Pith. sign in

REVIEW 3 major objections 4 minor 21 references

An agentic workflow can be adapted per input at inference time by grafting better region-level operators onto an offline-optimized structure, without retraining or whole-workflow search.

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 08:48 UTC pith:A5MTFBLA

load-bearing objection GRAFT is a genuinely new inference-time workflow adaptation idea with strong math/code ablations, but the headline QA gains are not causally supported by the method's own proxy-fidelity data. the 3 major comments →

arxiv 2608.02353 v1 pith:A5MTFBLA submitted 2026-08-03 cs.CL

Global Optimization and Inference-Time Region Grafting for Agentic Workflows

Donghyeok Koh , Gyuwan Kim , Jinyeong Bak , Seung-Hoon Na , Tao Yang , Haneol Jang , Cheoneum Park This is my paper
classification cs.CL
keywords agentic workflowsinference-time adaptationworkflow optimizationSESE region decompositionlabel-free quality proxyself-consistencyconfiguration memorylarge language models
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 claims that a workflow for an LLM agent—the ordered set of operators that plan, retrieve, reason, verify, and format—need not be frozen after it is optimized offline. Instead, at inference time, each input can trigger local 'grafting': only the regions of the workflow that are failing (according to label-free quality signals) are swapped for alternative operators, while the rest of the globally validated structure stays intact. The paper shows this training-free, per-instance adaptation improves average task scores over prior workflow-optimization methods by about 3.8 points under matched settings, and that the gains are driven by the label-free proxy signal and the local grafting mechanism, not merely by the initial workflow. It also shows the same optimized workflow transfers to stronger executor models without re-optimization, suggesting workflows function as adaptable execution policies rather than static artifacts.

Core claim

On its own terms, the paper's central discovery is that a task-level optimal workflow provides a structural prior, not a final answer: decomposing it into single-entry/single-exit regions and re-selecting only the operator within each region per input, using a label-free quality score Q and a coupling guard that blocks replacements degrading cross-boundary support, yields consistent gains over both fixed-workflow search and query-specific supernet sampling. The authors demonstrate this by exhaustive/coordinate-ascent global search to obtain a seed workflow, then a staged local edit-space expansion, a UCB-based proposer, and a configuration memory that reuses winning region assignments across

What carries the argument

The load-bearing machinery is the SESE (single-entry, single-exit) region decomposition of the workflow, which lets a replacement inside a region preserve the workflow's connectivity; the label-free local quality proxy Q (weighted self-consistency, groundedness measured as lexical evidence coverage, and verifier signals such as code test pass rate) that scores each candidate without ground truth; and the coupling guard that accepts a replacement only if it strictly improves Q and does not reduce boundary support by more than epsilon. A UCB-style proposer and a configuration memory keyed by input signature amortize the search across queries.

Load-bearing premise

The method assumes its label-free quality score Q ranks candidate region configurations in the same order as true answer correctness; on knowledge-intensive tasks the paper's own measurements put that correlation at 0.17 or below, so the grafting decisions there rest on a weak signal.

What would settle it

Run GRAFT on a knowledge-intensive benchmark (one with measured proxy-correctness correlation at or below 0.17) with the proxy score Q replaced by a random score for candidate ranking. If end-task accuracy does not drop compared with using the real Q, then the reported gains on such tasks are not attributable to the label-free proxy. Conversely, if accuracy drops sharply despite Q's low measured correlation, then the correlation metric is not capturing the signal that matters.

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

If this is right

  • Per-input adaptation can be added to any pre-optimized workflow without retraining, by limiting search to role-scoped regions rather than re-searching the whole graph.
  • Removing either local grafting or the label-free proxy drops average performance by roughly 4 and 3.5 points respectively, so the gains are attributable to the inference-time mechanism itself.
  • Winning region configurations generalize across similar inputs: reuse rates reach 0.86–0.97 on uniform math tasks, cutting search cost by about half after the first decile of inputs.
  • Swapping in a stronger executor on the same seed workflow improves accuracy without re-optimization, especially on reasoning-heavy tasks; the workflow thus behaves as an adaptive execution policy rather than a static artifact.
  • The label-free proxy is far weaker on knowledge-intensive QA (correlation at most 0.17 with correctness) than on math and code (0.23–0.62), so the method's applicability to such tasks is conditioned on proxy fidelity.

Where Pith is reading between the lines

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

  • A testable extension: on knowledge-heavy tasks with near-zero proxy-correctness correlation, GRAFT's selection could be replaced with uncertainty-aware or retrieval-faithfulness signals; if gains persist, the proxy assumption is less load-bearing than Table 3 suggests.
  • The SESE decomposition suggests a broader principle: any structured generation pipeline with well-defined interfaces can be adapted at inference time by local bandit-style search, not just agent workflows; the same machinery could apply to RAG pipelines or multi-stage code generation.
  • The memory of winning configurations is effectively a form of transductive learning over workflows; one could pre-populate it from a development set and then measure cold-start versus warm-start performance, a comparison the paper does not directly report.
  • If proxy quality and correctness are only weakly correlated, the coupling guard's epsilon threshold becomes the de facto controller of risk; tuning epsilon per task may matter more than the paper's fixed 0.02 suggests.

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 GRAFT, a training-free inference-time adaptation method that takes a globally optimized agentic workflow Wc*, decomposes it into SESE regions, and for each input locally replaces region operators with alternatives from a role-specific library, guided by label-free quality proxies Q (self-consistency, groundedness, code-test verifier) and a coupling guard. The global workflow is searched offline once per task; at inference, a proposer prioritizes regions, a memory caches winning configurations, and a bounded number of passes propagates changes. Experiments compare GRAFT with prior workflow-optimization methods under two setups: the MaAS setup (math/code, gpt-4o-mini) and the BayesFlow setup (math, QA, knowledge, Claude-Sonnet). The paper reports average 87.44 on the MaAS setup, 3.85 points above MaAS, and 84.1 on the BayesFlow setup, 3.3 above BayesFlow. It also reports ablations showing local grafting and the proxy contribute to gains on the MaAS suite, and an executor-transfer analysis.

Significance. If established, the contribution is meaningful: it moves agentic workflow optimization from a one-time static structure to an instance-wise adaptable policy without training, using a principled region decomposition and label-free scoring. The paper is honest enough to include a proxy-fidelity table (Table 3) that quantifies the correlation between Q and correctness, and it shows a clear cost-accuracy tradeoff via configuration memory and executor transfer. The SESE-region grafting idea and the coupling guard are novel and plausible. However, the evidence for the headline generality claim across QA and knowledge-intensive tasks is incomplete, because the only component ablation is on the math/code suite, and the paper's own diagnostic predicts poor proxy behavior on QA.

major comments (3)
  1. [§4.3 / Table 2] The component ablation that attributes gains to local grafting and the label-free proxy is run only on the five math/code benchmarks of the MaAS suite. The claim in §4.2 that 'inference-time local grafting is also effective for QA and knowledge-intensive tasks' rests solely on the overall BayesFlow numbers in Table 1b, with no matched ablation on HotpotQA, DROP, MMLU-Pro, or GPQA. This is load-bearing because Eq. (3) selects candidates by arg-max over Q, and Table 3 reports Spearman correlations of at most 0.17 (GPQA 0.04, HotpotQA 0.06) for exactly those tasks; with Q near random, the mechanism should not systematically improve answers. Without a QA/knowledge ablation, the gains in Table 1b could come from the offline Wc* or from baseline differences rather than from grafting. Please add matched ablations (e.g., '−local grafting' and '−label-free proxy') for the QA/knowledge benchmarks,
  2. [Table 3 vs. Table 1b] Table 3 lists GRAFT operating statistics with HotpotQA F1 74.58, DROP F1 85.55, MMLU-Pro 62.87, and GPQA 37.74, while Table 1b reports GRAFT scores of 80.0, 92.7, 83.4, and 74.2 for the same datasets. The setup for Table 3 is not stated. If these are from the BayesFlow setup, they are inconsistent with Table 1b; if from a different setup, the proxy-fidelity analysis cannot be used to interpret the main reported QA gains. Please specify the executor/optimizer for Table 3 and reconcile the two sets of numbers.
  3. [§4.1 / Baselines] The paper states that 'baseline results are taken from the corresponding papers, while GRAFT is evaluated on the same public benchmark splits.' This means baselines are not run under the same evaluation harness, answer-parsing rules, prompt formatting, or averaging procedure. Since the paper's headline is 'matched optimizer and executor settings,' the comparison would be much stronger if the baselines were rerun under the identical protocol or if the paper demonstrated protocol equivalence explicitly. As written, a portion of the reported 3.85-point margin could be an artifact of evaluation differences rather than the method.
minor comments (4)
  1. [§5.4 / Table 4] The operator trace in Table 4 is a single anecdotal example; it is useful for intuition but does not quantify how often grafting corrects such failures on QA/knowledge tasks. A brief frequency statement (e.g., among corrected instances, what fraction involved reason-region replacement) would help.
  2. [§4.1 / Implementation Details] The per-task selection of P and the use of code repair patterns is a tuning loop on the validation set. This is acceptable, but it should be listed explicitly among the free parameters when discussing generalization, since the method is otherwise presented as training-free with shared hyperparameters.
  3. [Table 3 / Footnote] MultiArith is measured with the GSM8K-base workflow; it would be clearer to also state whether the QA rows in Table 3 use the BayesFlow Wc* or a different global workflow, since this affects interpretation of the reuse and correlation statistics.
  4. [§5.3 / Figure 3] Figure 3 shows monotonic accuracy improvements when swapping executors, but no variance or number of runs is given; adding error bars or at least a note on run-to-run variability would strengthen the executor-transfer claim.

Circularity Check

0 steps flagged

No circularity: the grafting mechanism is evaluated against held-out benchmark correctness and isolated by counterfactual ablations; low proxy fidelity on QA is a correctness risk, not a circular reduction.

full rationale

The central derivation chain is not circular. The global workflow W_c* is selected offline with ground truth (Eq. 2), and inference-time grafting selects region configurations by maximizing the label-free proxy Q (Eq. 3) subject to a coupling guard (Eq. 6). Crucially, the reported scores are external benchmark correctness (solve rate, pass@1, F1, accuracy), not Q itself. The ablation in Table 2 empirically isolates the contribution of grafting (removing it drops the average by 4.10) and of the proxy (random scores drop it by 3.48), so the claim that grafting and the proxy drive the gains is supported by counterfactual comparison, not by construction. The SESE decomposition is cited to an external program-analysis source, and no load-bearing self-citation or uniqueness theorem is invoked. The paper's own Table 3 shows the proxy-correctness Spearman correlation is weak on QA/knowledge tasks (≤0.17), and the matched component ablation in Table 2 covers only math/code benchmarks. This weakens the causal attribution of the QA gains and is a real correctness risk, but it does not make the derivation circular: Eq. (3) does not define correctness, and no fitted parameter is renamed as a prediction. The per-task selection of P and code-repair patterns on the validation set is standard model selection, not a statistical forcing of the reported test results.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The method rests on several hand-set parameters (epsilon, lambdas, per-task P, per-task code-repair pattern) and on the domain assumption that label-free proxy signals can guide local search. The paper's own analysis (Table 3) admits the proxy is weak for QA/knowledge tasks, which is the strongest caveat in the ledger. No new physical entities are introduced; the memory and coupling guard are algorithmic mechanisms.

free parameters (5)
  • epsilon (guard slack) = 0.02
    Set globally for all tasks (Section 4.1); controls how much boundary-support degradation is tolerated. Chosen by hand, not derived.
  • lambda_tok / lambda_loc = 1e-5
    Weights for token/latency cost in both global objective Eq. (2) and local objective Eq. (3); fixed by hand (Section 4.1).
  • max passes P = per-task P* in {1,2,3}
    Selected per task on validation set using the same offline search protocol (Section 3.4, Section 4.1). Directly affects how many downstream regions are re-evaluated and therefore the final answer.
  • per-task code repair pattern usage = on/off per task
    Selected per task on validation set (Section 4.1). Affects code benchmark results.
  • signal weights alpha_k = fixed constants shared across tasks
    Weights in Eq. (4) are fixed but not enumerated; their values affect which candidate is selected in Eq. (3).
axioms (4)
  • domain assumption Label-free proxy Q correlates with answer correctness enough to guide selection (Eq. (3)-(4)).
    Section 5.1 Table 3 measures this and finds correlation <= 0.17 on QA/knowledge tasks, so the assumption is explicitly weakened by the paper's own data on a large part of the claimed task range.
  • domain assumption SESE region decomposition preserves workflow-level consistency and boundaries (Section 3.1).
    The paper cites Johnson, Pearson, and Pingali (1994) and assumes single-entry/single-exit replacement preserves interfaces; no proof is given that the semantic coupling is fully captured by the coupling guard.
  • ad hoc to paper The offline global workflow Wc* is a good prior for the input distribution (Eq. (2)).
    Wc* is selected on the same validation set used to tune P; the paper's task-level optimum claim is assumed rather than proven, and the baseline comparison inherits this validation-split selection.
  • domain assumption Lexical groundedness (Eq. (5)) is a valid signal for evidence-supported QA.
    Groundedness is defined as lexical content-word coverage; no semantic entailment is used. The paper's own Table 3 shows poor proxy correlation on QA tasks, so this assumption is partially contradicted by the data.
invented entities (2)
  • Configuration memory (winning configuration library) independent evidence
    purpose: Caches accepted configurations keyed by input-type signature to avoid repeated local search.
    Its reuse rate and search-cost reduction are measured in Table 3 and Figure 2; the entity is an algorithmic structure with directly observable effects, though the implementation details (signature definition) are not fully specified.
  • Coupling guard (boundary-support score Coh) independent evidence
    purpose: Rejects region replacements that degrade boundary evidence support by more than epsilon.
    The ablation removing it shows a 0.70 average drop (Table 2), providing an empirical handle. It is not a physical entity but a new mechanism introduced by the paper.

pith-pipeline@v1.3.0-daily-deepseek · 13013 in / 8626 out tokens · 56442 ms · 2026-08-04T08:48:03.622255+00:00 · methodology

0 comments
read the original abstract

Recent advances in agentic workflow optimization automate workflow design through task-specific workflow search or input-conditioned architecture selection. However, they determine the workflow before execution and cannot adapt failed workflow regions using execution-time label-free quality signals. Naively enabling such inference-time adaptation through whole-workflow re-optimization would be computationally prohibitive. To tackle this challenge, we introduce GRAFT, which preserves a globally optimized workflow while locally replacing only selected regions for each input. Without parameter training, GRAFT evaluates region-level alternatives using label-free execution-quality signals and accepts only replacements that improve local quality while preserving workflow-level consistency, thereby enabling instance-wise adaptation without whole-workflow re-optimization. GRAFT applies without modification across a range of tasks spanning mathematical reasoning, code generation, and multi-hop and knowledge-intensive question answering. Under matched optimizer and executor settings, it improves over the strongest prior workflow-optimization method, MaAS, by 3.85 points on average. Replacing only the executor with a stronger model yields further gains without re-optimizing the global workflow. This suggests that an optimized workflow is not merely a static optimization artifact, but an adaptable execution policy that can evolve with inference-time feedback and stronger executors.

Figures

Figures reproduced from arXiv: 2608.02353 by Cheoneum Park, Donghyeok Koh, Gyuwan Kim, Haneol Jang, JinYeong Bak, Seung-Hoon Na, Tao Yang.

Figure 1
Figure 1. Figure 1: Overview of GRAFT. Left: the offline-searched global workflow [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Per-query search cost (left, relative to the first [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Per-dataset cost-accuracy when the executor is [PITH_FULL_IMAGE:figures/full_fig_p007_3.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

21 extracted references · 17 linked inside Pith

  1. [3]

    Chen, W.; Su, Y.; Zuo, J.; Yang, C.; Yuan, C.; Chan, C.-M.; Yu,H.;Lu,Y.;Hung,Y.-H.;Qian,C.;Qin,Y.;Cong,X.;Xie, R.; Liu, Z.; Sun, M.; and Zhou, J

    Evaluating Large Language Models Trained on Code.arXiv preprint arXiv:2107.03374. Chen, W.; Su, Y.; Zuo, J.; Yang, C.; Yuan, C.; Chan, C.-M.; Yu,H.;Lu,Y.;Hung,Y.-H.;Qian,C.;Qin,Y.;Cong,X.;Xie, R.; Liu, Z.; Sun, M.; and Zhou, J

    Pith/arXiv arXiv

  2. [5]

    Du,Y.;Li,S.;Torralba,A.;Tenenbaum,J.B.;andMordatch, I

    Train- ing Verifiers to Solve Math Word Problems.arXiv preprint arXiv:2110.14168. Du,Y.;Li,S.;Torralba,A.;Tenenbaum,J.B.;andMordatch, I

    Pith/arXiv arXiv

  3. [7]

    arXiv:2308.00352

    MetaGPT: MetaProgrammingforAMulti-AgentCollaborativeFrame- work. arXiv:2308.00352. Hu, S.; Lu, C.; and Clune, J

    Pith/arXiv arXiv

  4. [10]

    arXiv:2303.17760

    CAMEL: Communicative Agents for "Mind" Exploration of Large Language Model Society. arXiv:2303.17760. Li,Y.;Li,L.;Wu,Z.;Liao,Q.;Hao,J.;Shao,K.;Xu,F.;and Li, Y

    Pith/arXiv arXiv

  5. [11]

    arXiv:2506.06017

    AgentSwift: Efficient LLM Agent Design via Value-guided Hierarchical Search. arXiv:2506.06017. Liu, B.; Jiang, Y.; Zhang, X.; Liu, Q.; Zhang, S.; Biswas, J.; and Stone, P

    arXiv

  6. [12]

    arXiv:2304.11477

    LLM+P: Empowering Large Language Models with Optimal Planning Proficiency. arXiv:2304.11477. Liu, Y.; et al

    Pith/arXiv arXiv

  7. [13]

    arXiv:2310.02170

    A Dynamic LLM-Powered Agent Network for Task-Oriented Agent Collaboration. arXiv:2310.02170. Madaan, A.; Tandon, N.; Gupta, P.; et al

    Pith/arXiv arXiv

  8. [14]

    InAdvances in Neural Information Processing Systems (NeurIPS)

    Self-Refine: Iterative Refinement with Self-Feedback. InAdvances in Neural Information Processing Systems (NeurIPS). Minaee,S.;Mikolov,T.;Nikzad,N.;Chenaghlu,M.;Socher, R.;Amatriain,X.;andGao,J.2025.LargeLanguageModels: A Survey. arXiv:2402.06196. Niu, B.; Song, Y.; Lian, K.; Shen, Y.; Yao, Y.; Zhang, K.; and Liu, T

    Pith/arXiv arXiv 2025

  9. [15]

    arXiv:2501.07834

    Flow: Modularized Agentic Workflow Automation. arXiv:2501.07834. Rein, D.; Hou, B. L.; Stickland, A. C.; et al

    Pith/arXiv arXiv

  10. [16]

    arXiv:2410.06153

    AgentSquare: Automatic LLM Agent Search in Modular Design Space. arXiv:2410.06153. Shinn, N.; Cassano, F.; Gopinath, A.; Narasimhan, K.; and Yao, S

    Pith/arXiv arXiv

  11. [17]

    InInternational Conference on Learning Representations (ICLR)

    Self- Consistency Improves Chain of Thought Reasoning in Lan- guage Models. InInternational Conference on Learning Representations (ICLR). Wang,X.;Yang,Z.;Hong,Z.;Zhang,S.;andWei,W.2026. FusionFlow: Enabling Deep Structural Exploration for Au- tomatedAgenticWorkflowGeneration. InLiakata,M.;Mor- eira, V. P.; Zhang, J.; and Jurgens, D., eds.,Proceedings of ...

    arXiv 2026

  12. [19]

    arXiv:2309.07864

    The Rise and Potential of Large Language Model Based Agents: A Survey. arXiv:2309.07864. Yang, Z.; Qi, P.; Zhang, S.; et al

    Pith/arXiv arXiv

  13. [21]

    arXiv:2303.18223

    A Survey of Large Language Models. arXiv:2303.18223. Zhuge,M.;Wang,W.;Kirsch,L.;Faccio,F.;Khizbullin,D.; and Schmidhuber, J

    Pith/arXiv arXiv

  14. [1994]

    In Sarkar, V.; Ryder, B

    The Pro- gram Structure Tree: Computing Control Regions in Linear Time. In Sarkar, V.; Ryder, B. G.; and Soffa, M. L., eds., Proceedings of the ACM SIGPLAN’94 Conference on Pro- gramming Language Design and Implementation (PLDI), Orlando, Florida, USA, June 20-24, 1994, 171–185. ACM. Lewis, P.; Perez, E.; Piktus, A.; et al

    1994

  15. [2012]

    Cemri,M.;Pan,M.Z.;Yang,S.;etal.2025

    A Survey of Monte Carlo Tree Search Methods.IEEE Transactions on Compu- tational Intelligence and AI in Games, 4(1): 1–43. Cemri,M.;Pan,M.Z.;Yang,S.;etal.2025. WhyDoMulti- AgentLLMSystemsFail?arXivpreprintarXiv:2503.13657. Chen, M.; Tworek, J.; Jun, H.; et al

    Pith/arXiv arXiv 2025

  16. [2021]

    Browne, C

    Program Synthesis with Large Language Models.arXiv preprint arXiv:2108.07732. Browne, C. B.; Powley, E.; Whitehouse, D.; Lucas, S. M.; Cowling, P. I.; Rohlfshagen, P.; Tavener, S.; Perez, D.; Samothrakis, S.; and Colton, S

    Pith/arXiv arXiv

  17. [2022]

    InAdvancesinNeuralInformationProcessingSys- tems (NeurIPS)

    Chain- of-Thought Prompting Elicits Reasoning in Large Language Models. InAdvancesinNeuralInformationProcessingSys- tems (NeurIPS). Wu,Q.;Bansal,G.;Zhang,J.;etal.2023.AutoGen:Enabling Next-GenLLMApplicationsviaMulti-AgentConversation. arXiv preprint arXiv:2308.08155. Xi,Z.;Chen,W.;Guo,X.;He,W.;Ding,Y.;Hong,B.;Zhang, M.; Wang, J.; Jin, S.; Zhou, E.; Zheng,...

    Pith/arXiv arXiv 2023

  18. [2023]

    arXiv:2308.10848

    AgentVerse: Facil- itating Multi-Agent Collaboration and Exploring Emergent Behaviors. arXiv:2308.10848. Cobbe, K.; Kosaraju, V.; Bavarian, M.; et al

    Pith/arXiv arXiv

  19. [2024]

    JMLR.org

    Improving factuality and reasoning in language modelsthroughmultiagentdebate.InProceedingsofthe41st International Conference on Machine Learning, ICML’24. JMLR.org. Dua,D.;Wang,Y.;Dasigi,P.;etal.2019. DROP:AReading Comprehension Benchmark Requiring Discrete Reasoning Over Paragraphs. InNorth American Chapter of the Associ- ation for Computational Linguist...

    2019

  20. [2025]

    arXiv:2408.08435

    Automated Design of Agentic Systems. arXiv:2408.08435. Johnson, R.; Pearson, D.; and Pingali, K

    Pith/arXiv arXiv

  21. [2026]

    BayesFlow: A Probability Inference Framework for Meta-Agent Assisted Workflow Generation. In Demberg, V.; Inui, K.; and Marquez, L., eds.,Findings of the Association for Computational Linguistics: EACL 2026, 3151–3179.Rabat,Morocco:AssociationforComputational Linguistics. ISBN 979-8-89176-386-9. Zhang, G.; Niu, L.; Fang, J.; Wang, K.; Bai, L.; and Wang, X...

    Pith/arXiv arXiv 2026