pith. sign in

arxiv: 2606.18620 · v1 · pith:HS7BUHBVnew · submitted 2026-06-17 · 💻 cs.CL · cs.AI

BCL: Bayesian In-Context Learning Framework for Information Extraction

Haoliang Liu , Chengkun Cai , Xu Zhao , Han Zhu , Shizhou Huang , Xinglin Zhang , Tao Chen , Jenq-Neng Hwang
show 2 more authors
Zhang Huaping Lei Li
This is my paper

Pith reviewed 2026-06-26 21:09 UTC · model grok-4.3

classification 💻 cs.CL cs.AI
keywords Bayesian in-context learningparticle filteringinformation extractionsequence labelingrelation classificationlarge language modelslabel representation refinement
0
0 comments X

The pith

BCL applies particle filtering and Bayesian updates to refine label representations for information extraction tasks with large language models.

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

The paper presents BCL as a new optimization method for in-context learning in information extraction. It treats label representations as particles that undergo initialization, observation, weight updates via Bayesian rules, and resampling to produce better outputs. The goal is to fix inconsistent results across model sizes and provide a single method that works for both sequence labeling and relation classification. Experiments indicate these steps deliver steady gains over standard in-context approaches without heavy task-specific changes.

Core claim

BCL is the first optimization framework that uses particle filtering with Bayesian updates to systematically refine label representations across IE tasks. Through four steps of initialization, observation, weight update, and resampling, BCL generalizes to both sequence labeling and relation classification paradigms and yields substantial and consistent improvements over existing approaches.

What carries the argument

The four-step particle filtering process (initialization, observation, weight update, and resampling) combined with Bayesian updates, which iteratively refines label representations inside in-context prompts.

If this is right

  • The same particle-filtering steps can be applied to both sequence labeling and relation classification without redesigning the core loop.
  • Label representations become more stable across different large language model scales.
  • Systematic refinement reduces the need for manual prompt engineering in information extraction.
  • The Bayesian weight update step allows incremental incorporation of new observations into the label set.

Where Pith is reading between the lines

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

  • The approach might transfer to other structured prediction tasks such as named entity recognition or event extraction by reusing the same four steps.
  • If resampling introduces little bias, BCL could serve as a lightweight alternative to full fine-tuning for domain adaptation in extraction.
  • Testing the framework on non-English or low-resource languages would reveal whether the Bayesian updates remain effective when initial label quality varies.

Load-bearing premise

The four-step particle filtering process will produce reliable improvements without requiring task-specific tuning or introducing bias from the choice of initial particles or resampling strategy.

What would settle it

A controlled test on standard IE benchmarks where BCL shows no gain or a decline relative to plain in-context learning, or where performance shifts sharply when the initial particle set or resampling method is altered.

Figures

Figures reproduced from arXiv: 2606.18620 by Chengkun Cai, Han Zhu, Haoliang Liu, Jenq-Neng Hwang, Lei Li, Shizhou Huang, Tao Chen, Xinglin Zhang, Xu Zhao, Zhang Huaping.

Figure 1
Figure 1. Figure 1: Comparison of previous approaches and our work. Top: Previous methods convert IE tasks to leverage models’ stronger code or chat capabilities, mak￾ing performance dependent on model-specific strengths. Bottom: Our approach directly improves IE perfor￾mance via automatically generated semantic patterns, re￾gardless of models’ relative strengths across task types. underperforms one-shot prompting by a substa… view at source ↗
Figure 2
Figure 2. Figure 2: Performance comparison of different methods [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overall framework of BCL showing the particle-based rule optimization pipeline with iterative generation [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Case study showing the prompt template for subcategory generation. The template guides the model to [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Case study demonstrating posterior probability calculation through likelihood estimation. The model [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Case study demonstrating the resampling stage with particle mutation. [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
read the original abstract

Existing information extraction (IE) tasks increasingly adopt in-context learning (ICL) with large language models. However, current approaches either show inconsistent performance across model scales or lack systematic optimization and generalizability. Building on this, we propose BCL (Bayesian In-Context Learning Framework for Information Extraction), the first optimization framework that uses particle filtering with Bayesian updates to systematically refine label representations across IE tasks. Through four steps initialization, observation, weight update, and resampling, BCL generalizes to both sequence labeling and relation classification paradigms. Extensive experiments demonstrate substantial and consistent improvements over existing approaches.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 2 minor

Summary. The paper proposes BCL, a Bayesian In-Context Learning framework for information extraction that applies particle filtering with Bayesian updates via four steps (initialization, observation, weight update, resampling) to refine label representations. It claims to be the first such optimization framework, generalizes to sequence labeling and relation classification, and reports substantial consistent improvements over existing ICL approaches on standard benchmarks.

Significance. If the empirical gains are reproducible and the Bayesian updates operate independently of fitted quantities, BCL could provide a systematic wrapper for optimizing ICL in IE tasks, addressing performance inconsistency across model scales. The particle-filtering integration is a potentially useful contribution if the four-step process yields reliable refinements without task-specific tuning.

minor comments (2)
  1. The abstract and introduction should include a brief comparison table or explicit statement of how BCL's particle-filtering steps differ from prior ICL optimization methods referenced in the framework description.
  2. Clarify in the method section whether the resampling strategy or initial particle choice introduces any bias, with a short ablation or sensitivity analysis to support the claim of reliable improvements.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive summary of BCL and the recommendation of minor revision. We appreciate the recognition that the particle-filtering approach could serve as a useful wrapper for ICL optimization in IE tasks.

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper introduces BCL as a new wrapper framework around existing ICL methods, defining it explicitly via a four-step particle filtering process (initialization, observation, weight update, resampling) that is presented as an independent construction rather than derived from or fitted to the target outputs. No equations, self-citations, or prior results from the same authors are invoked as load-bearing premises that reduce the central claim to a tautology or renaming. The generalization across IE paradigms and empirical improvements are asserted via the method definition and benchmark results, with no evidence of self-definitional steps, fitted inputs relabeled as predictions, or ansatzes smuggled through citations. The derivation chain is therefore self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no free parameters, axioms, or invented entities can be identified from the provided text.

pith-pipeline@v0.9.1-grok · 5643 in / 1125 out tokens · 21125 ms · 2026-06-26T21:09:38.259162+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

76 extracted references · 6 canonical work pages

  1. [1]

    Advances in neural information processing systems , volume=

    Direct preference optimization: Your language model is secretly a reward model , author=. Advances in neural information processing systems , volume=

  2. [2]

    2001 , publisher=

    Sequential Monte Carlo methods in practice , author=. 2001 , publisher=

    2001

  3. [3]

    Proceedings of the IEEE , volume=

    Taking the human out of the loop: A review of Bayesian optimization , author=. Proceedings of the IEEE , volume=. 2015 , publisher=

    2015

  4. [4]

    , author=

    Handbook of Approximate Bayesian Computation. , author=. 2020 , publisher=

    2020

  5. [5]

    Genetics , volume=

    Approximate Bayesian computation in population genetics , author=. Genetics , volume=. 2002 , publisher=

    2002

  6. [6]

    arXiv preprint arXiv:1807.02811 , year=

    A tutorial on Bayesian optimization , author=. arXiv preprint arXiv:1807.02811 , year=

    Pith/arXiv arXiv

  7. [7]

    IEE proceedings F (radar and signal processing) , volume=

    Novel approach to nonlinear/non-Gaussian Bayesian state estimation , author=. IEE proceedings F (radar and signal processing) , volume=. 1993 , organization=

    1993

  8. [8]

    gradient descent

    Automatic prompt optimization with" gradient descent" and beam search , author=. arXiv preprint arXiv:2305.03495 , year=

    arXiv

  9. [9]

    2021 , publisher=

    Feedback systems: an introduction for scientists and engineers , author=. 2021 , publisher=

    2021

  10. [10]

    Advances in neural information processing systems , volume=

    Training language models to follow instructions with human feedback , author=. Advances in neural information processing systems , volume=

  11. [11]

    International conference on machine learning , pages=

    Calibrate before use: Improving few-shot performance of language models , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  12. [12]

    The eleventh international conference on learning representations , year=

    Large language models are human-level prompt engineers , author=. The eleventh international conference on learning representations , year=

  13. [13]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

  14. [14]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

  15. [15]

    and De Meulder, Fien

    Tjong Kim Sang, Erik F. and De Meulder, Fien. Introduction to the C o NLL -2003 Shared Task: Language-Independent Named Entity Recognition. Proceedings of the Seventh Conference on Natural Language Learning at HLT - NAACL 2003. 2003

    2003

  16. [16]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

    1980

  17. [17]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

    Robinson, Arthur L. New Ways to Make Microcircuits Smaller---Duplicate Entry. Science

  18. [18]

    Clancey and Glenn Rennels , abstract =

    Diane Warner Hasling and William J. Clancey and Glenn Rennels , abstract =. Strategic explanations for a diagnostic consultation system , journal =. 1984 , issn =. doi:https://doi.org/10.1016/S0020-7373(84)80003-6 , url =

    work page doi:10.1016/s0020-7373(84)80003-6 1984

  19. [19]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

  20. [20]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Xu, Tianyi and Liu, Jiaxin and Mattei, Nicholas and Zheng, Zizhan , title =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2026 , month = mar, doi =

    2026

  21. [21]

    2026 , eprint=

    Discovering What You Can Control: Interventional Boundary Discovery for Reinforcement Learning , author=. 2026 , eprint=

    2026

  22. [22]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

  23. [23]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

  24. [24]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  25. [25]

    2017 , eprint=

    Attention Is All You Need , author=. 2017 , eprint=

    2017

  26. [26]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet

  27. [27]

    arXiv preprint arXiv:2301.00234 , year=

    A survey on in-context learning , author=. arXiv preprint arXiv:2301.00234 , year=

    Pith/arXiv arXiv

  28. [28]

    International Conference on Learning Representations (ICLR) , year =

    Intrinsic Entropy of Context Length Scaling in LLMs , author =. International Conference on Learning Representations (ICLR) , year =

  29. [29]

    Annual Meeting of the Association for Computational Linguistics (ACL) , year =

    The Role of Deductive and Inductive Reasoning in Large Language Models , author =. Annual Meeting of the Association for Computational Linguistics (ACL) , year =

  30. [30]

    DATA INTELLIGENCE , Year =

    Luo, Man and Xu, Xin and Dai, Zhuyun and Pasupat, Panupong and Kazemi, Mehran and Baral, Chitta and Imbrasaite, Vaiva and Zhao, Vincent , Title =. DATA INTELLIGENCE , Year =. doi:10.3724/2096-7004.di.2024.0012 , Keywords =

    work page doi:10.3724/2096-7004.di.2024.0012 2096

  31. [31]

    DATA INTELLIGENCE , Year =

    Chen, Chao and Luo, Pengfei and Feng, Changkai and Wu, Tian and Jiang, Wenbin and Xu, Tong , Title =. DATA INTELLIGENCE , Year =. doi:10.3724/2096-7004.di.2025.0025 , Unique-ID =

    work page doi:10.3724/2096-7004.di.2025.0025 2096

  32. [32]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Multiple Human Motion Understanding , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  33. [33]

    DATA INTELLIGENCE , Year =

    Wang, Zihang and Liang, Ye and Sun, Wenwei and Lin, Qicong and Xu, Chao and Zhang, Yong , Title =. DATA INTELLIGENCE , Year =. doi:10.3724/2096-7004.di.2025.0078 , Unique-ID =

    work page doi:10.3724/2096-7004.di.2025.0078 2096

  34. [34]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Human Motion Instruction Tuning , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  35. [35]

    The VLDB Journal , volume =

    Li, Shuaimin and Chen, Xuanang and Song, Yuanfeng and Song, Yunze and Zhang, Chen Jason and Hao, Fei and Chen, Lei , title =. The VLDB Journal , volume =. 2025 , month =

    2025

  36. [36]

    Proceedings of the 34th ACM International Conference on Information and Knowledge Management , series =

    Xiong, Xubang and Wong, Raymond Chi-Wing and Song, Yuanfeng , title =. Proceedings of the 34th ACM International Conference on Information and Knowledge Management , series =. 2025 , month =

    2025

  37. [37]

    Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =

    Ji, Yuelyu and Li, Zhuochun and Meng, Rui and He, Daqing , title =. Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =. 2025 , isbn =. doi:10.1145/3726302.3730070 , abstract =

    work page doi:10.1145/3726302.3730070 2025

  38. [38]

    Findings of the Association for Computational Linguistics: EMNLP , year =

    Attention Consistency for LLMs Explanation , author =. Findings of the Association for Computational Linguistics: EMNLP , year =

  39. [39]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    CountLLM: Towards Generalizable Repetitive Action Counting via Large Language Model , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  40. [40]

    Annual Meeting of the Association for Computational Linguistics (ACL) , year =

    Bayesian Optimization for Controlled Image Editing via LLMs , author =. Annual Meeting of the Association for Computational Linguistics (ACL) , year =

  41. [41]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages =

    Is Meta-Learning Out? Rethinking Unsupervised Few-Shot Classification with Limited Entropy , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages =

  42. [42]

    Advances in neural information processing systems , volume=

    Language models are few-shot learners , author=. Advances in neural information processing systems , volume=

  43. [43]

    Proceedings of the AAAI Conference on Artificial Intelligence , author=

    Beyond Step Pruning: Information Theory Based Step-level Optimization for Self-Refining Large Language Models , volume=. Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2026 , month=. doi:10.1609/aaai.v40i41.40798 , number=

    work page doi:10.1609/aaai.v40i41.40798 2026

  44. [44]

    arXiv preprint arXiv:2509.24253 , year=

    MRAG-Suite: A Diagnostic Evaluation Platform for Visual Retrieval-Augmented Generation , author=. arXiv preprint arXiv:2509.24253 , year=

    arXiv

  45. [45]

    Pretraining on the Test Set Is No Longer All You Need: A Debate-Driven Approach to

    Linbo Cao and Jinman Zhao , booktitle=. Pretraining on the Test Set Is No Longer All You Need: A Debate-Driven Approach to. 2025 , url=

    2025

  46. [46]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    GuideNER: Annotation Guidelines Are Better than Examples for In-Context Named Entity Recognition , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  47. [47]

    ACM computing surveys , volume=

    Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing , author=. ACM computing surveys , volume=. 2023 , publisher=

    2023

  48. [48]

    arXiv preprint arXiv:2109.01652 , year=

    Finetuned language models are zero-shot learners , author=. arXiv preprint arXiv:2109.01652 , year=

    Pith/arXiv arXiv

  49. [49]

    arXiv preprint arXiv:2101.06804 , year=

    What Makes Good In-Context Examples for GPT- 3 ? , author=. arXiv preprint arXiv:2101.06804 , year=

    arXiv

  50. [50]

    arXiv preprint arXiv:2112.08633 , year=

    Learning to retrieve prompts for in-context learning , author=. arXiv preprint arXiv:2112.08633 , year=

    arXiv

  51. [51]

    arXiv preprint arXiv:2305.04320 , year=

    Unified demonstration retriever for in-context learning , author=. arXiv preprint arXiv:2305.04320 , year=

    arXiv

  52. [52]

    arXiv , year=

    Learning an Efficient Optimizer via Hybrid-Policy Sub-Trajectory Balance , author=. arXiv , year=

  53. [53]

    arXiv preprint arXiv:2412.07779 , year=

    Evolution of Thought: Diverse and High-Quality Reasoning via Multi-Objective Optimization , author=. arXiv preprint arXiv:2412.07779 , year=

    arXiv

  54. [54]

    arXiv preprint arXiv:2203.15556 , year=

    Training compute-optimal large language models , author=. arXiv preprint arXiv:2203.15556 , year=

    Pith/arXiv arXiv

  55. [55]

    B ayesian Optimization for Controlled Image Editing via LLM s

    Cai, Chengkun and Liu, Haoliang and Zhao, Xu and Jiang, Zhongyu and Zhang, Tianfang and Wu, Zongkai and Lee, John and Hwang, Jenq-Neng and Li, Lei. B ayesian Optimization for Controlled Image Editing via LLM s. Findings of the Association for Computational Linguistics: ACL 2025. 2025

    2025

  56. [56]

    arXiv preprint cs/0306050 , year=

    Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition , author=. arXiv preprint cs/0306050 , year=

    Pith/arXiv arXiv 2003

  57. [57]

    arXiv preprint arXiv:2601.00536 , year=

    Retrieval--Reasoning Processes for Multi-hop Question Answering: A Four-Axis Design Framework and Empirical Trends , author=. arXiv preprint arXiv:2601.00536 , year=

    arXiv

  58. [58]

    (No Title) , year=

    Ace 2005 multilingual training corpus , author=. (No Title) , year=

    2005

  59. [59]

    Bioinformatics , volume=

    GENIA corpus—a semantically annotated corpus for bio-textmining , author=. Bioinformatics , volume=. 2003 , publisher=

    2003

  60. [60]

    Joint European conference on machine learning and knowledge discovery in databases , pages=

    Modeling relations and their mentions without labeled text , author=. Joint European conference on machine learning and knowledge discovery in databases , pages=. 2010 , organization=

    2010

  61. [61]

    A linear programming formulation for global inference in natural language tasks , author=

  62. [62]

    arXiv preprint arXiv:2206.07682 , year=

    Emergent abilities of large language models , author=. arXiv preprint arXiv:2206.07682 , year=

    Pith/arXiv arXiv

  63. [63]

    arXiv preprint arXiv:2202.12837 , year=

    Rethinking the role of demonstrations: What makes in-context learning work? , author=. arXiv preprint arXiv:2202.12837 , year=

    Pith/arXiv arXiv

  64. [64]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    GPT-RE: In-context Learning for Relation Extraction using Large Language Models , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    2023

  65. [65]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Revisiting Relation Extraction in the era of Large Language Models , author=. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  66. [66]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    CodeIE: Large Code Generation Models are Better Few-Shot Information Extractors , author=. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  67. [67]

    International Conference on Intelligent Computing , pages=

    Reassessing the role of chain-of-thought in sentiment analysis: Insights and limitations , author=. International Conference on Intelligent Computing , pages=. 2025 , organization=

    2025

  68. [68]

    arXiv preprint arXiv:2412.03944 , year=

    Chain-of-Thought in Large Language Models: Decoding, Projection, and Activation , author=. arXiv preprint arXiv:2412.03944 , year=

    arXiv

  69. [69]

    arXiv preprint arXiv:2410.02892 , year=

    The role of deductive and inductive reasoning in large language models , author=. arXiv preprint arXiv:2410.02892 , year=

    arXiv

  70. [70]

    arXiv preprint arXiv:2502.18180 , year=

    Chatmotion: A multimodal multi-agent for human motion analysis , author=. arXiv preprint arXiv:2502.18180 , year=

    arXiv

  71. [71]

    arXiv preprint arXiv:2412.00091 , year=

    Graph canvas for controllable 3d scene generation , author=. arXiv preprint arXiv:2412.00091 , year=

    arXiv

  72. [72]

    Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    C-ICL: Contrastive In-context Learning for Information Extraction , author=. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    2024

  73. [73]

    Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    A Simple but Effective Approach to Improve Structured Language Model Output for Information Extraction , author=. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    2024

  74. [74]

    Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence , pages=

    Recall, retrieve and reason: towards better in-context relation extraction , author=. Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence , pages=

  75. [75]

    arXiv preprint arXiv:2302.10205 , year=

    Chatie: Zero-shot information extraction via chatting with chatgpt , author=. arXiv preprint arXiv:2302.10205 , year=

    arXiv

  76. [76]

    arXiv preprint arXiv:2410.21155 , year=

    SciER: An entity and relation extraction dataset for datasets, methods, and tasks in scientific documents , author=. arXiv preprint arXiv:2410.21155 , year=

    arXiv