmic details, and experimental results that include Frontier Math and nonlinear optimal control problems. Compared to RL with vanilla transformer, the hyperbolic RL largely improves accuracy by (32%~44%) on FrontierMath benchmark, (43%~45%) on nonlinear optimal control benchmark, while achieving impressive reduction in computational time by (16%~32%) on FrontierMath benchmark, (16%~17%) on nonlinear optimal control benchmark. Our work demonstrates the potential of hyperbolic Transformers in reinforcement learning, particularly for multi-step reasoning tasks that involve hierarchical structures.

Figures

Figures reproduced from arXiv: 2507.16864 by the authors.

Figure 1
Figure 1. Architecture of hyperbolic transformers 4. Hyperbolic (Poincaré ball) Transformer for Transition Function A hyperbolic transformer can be employed to learn the dynamics of the environment by modeling the transition function in RL, which describes how the environment transitions from the current state 𝑠 to the next state 𝑠′ and issues rewards 𝑟 in response to the actions 𝑎 taken by the agent ( [PITH_FULL_IMAGE:figur… view at source ↗
Figure 2
Figure 2. Transition Function Learning. 4.1. Hyperbolic Transformer Architecture [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Outline of MoE. 5.1.3.1.Basic Architecture Consider the 𝑙 𝑡ℎ layer. Let 𝑢𝑡 𝑙 be the FFN input of the 𝑡 𝑡ℎ token, 𝑁𝑠 and 𝑁𝑟 be the numbers of shared experts and routed experts, respectively, FFN𝑖 (𝑠) (⋅) and FFN𝑖 (𝑟) (⋅) denote the 𝑖 𝑡ℎ shared expert and the 𝑖 𝑡ℎ routed expert, respectively. The FFN output ℎ𝑡 (𝑙) is given by ℎ𝑡 𝑙 = 𝑢𝑡 𝑙 + ∑ FFN𝑖 (𝑠) (𝑢𝑡 𝑙 ) + ∑ 𝑔𝑖,𝑡FFN𝑖 (𝑟) (𝑢𝑡 𝑙 ) 𝑁𝑟 𝑖=1 𝑁𝑠 𝑖=1 , (25) 𝑔𝑖,𝑡 = { 𝑠𝑖,𝑡 … view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: and outlined as follows [PITH_FULL_IMAGE:figures/full_fig_p029_4.png]
Figure 6
Figure 6. Figure 6: The increased accuracy of the Hyper-T vs Vanilla-T [PITH_FULL_IMAGE:figures/full_fig_p043_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. HypEHR: Hyperbolic Modeling of Electronic Health Records for Efficient Question Answering

    cs.AI 2026-04 unverdicted novelty 6.0 of 10

    HypEHR is a hyperbolic embedding model for EHR data that uses Lorentzian geometry and hierarchy-aware pretraining to answer clinical questions nearly as well as large language models but with much smaller size.

  2. HypEHR: Hyperbolic Modeling of Electronic Health Records for Efficient Question Answering

    cs.AI 2026-04 conditional novelty 5.0 of 10

    A 22M-parameter hyperbolic model answers structured EHR questions with accuracy close to LLM-based systems (EHRXQA 89.5%, MIMIC-Instr 76.0%).

Reference graph

Works this paper leans on

30 extracted references · 15 canonical work pages · cited by 1 Pith paper

  1. [1]

    Patil A. 2025. Advancing Reasoning in Large Language Models: Promising Methods and Approaches. arXiv preprint arXiv:2502.03671

    arXiv 2025

  2. [2]

    Multiworks. 2015. What is multi-hop reasoning. https://www.moveworks.com/us/en/resources/ai-terms-glossary/multi-hop-reasoning

    work page 2015

  3. [3]

    Abnave. 2025. OpenAI’s Deep Research: A Leap Towards AGI. . https://medium.com/@pratikabnave97/openais-deep-research-a-leap-towards-agi- e05339823715

    work page 2025

  4. [5]

    Multi-step reinforcement learning: A unifying algorithm

    De Asis K, Hernandez-Garcia J, Holland G, Sutton R. Multi-step reinforcement learning: A unifying algorithm. Proc. Proceedings of the AAAI conference on artificial intelligence, 2018, 32:

    work page 2018

  5. [6]

    Deepseek-AI, Liu A, Feng B, Wang B, Wang B, et al. 2024. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model. arXiv preprint arXiv:2405.04434

    arXiv 2024

  6. [7]

    Ye Y, Zhang T, Jiang W, Huang H. 2025. Process-Supervised Reinforcement Learning for Code Generation. arXiv preprint arXiv:2502.01715

    arXiv 2025

  7. [8]

    Kim S, Kim S. 2024. System-2 Reasoning via Generality and Adaptation. arXiv preprint arXiv:2410.07866

    arXiv 2024

  8. [9]

    Bereska L, Gavves E. 2024. Mechanistic Interpretability for AI Safety--A Review. arXiv preprint arXiv:2404.14082

    arXiv 2024

Show all 30 references
  1. [10]

    Agarwal P, Rahman AA, St-Charles P-L, Prince SJ, Kahou SE. 2023. Transformers in reinforcement learning: a survey. arXiv preprint arXiv:2307.05979

    2023 arXiv

  2. [11]

    Li W, Luo H, Lin Z, Zhang C, Lu Z, Ye D. 2023. A survey on transformers in reinforcement learning. arXiv preprint arXiv:2301.03044

    2023 arXiv

  3. [12]

    Esslinger K, Platt R, Amato C. 2022. Deep transformer q-networks for partially observable reinforcement learning. arXiv preprint arXiv:2206.01078

    2022 arXiv

  4. [13]

    Barto AG, Mahadevan S. 2003. Recent advances in hierarchical reinforcement learning. Discrete event dynamic systems 13:341-79

    2003

  5. [14]

    Chen C, Wu Y-F, Yoon J, Ahn S. 2022. Transdreamer: Reinforcement learning with transformer world models. arXiv preprint arXiv:2202.09481

    2022 arXiv

  6. [15]

    Nickel M, Kiela D. 2017. Poincaré embeddings for learning hierarchical representations. Advances in neural information processing systems 30

    2017

  7. [16]

    Ganea O, Bécigneul G, Hofmann T. 2018. Hyperbolic entailment cones for learning hierarchical embeddings. Proc. International conference on machine learning, 2018:1646-55: PMLR

    2018

  8. [17]

    Wang Z, Wang Y, Zhang Z, Zhou Z, Jin H, et al. 2024. Towards understanding how transformer perform multi-step reasoning with matching operation. arXiv e-prints:arXiv: 2405.15302

    2024 arXiv

  9. [18]

    Liu G, Ji K, Zheng R, Wu Z, Dun C, et al. 2024. Enhancing multi-step reasoning abilities of language models through direct q-function optimization. arXiv preprint arXiv:2410.09302 53

    2024 arXiv

  10. [19]

    van Spengler M, Mettes P. 2025. Distortion-free and GPU-compatible Tree Embeddings in Hyperbolic Space. In ICLR. Singapore

    2025

  11. [20]

    Yang M, Verma H, Zhang DC, Liu J, King I, Ying R. 2024. Hypformer: Exploring efficient transformer fully in hyperbolic space. Proc. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024:3770-81:

    2024

  12. [21]

    Hood. 2021. Reinforcement Learning: Temporal Difference (TD) Learning. . https://www.lancaster.ac.uk/stor-i-student-sites/jordan-j-hood/2021/04/12/reinforcement- learning-temporal-difference-td-learning/

    2021

  13. [22]

    Khrulkov V, Mirvakhabova L, Ustinova E, Oseledets I, Lempitsky V. 2020. Hyperbolic image embeddings. Proc. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020:6418-28:

    2020

  14. [23]

    Tifrea A, Bécigneul G, Ganea O-E. 2018. Poincar\'e glove: Hyperbolic word embeddings. arXiv preprint arXiv:1810.06546

    2018 arXiv

  15. [24]

    Nickel M, Kiela D. 2018. Learning continuous hierarchies in the lorentz model of hyperbolic geometry. Proc. International conference on machine learning, 2018:3779- 88: PMLR

    2018

  16. [25]

    Meng F, Yao Z, Zhang M. 2025. TransMLA: Multi-head Latent Attention Is All You Need. arXiv preprint arXiv:2502.07864

    2025 arXiv

  17. [26]

    Su J, Ahmed M, Lu Y, Pan S, Bo W, Liu Y. 2024. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing 568:127063

    2024

  18. [27]

    Wembo. 2025. DeepSeekMoE: Bridging Efficiency and Capacity in Large Language Models using DeepSeek Model from China. https://joelotepawembo.medium.com/deepseekmoe-bridging-efficiency-and-capacity-in- large-language-models-using-deepseek-model-from-dbd4e852a637

    2025

  19. [28]

    Grootendorst M. 2024. A Visual Guide to Mixture of Experts (MoE)

    2024

  20. [29]

    Wang H, Hao S, Dong H, Zhang S, Bao Y, et al. 2024. Offline Reinforcement Learning for LLM Multi-Step Reasoning. arXiv preprint arXiv:2412.16145

    2024 arXiv

  21. [30]

    Agarwal P, Rahman AA, St-Charles PL, Prince SJD, Kahou, SE. 2023. Transformers in reinforcement learning: a survey. https://doi.org/10.48550/arXiv.2307.05979

    2023 doi arXiv

  22. [31]

    Yong. 2023. Reinforcement Learning: Implementing TD(λ) with function approximation. https://medium.com/mitb-for-all/reinforcement-learning-implementing-td-%CE%BB- with-function-approximation-9b5f9f640aa1

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.