Pith. sign in

REVIEW 3 major objections 4 minor 49 references

GraphIR claims that feeding an LLM a mutation-oriented architecture state—computation skeleton, mutation surface, validity envelope—alongside executable code improves neural architecture search across multiple benchmarks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 23:51 UTC pith:QIKSB3SC

load-bearing objection GraphIR is a sensible representation idea with a useful diagnostic benchmark, but the CLRS evidence for its central claim is undermined by a constant +95 parameter change on all 30 tasks. the 3 major comments →

arxiv 2608.01633 v1 pith:QIKSB3SC submitted 2026-08-03 cs.LG

GraphIR: Architecture-Level Search States for LLM-Guided Neural Architecture Evolution

Zhen Liu , Wanqi Zhou , Shuanghao Bai , Yuhan Liu , Jinjun Wang , Jingwen Fu This is my paper
classification cs.LG
keywords neural architecture searchlarge language modelsintermediate representationcode-based NASprogram mutationdependency reasoningarchitecture evolutionexecutable neural programs
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that LLM-guided neural architecture search stumbles not because the search space is too small, but because the architecture itself is hidden inside executable code. It introduces GraphIR, an intermediate representation that makes the architecture explicit through three complementary views: a computation skeleton of tensor flow, a mutation surface of editable components, and a validity envelope of interface and shape contracts. The central claim is that appending this state to the mutation prompt improves dependency reasoning, search quality, and end-to-end efficiency. The paper builds a 120-question dependency benchmark and reports the best overall search performance across six downstream benchmarks. If right, the representation of candidate architectures—not just the search algorithm—becomes the decisive design choice in LLM-guided architecture evolution.

Core claim

On its own terms, the paper's discovery is that the obstruction in LLM-guided code-based neural architecture search is representational: source code is written to execute, not to reveal which components are editable, how tensor dimensions are coupled, or which downstream consumers break when a module changes. GraphIR supplies the missing state by extracting a compact, mutation-aligned description from each candidate program and rendering it as additional context for the LLM. The paper reports that this state lets the LLM localize exact producers, trace dependency propagation, and diagnose interface and failure risks, and that these reasoning gains translate into better evolved architectures

What carries the argument

The central object is the GraphIR state, G(P) = (I, O, M, T, F, E), where I and O are input/output interface contracts, M is the set of architecture modules, T is the set of tensor operations, F is deterministic structure facts, and E is producer-consumer dependencies. These six fields compose into three views: the computation skeleton, the mutation surface, and the validity envelope. The state is produced deterministically by combining Python abstract-syntax-tree analysis with PyTorch FX symbolic tracing and shape propagation, then serialized in a fixed field order and appended to the mutation prompt while the source program remains the editing anchor.

Load-bearing premise

The load-bearing premise is that the CLRS experiment measures architecture evolution; the reported model sizes show GraphIR's candidate is exactly 95 parameters larger than the initial model on all 30 tasks, so the comparison may not exercise the mutation mechanism at all.

What would settle it

Inspect the per-task parameter tables on CLRS: if GraphIR's model size minus the initial model size equals 95 for every task, then the claimed 22/30 wins and 89.21% average would need to be rechecked with a controlled search where the evolved models genuinely differ in architecture. A single task where the difference is not 95 would weaken this artifact explanation.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If GraphIR works as claimed, LLM-guided neural architecture search can support open-ended code edits without sacrificing explicit architecture reasoning.
  • Dependency reasoning improves specifically on exact producer localization, dependency next-hop prediction, and interface-failure diagnosis.
  • The reported gains appear with comparable model size and lower end-to-end search time than generic code-graph contexts.
  • Because GraphIR is reconstructed for every child program, newly introduced modules, operations, and dataflow edges become visible in later evolution iterations.
  • The representation is not tied to one network family, so it should transfer across linear, MLP, CNN, GRU, and graph-structured candidates.

Where Pith is reading between the lines

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

  • Beyond the paper's claims, the reported per-task parameter counts on CLRS show that the evolved GraphIR model differs from the benchmark-initial architecture by exactly 95 parameters on every one of the 30 tasks; if this constant offset is real, the CLRS comparison does not test mutation guidance, and the 22/30 wins would need reattribution.
  • A testable extension is to apply GraphIR's three-view state to other executable-program evolution tasks, such as algorithm discovery or code optimization, and measure whether the same dependency-reasoning gains appear.
  • The NAS-Dependency benchmark is generated from CNN evolution traces in the 1D classification setting, so whether the six reasoning dimensions transfer to recurrent, attention-based, or graph neural network programs remains untested by the paper.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes GraphIR, an intermediate representation for LLM-guided code-based neural architecture search. GraphIR supplements the executable source program with three views: a computation skeleton, a mutation surface, and a validity envelope. The authors introduce NAS-Dependency, a 120-question benchmark for architecture-level dependency reasoning, and evaluate GraphIR on six downstream search benchmarks including CLRS, MNIST1D-Shuffle, 20 Newsgroups, CartPole-v1, Yeast, and Breast Cancer Wisconsin. The paper claims that GraphIR achieves the best overall search performance, with 89.21% average accuracy on the 30-task CLRS suite and 22/30 per-task wins.

Significance. The core idea—making architecture-level state explicit for LLM mutation—is plausible and potentially useful. The paper includes controlled comparisons on MNIST1D-Shuffle, ablations of state components, and a deterministic benchmark construction that does not use an LLM to generate gold answers. If the central claims held, the work would be a meaningful contribution to LLM-guided NAS. However, the CLRS results, which are the primary evidence for the central claim, are undermined by a systematic anomaly in the reported model sizes and by uncontrolled baselines. The current evidence does not support the abstract's claim that GraphIR provides an effective interface for LLM-guided architecture evolution.

major comments (3)
  1. [RQ2, Tables 21–23] In every one of the 30 CLRS tasks, the GraphIR model size equals the Initial architecture plus exactly 95 parameters (e.g., 531,913→532,008 for articulation_points; 262,204→262,299 for activity_selector). This constant delta indicates that the 'evolved' architecture is a fixed near-constant perturbation, not the output of a diverse mutation-guided search. The reported 89.21% average and 22/30 wins are therefore the performance of a single fixed modification and cannot be attributed to GraphIR's mutation state. This is load-bearing for the central claim.
  2. [Table 2] EvoPrompting and Triplet-GMPNN results are taken from their original papers rather than re-run under the same evaluation protocol as the other methods. The comparison is therefore not controlled, and the claim that GraphIR achieves 'the best overall search performance' is not established against these baselines. The win counts and average accuracy comparisons are unreliable without a common protocol.
  3. [RQ1, Table 1] NAS-Dependency's six dimensions (producer occurrences, dependency propagation, interface/failure risks) are exactly the information categories GraphIR is designed to emit. The benchmark advantage is partly by construction, so the diagnostic results provide only weak evidence that the representation improves mutation reasoning. The downstream experiments are the relevant test, and they are compromised by the CLRS anomaly noted above.
minor comments (4)
  1. [Table 3] The CNN row reports '1071.55±2.14'; this should be '71.55±2.14' to match Table 6 and the text.
  2. [Table 4] The columns for Invalid and Shape Err are visually merged ('216', '1817'). Please separate the values for readability.
  3. [Global] Terminology is inconsistent: 'CodeRAG-style' in Table 1 and 'CodeGRAG-style' in Table 6 and elsewhere. Please unify.
  4. [Figure 1] Figure 1(b) references supplementary material for details, but the supplement only provides aggregated values for the four benchmarks; per-run raw results would strengthen reproducibility.

Circularity Check

1 steps flagged

NAS-Dependency advantage is partly by construction; downstream search evidence is independent but weakened by uncontrolled CLRS comparison.

specific steps
  1. self definitional [RQ1: Architecture-Level Reasoning; Table 11; Method 'GraphIR State Construction' (Eq. 7)]
    "GraphIR constructs an LLM-visible architecture state G(P) = (I,O,M,T,F,E) ... E contains selected producer-consumer dependencies. ... Producer Occurrence: Identify the exact outermost operation occurrence that directly produces a tensor version. ... Interface/Failure: Infer an observed output shape, affected interface, or deterministic failure category."

    The NAS-Dependency benchmark tasks are defined on exactly the information categories GraphIR is designed to emit: producer-consumer edges (E) and input/output interface contracts (I,O). GraphIR renders these fields directly into the mutation prompt, so answering 'which operation produces tensor v' or 'what is the observed output shape' reduces to reading the supplied state. The Gold answers are independently generated by SourceOracleV2, so the benchmark is not fully tautological, but GraphIR's diagnostic advantage on the highlighted dimensions is substantially by construction: the state is the answer key. This supports the mechanism claim only in the sense that GraphIR exposes what it was built to expose; it does not independently validate that mutation-oriented search improves downstream

full rationale

The paper's central downstream claim -- that GraphIR achieves the best overall search performance -- is not itself circular: it rests on external benchmarks (MNIST1D-Shuffle, 20News, CartPole-v1, Yeast, Breast Cancer) where GraphIR is compared under the same pipeline and the Gold answers do not come from GraphIR. The most significant circularity concern is the NAS-Dependency diagnostic: its six task families (typed edge, fan-out, partition, producer occurrence, next-hop, interface/failure) map directly onto the fields GraphIR emits (I,O,M,T,F,E), so the reported RQ1 advantage is partly by construction rather than an independent test of mutation reasoning. The CLRS evidence has a serious experimental-validity problem that is not circularity: Tables 21-23 show GraphIR's per-task parameter count is exactly Initial+95 on all 30 tasks, and the EvoPrompting/Triplet-GMPNN numbers are imported from their original papers, so the 'best overall search performance' claim is less controlled than presented. These are correctness risks, not derivation-chain circularity. Self-citations to Spark and related prior work are used as baselines or related work, not as load-bearing justification for GraphIR's design, so they do not raise the circularity score. Overall, partial circularity in the diagnostic plus weakened independent support for the central claim justify a score of 5, not higher, because the central derivation is not fully equivalent to its inputs.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

GraphIR is a representation, not a new physical or mathematical entity; SourceOracleV2 and the GraphIR state are software components, not postulated entities. The method relies on hand-chosen vocabularies and contracts, and the CLRS evaluation depends on an unverified assumption of protocol equivalence. The model-size anomaly suggests the CLRS experiments may not measure architecture evolution at all.

free parameters (2)
  • default_feature_width = unspecified constant
    Used to synthesize an example input tensor when the input width cannot be inferred from the first linear module (Method, GraphIR State Construction).
  • rendered entity bound = unspecified
    The number of modules, operations, and edges emitted to the LLM is bounded to keep the mutation context compact; the bound is a hand-chosen design constant.
axioms (5)
  • domain assumption The three-view decomposition (computation skeleton, mutation surface, validity envelope) is the correct abstraction for mutation-aligned NAS state.
    Central design premise from which all subsequent experiments follow.
  • domain assumption torch.fx symbolic tracing recovers all architecture-relevant dataflow, and when it fails, static assignment-level dependencies are a faithful fallback.
    Used in Algorithm 1 steps 5-8; if tracing misses dependencies, the validity envelope is incomplete.
  • domain assumption The fixed neural-module and operation vocabularies cover the mutations that matter in the evaluated search spaces.
    Extraction of M and T depends on matching against these predefined lists; anything outside is invisible to the LLM.
  • domain assumption SourceOracleV2's normalized graph is a correct oracle for dependency reasoning, and its Gold answers are unambiguous.
    The entire NAS-Dependency benchmark relies on this; the paper notes it is not a complete interpreter for arbitrary dynamic Python.
  • domain assumption All compared methods are trained and evaluated under the same CLRS protocol.
    The paper claims 'under Spark's experimental setting' but takes EvoPrompting and Triplet-GMPNN numbers from their original papers; this assumption is questionable.

pith-pipeline@v1.3.0-daily-deepseek · 18167 in / 15614 out tokens · 162431 ms · 2026-08-04T23:51:00.121365+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of GraphIR: Architecture-Level Search States for LLM-Guided Neural Architecture Evolution." pith.science (2026). https://pith.science/paper/QIKSB3SC

@misc{pith2026260801633,
  author       = {Pith},
  title        = {Pith review of: GraphIR: Architecture-Level Search States for LLM-Guided Neural Architecture Evolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QIKSB3SC}},
  note         = {Machine review of arXiv:2608.01633}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large language models (LLMs) enable neural architecture search (NAS) directly over executable neural network programs. However, code-level flexibility does not provide the architecture state needed for effective mutation: LLMs must infer tensor dependencies, editable components, and compatibility constraints from implementation details. To address this representation mismatch, we propose GraphIR, an architecture-aware intermediate representation that supplements executable programs with a mutation-aligned candidate state. GraphIR organizes each candidate through three complementary views: a computation skeleton describing tensor flow, a mutation surface exposing editable modules and operations, and a validity envelope capturing interface contracts, propagated shapes, and downstream dependencies. To evaluate our method, we construct NAS-Dependency, a 120-question benchmark covering six complementary dependency-reasoning dimensions. The diagnostic shows that GraphIR is particularly effective at identifying exact producer occurrences, tracing dependency propagation, and diagnosing interface and failure risks. Across six downstream benchmarks including CLRS, GraphIR achieves the best overall search performance while maintaining comparable model size and favorable end-to-end NAS efficiency when integrated into OpenEvolve. These results show that a mutation-oriented architecture state provides an effective interface between executable neural programs and LLM-guided architecture evolution.

Figures

Figures reproduced from arXiv: 2608.01633 by Jingwen Fu, Jinjun Wang, Shuanghao Bai, Wanqi Zhou, Yuhan Liu, Zhen Liu.

Figure 1
Figure 1. Figure 1: Overview and evaluation of GraphIR. (a) GraphIR extracts a compact, mutation-aligned architecture state from executable neural network code, exposing its computation skeleton, mutation surface, and validity envelope for LLM￾guided search. (b) Cross-task efficiency–performance com￾parison on 20News, CartPole-v1, Yeast, and Breast Cancer Wisconsin Diagnostic. Performance is normalized by the best result on e… view at source ↗

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

49 extracted references · 41 canonical work pages · 1 internal anchor

  1. [1]

    Le , title =

    Barret Zoph and Quoc V. Le , title =

    work page

  2. [2]

    Le , title =

    Esteban Real and Alok Aggarwal and Yanping Huang and Quoc V. Le , title =

    work page

  3. [3]

    Hanxiao Liu and Karen Simonyan and Yiming Yang , title =

    work page

  4. [4]

    So , title =

    Angelica Chen and David Dohan and David R. So , title =. NeurIPS , year =

    work page

  5. [5]

    Cleghorn , title =

    Muhammad Umair Nasir and Sam Earle and Julian Togelius and Steven James and Christopher W. Cleghorn , title =

    work page

  6. [6]

    Zipeng Ji and Guanghui Zhu and Chunfeng Yuan and Yihua Huang , title =

    work page

  7. [7]

    Guan and Barret Zoph and Quoc V

    Hieu Pham and Melody Y. Guan and Barret Zoph and Quoc V. Le and Jeff Dean , title =

    work page

  8. [8]

    Abdelfattah and Abhinav Mehrotra and Lukasz Dudziak and Nicholas Donald Lane , title =

    Mohamed S. Abdelfattah and Abhinav Mehrotra and Lukasz Dudziak and Nicholas Donald Lane , title =

    work page

  9. [9]

    arXiv preprint arXiv:1705.10823 , year=

    Accelerating neural architecture search using performance prediction , author=. arXiv preprint arXiv:1705.10823 , year=

    Pith/arXiv arXiv

  10. [10]

    IEEE Access , volume=

    Multi-fidelity neural architecture search with knowledge distillation , author=. IEEE Access , volume=. 2023 , publisher=

    work page 2023

  11. [11]

    European conference on computer vision , pages=

    Single path one-shot neural architecture search with uniform sampling , author=. European conference on computer vision , pages=. 2020 , organization=

    work page 2020

  12. [12]

    International conference on machine learning , pages=

    Nas-bench-101: Towards reproducible neural architecture search , author=. International conference on machine learning , pages=. 2019 , organization=

    work page 2019

  13. [13]

    arXiv preprint arXiv:2506.05176 , year=

    Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models , author=. arXiv preprint arXiv:2506.05176 , year=

    Pith/arXiv arXiv

  14. [14]

    2025 , publisher =

    OpenEvolve: an open-source evolutionary coding agent , author =. 2025 , publisher =

    2025

  15. [15]

    CoRR , volume =

    Mingkai Zheng and Xiu Su and Shan You and Fei Wang and Chen Qian and Chang Xu and Samuel Albanie , title =. CoRR , volume =

    work page

  16. [16]

    LM -Searcher: Cross-domain Neural Architecture Search with LLM s via Unified Numerical Encoding

    Hu, Yuxuan and Liu, Jihao and Wang, Ke and Zheng, Jinliang and Shi, Weikang and Zhang, Manyuan and Dou, Qi and Liu, Rui and Zhou, Aojun and Li, Hongsheng. LM -Searcher: Cross-domain Neural Architecture Search with LLM s via Unified Numerical Encoding. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025. doi:10.1865...

    work page doi:10.18653/v1/2025.emnlp-main.478 2025

  17. [17]

    Xun Zhou and Xingyu Wu and Liang Feng and Zhichao Lu and Kay Chen Tan , title =

    work page

  18. [18]

    CoRR , volume =

    Zhe Li and Zhiwei Lin and Yongtao Wang , title =. CoRR , volume =

    work page

  19. [19]

    Arjun Sridhar and Yiran Chen , title =

    work page

  20. [20]

    Scaling down deep learning with

    Greydanus, Sam and Kobak, Dmitry , booktitle=. Scaling down deep learning with

    work page

  21. [21]

    1997 , howpublished =

    Mitchell, Tom , title =. 1997 , howpublished =

    1997

  22. [22]

    Terry and John U

    Mark Towers and Ariel Kwiatkowski and Jordan K. Terry and John U. Balis and Gianluca De Cola and Tristan Deleu and Manuel Goul. Gymnasium:. CoRR , volume =

    work page

  23. [23]

    1991 , howpublished =

    Nakai, Kenta , title =. 1991 , howpublished =

    1991

  24. [24]

    and Mangasarian, Olvi L

    Wolberg, William H. and Mangasarian, Olvi L. and Street, Nick and Street, W. Nick , title =. 1993 , howpublished =

    work page 1993

  25. [25]

    Petar Velickovic and Adri. The

    work page

  26. [26]

    arXiv preprint arXiv:2605.04057 , year=

    Structured Progressive Knowledge Activation for LLM-Driven Neural Architecture Search , author=. arXiv preprint arXiv:2605.04057 , year=

    Pith/arXiv arXiv

  27. [27]

    A Generalist Neural Algorithmic Learner , booktitle =

    Borja Ibarz and Vitaly Kurin and George Papamakarios and Kyriacos Nikiforou and Mehdi Bennani and R. A Generalist Neural Algorithmic Learner , booktitle =. 2022 , url =

    work page 2022

  28. [28]

    Thomas Elsken and Jan Hendrik Metzen and Frank Hutter , title =. J. Mach. Learn. Res. , volume =

    work page

  29. [29]

    arXiv preprint arXiv:2301.08727 , year=

    Neural architecture search: Insights from 1000 papers , author=. arXiv preprint arXiv:2301.08727 , year=

    Pith/arXiv arXiv

  30. [30]

    Neural Architecture Search with Bayesian Optimisation and Optimal Transport , booktitle =

    Kirthevasan Kandasamy and Willie Neiswanger and Jeff Schneider and Barnab. Neural Architecture Search with Bayesian Optimisation and Optimal Transport , booktitle =

    work page

  31. [31]

    Colin White and Willie Neiswanger and Yash Savani , title =

    work page

  32. [32]

    Zekang Yang and Wang Zeng and Sheng Jin and Chen Qian and Ping Luo and Wentao Liu , title =

    work page

  33. [33]

    Ottenstein and Joe D

    Jeanne Ferrante and Karl J. Ottenstein and Joe D. Warren , title =

    work page

  34. [34]

    Fabian Yamaguchi and Nico Golde and Daniel Arp and Konrad Rieck , title =

    work page

  35. [35]

    Kechi Zhang and Wenhan Wang and Huangzhao Zhang and Ge Li and Zhi Jin , title =

    work page

  36. [36]

    Clement and Dawn Drain and Neel Sundaresan and Jian Yin and Daxin Jiang and Ming Zhou , title =

    Daya Guo and Shuo Ren and Shuai Lu and Zhangyin Feng and Duyu Tang and Shujie Liu and Long Zhou and Nan Duan and Alexey Svyatkovskiy and Shengyu Fu and Michele Tufano and Shao Kun Deng and Colin B. Clement and Dawn Drain and Neel Sundaresan and Jian Yin and Daxin Jiang and Ming Zhou , title =

    work page

  37. [37]

    Wei Liu and Ailun Yu and Daoguang Zan and Bo Shen and Wei Zhang and Haiyan Zhao and Zhi Jin and Qianxiang Wang , title =

    work page

  38. [38]

    CoRR , volume =

    Kounianhua Du and Renting Rui and Huacan Chai and Lingyue Fu and Wei Xia and Yasheng Wang and Ruiming Tang and Yong Yu and Weinan Zhang , title =. CoRR , volume =

    work page

  39. [39]

    Siru Ouyang and Wenhao Yu and Kaixin Ma and Zilin Xiao and Zhihan Zhang and Mengzhao Jia and Jiawei Han and Hongming Zhang and Dong Yu , title =

    work page

  40. [40]

    Wei Li and Shaogang Gong and Xiatian Zhu , title =

    work page

  41. [41]

    Xuefei Ning and Yin Zheng and Tianchen Zhao and Yu Wang and Huazhong Yang , title =

    work page

  42. [42]

    Reed and Zachary DeVito and Horace He and Ansley Ussery and Jason Ansel , title =

    James K. Reed and Zachary DeVito and Horace He and Ansley Ussery and Jason Ansel , title =. MLSys , publisher =

    work page

  43. [43]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    Pith/arXiv arXiv

  44. [44]

    2026 , eprint =

    Liu, Zhen and Liu, Yuhan and Wang, Jinjun and Liu, Jianyi and Song, Wei and Fu, Jingwen , title =. 2026 , eprint =

    work page 2026

  45. [45]

    2019 , eprint =

    Fu, Jingwen and Zhu, Xiaoyan and Li, Yingbin , title =. 2019 , eprint =

    work page 2019

  46. [46]

    2023 , eprint =

    Fu, Jingwen and Wang, Bohan and Zhang, Huishuai and Zhang, Zhizheng and Chen, Wei and Zheng, Nanning , title =. 2023 , eprint =

    work page 2023

  47. [47]

    Neurocomputing , volume =

    Fu, Jingwen and Zhang, Xiaoyi and Wang, Yuwang and Zeng, Wenjun and Zheng, Nanning , title =. Neurocomputing , volume =. 2024 , doi =

    work page 2024

  48. [48]

    2026 , eprint =

    Fu, Jingwen and Liu, Zhen and Liu, Yuhan and Zhang, He and Zheng, Nanning , title =. 2026 , eprint =

    work page 2026

  49. [49]

    ELVA: Exploring Ranking-Driven Universal Multimodal Retrieval

    Liu, Yuhan and Fu, Pei and Li, Hang and Qi, Yukun and Jiang, Chao and Fu, Jingwen and Liu, Zhen and Qin, Bin and Luo, Zhenbo and Luan, Jian and Xin, Jingmin , title =. 2026 , eprint =. doi:10.48550/arXiv.2606.20280 , note =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2606.20280 2026