Pith. sign in

REVIEW 3 major objections 5 minor 42 references

Scientific inspiration retrieval works better when the system learns, for each target, which reusable principle a candidate paper actually offers.

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 · grok-4.5

2026-07-31 05:26 UTC pith:U46VDZRU

load-bearing objection Trained target-conditioned abstractions beat direct LLM SIR on ResearchBench by a real margin; the closed teacher–grader–student loop is the main caveat, not a reason to dismiss the result. the 3 major comments →

arxiv 2607.28498 v1 pith:U46VDZRU submitted 2026-07-30 cs.IR cs.AIcs.CL

TCA-SIR: Learning Target-Conditioned Abstractions for Scientific Inspiration Retrieval

Yuto Suzuki , Farnoush Banaei-Kashani This is my paper
classification cs.IR cs.AIcs.CL
keywords scientific inspiration retrievaltarget-conditioned abstractionhypothesis generationtransferability scoringremote analogyLLM fine-tuningResearchBench
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.

Finding papers that can seed a new scientific hypothesis is hard when the useful idea comes from another field and shares little topical overlap with the target problem. This paper argues that the right retrieval object is not the paper itself, but a transferable principle extracted from that paper specifically for the target. It trains a model to write a short target-conditioned reasoning-and-abstraction pair for each candidate and to score how transferable that principle is. On a fixed-pool benchmark spanning many scientific domains, the learned method recovers gold inspirations more often than direct large-language-model ranking and than prior abstraction-based approaches, and the written abstractions more clearly name the mechanism that could transfer. The practical payoff is both stronger ranking and an inspectable rationale that later hypothesis composition can use.

Core claim

Reformulating scientific inspiration retrieval as target-conditioned abstraction, and jointly learning to generate those abstractions and to score transferability from their hidden representations, improves recovery of gold inspirations on ResearchBench by more than ten percentage points in HitRate at the top 4% of each pool relative to strong direct LLM baselines, while producing more target-relevant portable mechanisms than an untrained version of the same interface.

What carries the argument

Target-conditioned abstraction (TCA): for each candidate–target pair the model emits a Reasoning/Abstraction pair Φ(i, b) = (r_i, a_i), then predicts a transfer score s_i from the hidden state at a readout token after that generation, and ranks candidates by s_i.

Load-bearing premise

The benchmark’s designated gold papers, plus transfer labels written by the same open model family used as the student, are a good enough stand-in for real scientific transferability.

What would settle it

On held-out ResearchBench-style pools, if a learned TCA ranker no longer beats direct LLM pair scoring and listwise screening on HitRate@top4% and MRR, or if its abstractions no more often name the target-relevant mechanism than an untrained TCA prompt, the central claim fails.

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

If this is right

  • SIR systems should retrieve and expose a target-specific transferable principle, not only rank raw papers by topical similarity.
  • Remote inspirations become rankable when the model selects the facet and depth of abstraction the target needs.
  • The generated Reasoning/Abstraction text can serve as an intermediate explanation for human review and for downstream hypothesis composition.
  • Untrained multi-level or domain-agnostic abstractions are weaker for this task than abstractions optimized jointly with a transfer score.
  • Inference cost scales with target–candidate pairs because each abstraction is conditioned on the target, so large pools need efficiency work.

Where Pith is reading between the lines

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

  • The same target-conditioning idea could be applied outside literature search—to lab protocols, patents, or code repositories—wherever a source must be remapped to a new problem.
  • Score–generation disagreements noted in the paper suggest a natural next test: require the predicted score and the written principle to agree before trusting either.
  • If graded mechanism labels transfer across model families, weaker students might inherit TCA ranking without needing the full teacher stack.

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 / 5 minor

Summary. The paper reformulates Scientific Inspiration Retrieval (SIR) as target-conditioned abstraction (TCA): rather than ranking candidate papers by topical similarity, the retrieval object is a transferable principle Φ(i,b)=(r_i,a_i) extracted from candidate i conditioned on target problem b, with transferability scored by τ(Φ(i,b),b). TCA-SIR jointly fine-tunes a LoRA student to generate Reasoning/Abstraction text and to predict a continuous transfer score from the last-layer hidden state at a controller-inserted readout token. Supervision is built with a label-aware R/A teacher, a generative judge that retains only polarity-consistent examples, and a 5-level mechanism-based grader, all on Llama-3.1-8B-Instruct. On ResearchBench test240 (fixed 75-candidate pools), TCA-SIR reports HitRate@top4% 0.481 vs 0.371 (MOOSE-Chem Direct LLM) and 0.379 (pair-score), with gains on HitRate@top20%, MRR, and NDCG@3, plus ablations and qualitative comparisons arguing that learned abstractions recover target-relevant mechanisms more clearly than untrained TCA prompts.

Significance. If the gains reflect genuine transferability rather than benchmark-specific fitting, the contribution is meaningful for AI-for-Science hypothesis pipelines: SIR is a recognized bottleneck, remote inspirations are poorly handled by topical retrieval, and an explicit, inspectable intermediate is valuable for trusted composition. Strengths include a clean problem reformulation (Eqs. 2–4), controlled same-backbone comparisons, bootstrap CIs, a structured ablation ladder (Table 3), qualitative mechanism examples (Table 2), sensitivity sweeps, and per-discipline breakdowns. The work is a solid empirical IR/methods paper on a timely task, not a foundational theoretical result.

major comments (3)
  1. [§4.2 Training Data Construction] §4.2 / evaluation design (§5.1): The central >10pp HitRate claim rests on a closed supervision loop—same Llama-3.1-8B-Instruct as teacher, judge, grader, and student; grader receives gold binary label/tier as reference metadata; examples are filtered so judged polarity matches ResearchBench gold pairs; success is recovering those same gold papers in fixed 75-pools. Direct LLM baselines lack this gold-conditioned supervision, so Table 1 gains may partly be self-distillation on benchmark labels rather than independent scientific transferability. This is load-bearing: please add at least one external check (held-out human transfer judgments, a stronger/different-family grader, or grader ablated of gold metadata) and report how much of the gap remains.
  2. [Table 3; Appendix B.2] Table 3 and Appendix B.2 (score–generation disagreement): Score-only is close to full TCA-SIR (0.443 vs 0.481 HitRate@top4%), and the appendix notes cases where high ranks coexist with cautious/scope-limited abstractions. The paper’s dual claim—better retrieval and an interpretable transferable rationale—is only partly supported if ranking is driven mainly by the score head. Quantify score–abstraction alignment (e.g., human or LLM rating of whether top-ranked abstractions state an actionable mechanism) and clarify when the generated R/A can be trusted as the explanation of the rank.
  3. [§5.1 Dataset; §6 Results] §5.1 / §6: All headline results use one backbone, one benchmark, and pools of 75 title+abstract candidates with 2–3 dataset-designated golds. The weakest assumption is that optimizing against these golds measures real inspiration quality. Without multi-backbone results, a second corpus, or evaluation beyond gold recovery (e.g., downstream hypothesis-composition quality), the scope of “outperforms prior SIR methods” should be narrowed in the abstract and conclusions, or additional evidence added.
minor comments (5)
  1. [Figure 1] Figure 1 is conceptually clear but low on operational detail; a short caption walk-through tying panels to Eqs. (2)–(4) would help.
  2. [§7 Conclusions] Inference cost of per-pair target-conditioned generation is acknowledged in §7 but not quantified (latency/tokens vs Direct LLM / SourceAbs). A small cost table would strengthen the practicality discussion.
  3. [§3–§4.3] Notation: τ is written both as τ(Φ(i,b),b) and as reading (r_i,a_i,b) from a hidden state; state the implemented interface once in §4.3.
  4. [§4.2; Appendix E] Appendix E prompts and supervision examples are useful; cite them explicitly from §4.2 so readers find the polarity-template controls and grader rubric.
  5. [Throughout] Minor prose issues: missing spaces in compounds (e.g., “seedahypothesis”, “target-conditionedabstractions”) from line-break artifacts; a pass for spacing and hyphenation is needed.

Circularity Check

0 steps flagged

No derivation circularity: held-out ResearchBench ranking is not forced by the training labels or by self-citation.

full rationale

TCA-SIR is a supervised retrieval method, not a first-principles derivation. The central quantitative claim is empirical HitRate/MRR/NDCG on ResearchBench test240 (held-out targets, fixed 75-candidate pools) versus Direct LLM and abstraction baselines on the same backbone. Train/val/test target sets are enforced disjoint (§5.1), so recovering gold inspirations at test time is not true by construction of the fit. Supervision in §4.2 (label-aware R/A teacher, judge filtered to gold polarity, grader given gold binary/tier as reference metadata, same Llama-3.1-8B checkpoint for teacher and student) is a closed-loop labeling design that can bias what the score head learns and raises external-validity questions about “true” transferability—but that is benchmark/self-label contamination risk, not circular reduction of a claimed prediction to its inputs. There is no self-definitional identity, no uniqueness theorem imported from these authors, and no load-bearing self-citation chain. Baselines are evaluated under the same task definition without that gold-conditioned SFT, so the reported gap is an experimental outcome, not an algebraic restatement of the training labels. Score 0; empty steps.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 3 invented entities

The central ranking claim rests on ResearchBench’s operationalization of inspiration, LLM-judged transfer labels, and several training/inference design choices rather than on a closed-form derivation. The ledger below lists the load-bearing modeling assumptions, fitted knobs, and constructed intermediates.

free parameters (6)
  • LoRA rank r and alpha = r=8, alpha=16
    Adapter capacity chosen as r=8, alpha=16 on q_proj/v_proj; affects what the student can represent.
  • Learning rate = 1e-5
    Selected via pilot12 sensitivity; reference checkpoint uses 1e-5 after comparing 5e-6 and 2e-5.
  • Loss weights lambda_LM and lambda_s = 1.0 / 1.0
    Balance generation vs score supervision; reported checkpoint sets both to 1.0 after weight sweep.
  • Five-level transfer grade set = {0.1,0.3,0.5,0.7,0.9}
    Discrete supervision values {0.1,0.3,0.5,0.7,0.9} are a hand-chosen rubric encoding intermediate transfer strength.
  • Judge keep/discard threshold
    Generative judge converts continuous usefulness into binary retention aligned with gold labels; threshold is part of data filtering.
  • Training epochs = 3
    Fine-tune length fixed at 3 epochs for the reported run.
axioms (6)
  • domain assumption ResearchBench gold inspirations in 75-paper pools are the correct retrieval targets for measuring scientific inspiration quality.
    All headline metrics are HitRate/MRR/NDCG against those gold sets (§5.1–5.2).
  • domain assumption Title+abstract text suffices to represent candidate inspirations for transfer judgments.
    Benchmark and method use only title and abstract, not full text or figures (§5.1).
  • ad hoc to paper A transferable principle for SIR can be adequately expressed as a short target-conditioned Reasoning/Abstraction text pair.
    This is the TCA reformulation in §3 and the training target of Φ in §4.
  • ad hoc to paper Llama-3.1-8B-Instruct teacher, judge, and grader labels are adequate supervision for transferability.
    §4.2 intentionally uses one open checkpoint for all label stages and student init.
  • ad hoc to paper Transfer score can be read from the last-layer hidden state at a controller-inserted 'Transfer score:' token after R/A generation.
    Architectural commitment in §4.3 realizing τ without a separate text classifier over R/A.
  • domain assumption Listwise two-round screening (75→ groups of 15 → shortlist → top 3) is a fair common protocol for comparing methods.
    All baselines and TCA-SIR use this harness at temperature 0 (§5.3).
invented entities (3)
  • Target-conditioned abstraction (TCA) retrieval object Φ(i,b)=(r_i,a_i) no independent evidence
    purpose: Replace atomic paper matching with a target-specific transferable principle used for ranking.
    Core reformulation in §3; not an external physical entity, but a new task intermediate the method optimizes.
  • Transfer-score readout head on a controller-inserted token no independent evidence
    purpose: Predict continuous transferability from post-generation hidden states.
    Introduced in §4.3 as the implementation of τ; validated only via ResearchBench ranking gains.
  • Label-aware R/A teacher + generative judge + 5-level grader supervision stack no independent evidence
    purpose: Construct training targets for joint generation and scoring without proprietary teachers.
    Described in §4.2; quality argued via distribution checks and manual inspection, not external human transfer labels.

pith-pipeline@v1.2.0-daily-grok45 · 26241 in / 3800 out tokens · 58553 ms · 2026-07-31T05:26:13.334750+00:00 · methodology

0 comments
read the original abstract

Scientific hypothesis generation for AI for Science typically involves Scientific Inspiration Retrieval (SIR) followed by hypothesis composition. Existing SIR methods rank papers by topical similarity and do not explicitly represent how a candidate inspiration transfers to a target problem. This is especially limiting for remote inspirations, whose value often lies in reusable problem-solving principles rather than topical overlap. Motivated by how humans abstract transferable aspects of a source and remap them to a new target, we reformulate SIR as target-conditioned abstraction (TCA). The retrieval object is a transferable abstract principle extracted from a candidate specifically for the target. We present TCA-SIR, which learns to generate target-conditioned abstractions and uses their representations to predict transferability. On ResearchBench, TCA-SIR outperforms prior SIR methods and direct LLM retrieval, improving HitRate@top4% over MOOSE-Chem by more than 10 percentage points. Learned abstractions also recover target-relevant mechanisms more clearly than an untrained TCA prompt, yielding both stronger retrieval and an interpretable rationale for scientific inspiration.

Figures

Figures reproduced from arXiv: 2607.28498 by Farnoush Banaei-Kashani, Yuto Suzuki.

Figure 1
Figure 1. Figure 1: Three formulations of scientific inspiration retrieval. (a) direct matching leaves the transferable principle implicit (Yang [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of TCA-SIR. Given a target problem and an inspiration candidate, an LLM fine-tuned with LoRA generates [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 5
Figure 5. Figure 5: Instruction-detail sensitivity on pilot12 (Hi [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 3
Figure 3. Figure 3: Learning-rate sensitivity on pilot12 (Hi [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Score-loss weight sensitivity on pilot12 (Hi [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 7
Figure 7. Figure 7: TCA-SIR training loss and validation trend. [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Trained vs. untrained HitRate@top4% on test240. [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: Overall HitRate@top4% ranking on test240. [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: Five-level supervision distributions over accepted training rows. Left: overall distribution. Middle/right: per-class [PITH_FULL_IMAGE:figures/full_fig_p019_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Summary views of teacher-supervision quality. Left: average graded score by class (positive [PITH_FULL_IMAGE:figures/full_fig_p019_13.png] 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

42 extracted references · 3 linked inside Pith

  1. [1]

    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)

  2. [2]

    Classification Problem Solving

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

  3. [3]

    , title =

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

    1980

  4. [4]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

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

  5. [5]

    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 =

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

  6. [6]

    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

  7. [7]

    Poligon: A System for Parallel Problem Solving

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

  8. [8]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

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

  9. [9]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  10. [10]

    2023 , eprint=

    Attention Is All You Need , author=. 2023 , eprint=

    2023

  11. [11]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet

  12. [12]

    Liu, Yujie and Yang, Zonglin and Xie, Tong and Ni, Jinjie and Gao, Ben and Li, Yuqiang and Tang, Shixiang and Ouyang, Wanli and Cambria, Erik and Zhou, Dongzhan , booktitle =

  13. [13]

    Yang, Zonglin and Liu, Wanhao and Gao, Ben and Xie, Tong and Li, Yuqiang and Ouyang, Wanli and Poria, Soujanya and Cambria, Erik and Zhou, Dongzhan , booktitle =

  14. [14]

    Sparking Scientific Creativity via

    Kargupta, Priyanka and Mehri, Shuhaib and Hakkani-Tur, Dilek and Han, Jiawei , journal =. Sparking Scientific Creativity via

  15. [15]

    Gu, Tianyang and Wang, Jingjin and Zhang, Zhihao and Li, HaoHong , journal =

  16. [16]

    1964 , publisher =

    The Act of Creation , author =. 1964 , publisher =

    1964

  17. [17]

    2004 , publisher =

    The Creative Mind: Myths and Mechanisms , author =. 2004 , publisher =

    2004

  18. [18]

    Garikaparthi, Aniketh and Patwardhan, Manasi and Kanade, Aditya Sanjiv and Hassan, Aman and Vig, Lovekesh and Cohan, Arman , booktitle =

  19. [19]

    Wang, Qingyun and Downey, Doug and Ji, Heng and Hope, Tom , booktitle =

  20. [20]

    Sternlicht, Noy and Hope, Tom , booktitle =

  21. [21]

    Proceedings of the 23rd

    Accelerating Innovation Through Analogy Mining , author =. Proceedings of the 23rd

  22. [22]

    Findings of the Association for Computational Linguistics:

    Abstraction-of-Thought Makes Language Models Better Reasoners , author =. Findings of the Association for Computational Linguistics:

  23. [23]

    Qu, Yuxiao and Singh, Anikait and Lee, Yoonho and Setlur, Amrith and Salakhutdinov, Ruslan and Finn, Chelsea and Kumar, Aviral , journal =

  24. [24]

    Gao, Silin and Bosselut, Antoine and Bengio, Samy and Abbe, Emmanuel , journal =

  25. [25]

    and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , journal =

    Hu, Edward J. and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , journal =

  26. [26]

    Lu, Chris and Lu, Cong and Lange, Robert Tjarko and Foerster, Jakob and Clune, Jeff and Ha, David , journal =. The

  27. [27]

    Yamada, Yutaro and Lange, Robert Tjarko and Lu, Cong and Hu, Shengran and Lu, Chris and Foerster, Jakob and Clune, Jeff and Ha, David , journal =. The

  28. [28]

    Towards End-to-End Automation of

    Lu, Chris and Lu, Cong and Lange, Robert Tjarko and Yamada, Yutaro and Hu, Shengran and Foerster, Jakob and Ha, David and Clune, Jeff , journal =. Towards End-to-End Automation of. 2026 , publisher =

    2026

  29. [29]

    Wei, Jiaqi and Yang, Yuejin and Zhang, Xiang and Chen, Yuhan and Zhuang, Xiang and Gao, Zhangyang and Zhou, Dongzhan and Wang, Guangshuai and Gao, Zhiqiang and Cao, Juntai and others , journal =. From

  30. [30]

    Si, Chenglei and Yang, Diyi and Hashimoto, Tatsunori , booktitle =. Can

  31. [31]

    Nature , volume =

    Mathematical Discoveries from Program Search with Large Language Models , author =. Nature , volume =. 2024 , publisher =

    2024

  32. [32]

    Yang, Zonglin and Liu, Wanhao and Gao, Ben and Liu, Yujie and Li, Wei and Xie, Tong and Bing, Lidong and Ouyang, Wanli and Cambria, Erik and Zhou, Dongzhan , journal =

  33. [33]

    arXiv preprint arXiv:2506.13131 , year =

    Novikov, Alexander and V. arXiv preprint arXiv:2506.13131 , year =

    Pith/arXiv arXiv

  34. [34]

    Towards Multimodal Data-Driven Scientific Discovery Powered by

    Liu, Fan and Zeng, Xiaozhao and Liu, Hao , booktitle =. Towards Multimodal Data-Driven Scientific Discovery Powered by

  35. [35]

    , booktitle =

    Radensky, Marissa and Shahid, Simra and Fok, Raymond and Siangliulue, Pao and Hope, Tom and Weld, Daniel S. , booktitle =

  36. [36]

    Baek, Jinheon and Jauhar, Sujay Kumar and Cucerzan, Silviu and Hwang, Sung Ju , booktitle =

  37. [37]

    Nature , volume =

    Autonomous Chemical Research with Large Language Models , author =. Nature , volume =. 2023 , publisher =

    2023

  38. [38]

    and Pak, John E

    Swanson, Kyle and Wu, Wesley and Bulaong, Nathan L. and Pak, John E. and Zou, James , journal =. The Virtual Lab:. 2024 , note =

    2024

  39. [39]

    and Bucci, Matthew and Nazir, Abbas and Chen, Bob and Trinh, Linh and Zhang, Serena and Huang, Kexin and Chandrasekar, Vineethkrishna and others , journal =

    Wang, Hanchen and He, Yichun and Coelho, Paula P. and Bucci, Matthew and Nazir, Abbas and Chen, Bob and Trinh, Linh and Zhang, Serena and Huang, Kexin and Chandrasekar, Vineethkrishna and others , journal =. 2025 , publisher =

    2025

  40. [40]

    arXiv preprint arXiv:2411.03743 , year =

    Automating Exploratory Proteomics Research via Language Models , author =. arXiv preprint arXiv:2411.03743 , year =

    Pith/arXiv arXiv

  41. [41]

    Ou, Jiefu and Walden, William Gantt and Sanders, Kate and Jiang, Zhengping and Sun, Kaiser and Cheng, Jeffrey and Jurayj, William and Wanner, Miriam and Liang, Shaobo and Morgan, Candice and others , journal =

  42. [42]

    arXiv preprint arXiv:2311.09706 , year =

    Towards Autonomous Hypothesis Verification via Language Models with Minimal Guidance , author =. arXiv preprint arXiv:2311.09706 , year =

    Pith/arXiv arXiv