Large Language Model Empowered Design of Fluid Antenna Systems: Challenges, Frameworks, and Case Studies for 6G

Chan-Byoung Chae; Chao Wang; Kai-Kit Wong; Liang Jin; Zan Li

REVIEW 3 major objections 5 minor 15 references

Large Language Model Empowered Design of Fluid Antenna Systems: Challenges, Frameworks, and Case Studies for 6G

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper argues that large language models can take over the hardest parts of fluid antenna system (FAS) design, and backs the claim with a multiuser case study where an LLM-refined genetic algorithm picks antenna ports needing up to…

desk verdict A plausible vision paper whose single quantitative claim—LLM-enhanced GA beats GA for FAS port selection—is not backed by a compute-controlled comparison, so the headline result is suggestive at best. read the letter →

arxiv 2506.14288 v1 pith:UNLVLNEX submitted 2025-06-17 cs.IT math.IT

Chao Wang , Kai-Kit Wong , Zan Li , Liang Jin , Chan-Byoung Chae This is my paper · ORCID

classification cs.ITmath.IT

keywords fluidantennasystemlargelanguagemodelportselectionprecoderdesignreflectiveevolutionheuristicoptimization6Gchannelprediction

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

The paper aims to establish that large language models can automate and improve the design of fluid antenna systems, a 6G technology where antennas can switch among many ports. FAS design involves jointly selecting ports and designing precoders, a combinatorial, nonconvex problem that conventional deep learning handles poorly because it generalizes badly across environments. To support this vision, the authors run a multiuser case study where an LLM acts as a hyper-heuristic, writing and evolving the crossover operator of a genetic algorithm, and show it consistently outperforms a standard genetic algorithm. If the claim holds, LLMs become a practical tool not just for tuning wireless parameters but for generating the search heuristics themselves.

What carries the argument

ReEvo (reflective evolution), a dual-layer LLM agent architecture that treats genetic-operation code snippets as evolvable individuals: a generation LLM writes crossover operators, a reflection LLM critiques their performance, and selection, short-term reflection, crossover, long-term reflection, and elite mutation refine them over generations. It is paired with an alternating optimization that first solves the combinatorial port-selection subproblem with this LLM-refined genetic algorithm and then solves the convex precoder subproblem as a second-order cone program.

What would settle it

Run the same FAS port-selection problem with both methods given equal total compute budgets, including LLM inference time for the ReEvo side, and a genetic algorithm whose population size and elite rate are tuned under the same budget; if the transmit-power gap drops below the reported 10–17.6% or reverses, the claim that LLM-generated heuristics are superior would be refuted. A simpler check: if a standard GA with a randomly sampled crossover operator matches ReEvo's performance at equal evaluation counts, then the advantage comes from search effort rather than the LLM's reflections.

Watch

Extended reading notes

Core claim

The paper's central claim is that an LLM can improve the heuristic algorithm itself, not just the final configuration. In the case study, the LLM writes, reflects on, and mutates the crossover operator of a genetic algorithm over multiple evolution cycles, and the resulting ReEvo-enhanced GA selects fluid antenna ports that require substantially less downlink transmission power to meet user SINR targets. Across the reported experiments the improvement over a conventional GA ranges from 10% to 17.6% in minimum transmit power, and the gain grows with the number of users. The paper presents this as evidence that LLM-driven automated heuristic design is a viable route for FAS optimization.

Load-bearing premise

The case study compares the LLM-refined genetic algorithm with a conventional genetic algorithm without charging the LLM's inference cost to the refined method, and the baseline uses a single hand-picked parameter set; if the reported gains come from extra compute or poor baseline tuning rather than better heuristics, the central claim is not established.

Editorial extensions

If this is right

Port selection and precoder design for FAS can be automated end-to-end from a scene description, using a multimodal LLM to build the optimization model and then solve it.
LLM-enhanced heuristics converge to lower-power solutions faster than a conventional GA in multiuser FAS setups, with gains growing with the number of users.
The same reflective-evolution mechanism could be applied to other NP-hard wireless design problems where crossover or neighborhood operators are currently hand-designed.
LLMs enable channel extrapolation from a few measured ports by casting CSI prediction as time-series forecasting, avoiding retraining for new environments.

Reading between the lines

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

If the reported gains survive a matched-compute comparison that charges LLM inference time to the ReEvo budget, the advantage would be genuinely algorithmic rather than an artifact of extra computation.
The paper sketches an actor-reflector split where a fast lightweight model makes real-time decisions and a slow LLM periodically revises the heuristic; benchmarking that split directly would test whether the approach is deployable at wireless latencies.
The same reflective-evolution loop could be transferred to cooperative multi-agent FAS, where the paper only outlines an LLM-based MARL framework rather than measuring it; extending the case study to multiple fluid-antenna access points would test whether the gains scale to joint action spaces.
LLM hallucination in problem formulation is a known risk, and the paper's proposed fix (solver-based validation plus retrieval-augmented generation) is plausible but untested; a concrete failure-mode catalog would make the automated modeling claim falsifiable.

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.

Desk Editor's Note

read the letter

You should know three things before reading this. First, it is a magazine-style vision article for IEEE Wireless Communications, not a rigorous algorithmic paper. Second, the genuinely new bit is the application of ReEvo (an existing LLM hyper-heuristic from NeurIPS 2024) to fluid antenna port selection, plus a framework that ties together multimodal LLMs, RAG, and ReEvo for FAS design. Third, the central quantitative claim—that ReEvo-enhanced GA beats a conventional GA by 10–17.6% in transmit power—is not actually established by the experiments as reported.

What the paper does well: the challenge taxonomy in Section II is clear and useful, the framework in Section III is sensible, and the authors are honest about known LLM limitations (hallucination in Remark 2, inference latency in Remark 1). They also cite the prior art fairly, including ReEvo, Port-LLM, and FAS-LLM; self-citation is not an issue here since those earlier results are legitimately prior work.

The soft spot is the case study. The stress-test note is spot on: ReEvo is an evolutionary hyper-heuristic, and each candidate crossover operator it evaluates is itself tested by running a GA. So the total number of fitness evaluations and wall-clock time for ReEvo is far larger than for a single run of the hand-picked GA baseline. The paper never reports computational budget, wall-clock time, number of LLM API calls, or even a definition of what an 'iteration' costs in Figure 4. Without equal total compute, or an ablation replacing the LLM with random code mutation or a parameter sweep, the observed gains cannot be attributed to LLM reasoning rather than to a larger search budget. The paper also gives no error bars or significance tests, and no code or data are released. Section IV.A's automated modeling pipeline is described but not validated; the actual case study appears to start from a hand-written mathematical formulation, not from the vision-to-model pipeline shown in Figure 3. The claims in the paper that these results 'validate the superior efficiency of LLM' are therefore overstrong.

That said, the flaws are fixable or at least reducible. For a magazine article, the framework and the ReEvo application are worth putting in front of reviewers, provided the authors either (a) rerun the experiments with matched compute and report error bars, or (b) explicitly reframe the case study as a proof-of-concept with no efficiency claim. I would not cite the quantitative results in my own work, but I might cite the framework discussion if I were writing about LLM-assisted wireless design.

Who is this for? Readers who want a broad, readable overview of how LLMs could be plugged into FAS design, and who are comfortable treating the case study as illustrative. It deserves a serious referee, but the referee should push hard on the experimental methodology and on the wording of the efficiency claim.

Referee Report

3 major / 5 minor

Summary. The paper proposes a vision and framework for using large language models (LLMs) to design fluid antenna systems (FAS) for 6G, covering channel extrapolation, flexible precoder design, and cooperative multi-agent FAS. The manuscript argues that conventional optimization and deep learning approaches are impractical for FAS due to nonconvexity, combinatorial port selection, and limited generalization, and that LLMs offer reasoning, few-shot adaptation, and automated problem modeling. The main quantitative evidence is a case study in Section IV in which a ReEvo-enhanced genetic algorithm (ReEvo) is compared with a conventional GA for multiuser FAS port selection and precoder optimization, reporting 10%-17.6% lower minimum transmission power in Figures 4 and 5. The paper also describes an automated problem-modeling pipeline based on Qwen-VL-Max and retrieval-augmented generation, though this pipeline is not quantitatively validated.

Significance. If the reported results were properly controlled, the paper would provide a useful demonstration that LLM-driven hyper-heuristics can improve heuristic design for the NP-hard joint port-selection and precoding problem in FAS. The conceptual organization of the paper—its identification of channel extrapolation, automatic problem modeling, and cooperative multi-agent challenges—is valuable for a magazine-style venue, and the use of an existing, public ReEvo framework for an open problem is timely. However, the central quantitative claim, 'These results validate the superior efficiency of LLM in optimizing heuristic algorithms for FAS design' (Section IV.B), is based on a comparison that does not hold the computational budget fixed and does not report statistical variation. The automated modeling pipeline in Section IV.A is also described but never demonstrated end-to-end, so the paper's broader framework claim is supported only by architectural arguments. The paper does not provide code, data, or detailed experiment logs, which further limits independent verification.

major comments (3)

[Section IV.B, Figures 4 and 5] The comparison between ReEvo and GA does not control for computational budget, which is load-bearing for the claim of 'superior efficiency'. The manuscript specifies the GA budget (population size 200, elite rate 20%, 50 generations) but never specifies the ReEvo budget: number of generated crossover-operator candidates, number of inner GA runs used to evaluate each candidate, number of LLM calls, or wall-clock time. Since ReEvo is an evolutionary hyper-heuristic, each candidate operator is itself evaluated by running a GA, so the total number of fitness evaluations is likely many times larger than for a single GA run. Without equal total fitness evaluations, an equal-time budget, or a control that replaces LLM-generated operators with random mutation or a simple parameter sweep, the reported 10%-17.6% power reduction cannot be attributed to LLM reasoning as opposed to a larger search effort. Additionally, the x-axis label 'Iteration Number' in Figure 4 is never defined in terms of per-algorithm cost, which obscures exactly this disparity. The paper must report equal-budget comparisons, include wall-clock and API-call overhead, and provide an ablation to support the stated conclusion.
[Section IV.B, Figures 4 and 5] The experimental evidence lacks statistical rigor for the strength of the claim. The paper reports 'average' curves but provides no error bars, no number of independent trials, no confidence intervals, and no significance tests. The GA baseline uses a single hand-picked parameter set (population 200, elite 20%, 50 generations), with no sensitivity analysis to show that the reported gains are not an artifact of a poorly tuned baseline. Since the central conclusion 'These results validate the superior efficiency of LLM' rests entirely on Figures 4 and 5, the authors should run multiple random seeds for both algorithms, report variance, and ideally perform a paired significance test across the same problem instances. This is a necessary condition for the claim to be reproducible and quantitatively meaningful.
[Section IV.A] The automated problem-modeling pipeline using Qwen-VL-Max, a terminology agent, and RAG is described in detail but never validated end-to-end. The text claims that 'LLM can automate problem modeling and solving' and that the problem description includes 'transmitter and receiver details... the system automatically generates optimization variables... along with the necessary constraints', but no actual generated problem model, constraint list, or objective function is shown, and no comparison against manually constructed models or solver validation is provided. Given that Section IV.A is one of the two pillars of the case study, the paper should include at least one concrete end-to-end example with the exact LLM outputs and a check of correctness (e.g., via a solver or expert review). Without this, the automated modeling component is a stated capability rather than a demonstrated result.

minor comments (5)

[Section IV.A] The text refers to 'IEEE Explore'; the correct name is IEEE Xplore.
[Section III.B.2 / Figure 2] The SOCP formulation displayed in Figure 2 is rendered with garbled mathematical symbols and is not readable. This should be typeset properly so that the constraints and objective are legible.
[Section IV.B] The phrase '2D Jake's fading model' should be '2D Jakes fading model' or 'Jakes' fading model' to match standard terminology.
[References] Reference [7] lists the authors as 'H. Yang, Zhao, S. Lambotharan, and M. Derakhshani', which appears to have an incomplete author entry; this should be corrected.
[Section IV.B] The caption of Figure 4 states '10 single-antenna User FAS System' but the x-axis and its related text call the setting a 10-user system; consider making the wording consistent and defining the channel generation procedure (e.g., number of channel realizations averaged over).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the claimed FAS gains are external simulation comparisons against a standard GA baseline, not fitted inputs or self-cited conclusions.

full rationale

The central quantitative claim in Section IV.B is that ReEvo-enhanced GA achieves 10% to 17.6% lower minimum transmission power than a conventional GA. This claim is supported by the simulations in Figs. 4 and 5, where ReEvo is an existing method imported from the external NeurIPS paper [15] and the GA is a standard baseline. No term in the reported result is fitted from the target outcome, no prediction is a renamed input parameter, and no load-bearing conclusion is justified solely by the authors' own prior work. The paper's self-citations, e.g., to [2], [3], and [4], support background assertions about FAS channel estimation and design challenges but do not determine the case-study outcome. The reported percentage gains are posterior performance differences, not constraints used to construct the methods. The paper acknowledges LLM inference latency in Remark 1 but does not include it in the comparison; that is a computational-fairness concern, which is a correctness risk rather than a circularity defect. No equation or definition in the paper makes the claimed result equivalent to its input by construction, and no uniqueness theorem or prior ansatz is invoked to force the outcome. Therefore no circular step can be identified with the specific reduction required by the review standard.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The case study relies on hand-picked GA parameters and a standard channel model; these are reasonable but unvalidated inputs. The automated modeling pipeline assumes LLMs and RAG produce correct problem formulations, which is not demonstrated. No new physical entities are introduced.

free parameters (3)

GA population size = 200
Hand-chosen in Section IV.B; the reported ReEvo gains may depend on this baseline setting, and no sensitivity analysis is provided.
GA elite rate = 20%
Hand-chosen in Section IV.B; affects the baseline GA performance against which ReEvo is compared.
Number of GA generations = 50
Hand-chosen in Section IV.B; determines the convergence comparison in Figure 4.

assumptions (4)

domain assumption Spatially correlated Rayleigh fading with 2D Jake's correlation models realistic FAS channels
Used in Section IV.B for the case study; if real FAS channels deviate, the reported gains may not transfer.
domain assumption Qwen-VL-Max and RAG can correctly extract scene information and retrieve appropriate problem models
Assumed in Section IV.A; the automated modeling pipeline is described but not quantitatively validated.
domain assumption The alternating optimization between port selection and precoder design converges to a good local optimum
Used in Section IV.B; no convergence proof is provided, though it is a standard heuristic.
standard math SOCP reformulation of the precoder design with fixed port selection is valid
Used in Section IV.B; the SOCP form is stated and relies on standard convex optimization.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Large Language Model Empowered Design of Fluid Antenna Systems: Challenges, Frameworks, and Case Studies for 6G." pith.science (2026). https://pith.science/paper/UNLVLNEX

@misc{pith2026250614288,
  author       = {Pith},
  title        = {Pith review of: Large Language Model Empowered Design of Fluid Antenna Systems: Challenges, Frameworks, and Case Studies for 6G},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UNLVLNEX}},
  note         = {Machine review of arXiv:2506.14288}
}

read the original abstract

The Fluid Antenna System (FAS), which enables flexible Multiple-Input Multiple-Output (MIMO) communications, introduces new spatial degrees of freedom for next-generation wireless networks. Unlike traditional MIMO, FAS involves joint port selection and precoder design, a combinatorial NP-hard optimization problem. Moreover, fully leveraging FAS requires acquiring Channel State Information (CSI) across its ports, a challenge exacerbated by the system's near-continuous reconfigurability. These factors make traditional system design methods impractical for FAS due to nonconvexity and prohibitive computational complexity. While deep learning (DL)-based approaches have been proposed for MIMO optimization, their limited generalization and fitting capabilities render them suboptimal for FAS. In contrast, Large Language Models (LLMs) extend DL's capabilities by offering general-purpose adaptability, reasoning, and few-shot learning, thereby overcoming the limitations of task-specific, data-intensive models. This article presents a vision for LLM-driven FAS design, proposing a novel flexible communication framework. To demonstrate the potential, we examine LLM-enhanced FAS in multiuser scenarios, showcasing how LLMs can revolutionize FAS optimization.

Figures

Figures reproduced from arXiv: 2506.14288 by the authors.

**Figure 2.** ReEvo-enhanced heuristic algorithm for FAS design. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** LLM-enabled intelligent FAS. decision-making by analyzing historical performance data, extracting insights from both successes and failures. LLM-Based Multi-Agent Reinforcement Learning: multi-agent reinforcement learning (MARL) has been extensively employed to develop autonomous agents capable of learning cooperative behaviors in complex environments. As a result, MARL can be effectively applied to train cooperati… view at source ↗

Figures from the paper (1 more)

**Figure 5.** Figure 5: Average Minimum Downlink Transmission Power versus [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 9 canonical work pages

[1]

A tutorial on fluid antenna system for 6G networks: Encompassing communication theory, optimization methods and hard- ware designs,

W. K. New et al., “A tutorial on fluid antenna system for 6G networks: Encompassing communication theory, optimization methods and hard- ware designs,” IEEE Commun. Surv. & Tut., doi:10.1109/COMST.2024. 3498855, 2024

doi:10.1109/comst.2024 2024
[2]

Virtual FAS by learning-based imaginary antennas,

K.-K. Wong et al., “Virtual FAS by learning-based imaginary antennas,” IEEE Wireless Commun. Lett., vol. 13, no. 6, pp. 1581–1585, Jun. 2024

work page 2024
[3]

Fluid antenna system liberating multiuser MIMO for ISAC via deep reinforcement learning,

C. Wang et al. , “Fluid antenna system liberating multiuser MIMO for ISAC via deep reinforcement learning,” IEEE Trans. Wireless Commun., vol. 23, no. 9, pp. 10 879–10 894, Sept. 2024

work page 2024
[4]

AI-empowered fluid antenna systems: Opportunities, challenges, and future directions,

C. Wang et al. , “AI-empowered fluid antenna systems: Opportunities, challenges, and future directions,” IEEE Wireless Commun. , vol. 31, no. 5, pp. 34–41, Oct. 2024

work page 2024
[5]

Generative AI enabled robust data augmentation for wireless sensing in ISAC networks,

J. Wang et al. , “Generative AI enabled robust data augmentation for wireless sensing in ISAC networks,” arXiv preprint, arXiv:2502.12622, Feb. 2025

arXiv 2025
[6]

Generative AI agents with large language model for satellite networks via a mixture of experts transmission,

R. Zhang et al. , “Generative AI agents with large language model for satellite networks via a mixture of experts transmission,” IEEE J. Sel. Areas Commun., vol. 42, no. 12, pp. 3581–3596, Dec. 2024

work page 2024
[7]

FAS-LLM: Large language model-based channel prediction for OTFS-enabled satellite- FAS links,

H. Yang, Zhao, S. Lambotharan, and M. Derakhshani, “FAS-LLM: Large language model-based channel prediction for OTFS-enabled satellite- FAS links,” arXiv preprint, arXiv:2505.09751, May 2025

arXiv 2025
[8]

Port-LLM: A port prediction method for fluid antenna based on large language models,

Y . Zhang, H. Yin, W. Li, E. Bj ¨ornson, and M. Debbah, “Port-LLM: A port prediction method for fluid antenna based on large language models,” arXiv preprint, arXiv:2502.09857, Feb. 2025

arXiv 2025

Show all 15 references

[9]

Channel knowledge map aided channel prediction with measurements-based evaluation,

X. Wang et al., “Channel knowledge map aided channel prediction with measurements-based evaluation,” IEEE Trans. Commun., vol. 73, no. 5, pp. 3622–3636, May 2025

2025
[10]

Knowing what not to do: Leverage language model insights for action space pruning in multi-agent reinforcement learning,

Z. Liu et al. , “Knowing what not to do: Leverage language model insights for action space pruning in multi-agent reinforcement learning,” arXiv preprint, arXiv:2405.16854, May 2024

2024 arXiv
[11]

A survey on multimodal large language models,

S. K. Yin et al., “A survey on multimodal large language models,” arXiv preprint, arXiv:2306.13549, Nov. 2024

2024 arXiv
[12]

Benchmarking large language models in retrieval-augmented generation,

J. Chen, H. Lin, X. Han, and L. Sun, “Benchmarking large language models in retrieval-augmented generation,”Proc. of AAAI Conf. Artificial Intelligence, vol. 38, no. 16, pp. 17 754–17 762, Dec. 2024

2024
[13]

Using large language models for hyperparameter optimization,

M. Zhang, N. Desai, J. Bae, J. Lorraine, and J. Ba, “Using large language models for hyperparameter optimization,” in Proc. NeurIPS Foundation Models for Decision Making Workshop , Dec. 2023

2023
[14]

Llm4cp: Adapting large language models for channel prediction,

B. Liu, X. Liu, S. Gao, X. Cheng, and L. Yang, “Llm4cp: Adapting large language models for channel prediction,” J. Commun. Inf. Netw. , vol. 9, no. 2, pp. 113–125, Jun. 2024

2024
[15]

Reevo: Large language models as hyper-heuristics with reflective evolution,

H. Ye et al. , “Reevo: Large language models as hyper-heuristics with reflective evolution,” in Proc. Adv. Neural Inf. Process. Syst., Dec. 2024, https://github.com/ai4co/reevo

2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

A tutorial on fluid antenna system for 6G networks: Encompassing communication theory, optimization methods and hard- ware designs,

W. K. New et al., “A tutorial on fluid antenna system for 6G networks: Encompassing communication theory, optimization methods and hard- ware designs,” IEEE Commun. Surv. & Tut., doi:10.1109/COMST.2024. 3498855, 2024

doi:10.1109/comst.2024 2024

[2] [2]

Virtual FAS by learning-based imaginary antennas,

K.-K. Wong et al., “Virtual FAS by learning-based imaginary antennas,” IEEE Wireless Commun. Lett., vol. 13, no. 6, pp. 1581–1585, Jun. 2024

work page 2024

[3] [3]

Fluid antenna system liberating multiuser MIMO for ISAC via deep reinforcement learning,

C. Wang et al. , “Fluid antenna system liberating multiuser MIMO for ISAC via deep reinforcement learning,” IEEE Trans. Wireless Commun., vol. 23, no. 9, pp. 10 879–10 894, Sept. 2024

work page 2024

[4] [4]

AI-empowered fluid antenna systems: Opportunities, challenges, and future directions,

C. Wang et al. , “AI-empowered fluid antenna systems: Opportunities, challenges, and future directions,” IEEE Wireless Commun. , vol. 31, no. 5, pp. 34–41, Oct. 2024

work page 2024

[5] [5]

Generative AI enabled robust data augmentation for wireless sensing in ISAC networks,

J. Wang et al. , “Generative AI enabled robust data augmentation for wireless sensing in ISAC networks,” arXiv preprint, arXiv:2502.12622, Feb. 2025

arXiv 2025

[6] [6]

Generative AI agents with large language model for satellite networks via a mixture of experts transmission,

R. Zhang et al. , “Generative AI agents with large language model for satellite networks via a mixture of experts transmission,” IEEE J. Sel. Areas Commun., vol. 42, no. 12, pp. 3581–3596, Dec. 2024

work page 2024

[7] [7]

FAS-LLM: Large language model-based channel prediction for OTFS-enabled satellite- FAS links,

H. Yang, Zhao, S. Lambotharan, and M. Derakhshani, “FAS-LLM: Large language model-based channel prediction for OTFS-enabled satellite- FAS links,” arXiv preprint, arXiv:2505.09751, May 2025

arXiv 2025

[8] [8]

Port-LLM: A port prediction method for fluid antenna based on large language models,

Y . Zhang, H. Yin, W. Li, E. Bj ¨ornson, and M. Debbah, “Port-LLM: A port prediction method for fluid antenna based on large language models,” arXiv preprint, arXiv:2502.09857, Feb. 2025

arXiv 2025