Pith. sign in

REVIEW 4 major objections 3 minor 47 references

ReasFlow: Assisting Reasoning-Centric Scientific Discovery in Applied Mathematics via a Knowledge-Based Multi-Agent System

T0 review · 4 major / 3 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read ReasFlow is an autonomous multi-agent system that claims to carry out end-to-end reasoning-centric scientific discovery in applied mathematics, producing five research papers with new theoretical convergence results that in several cases su

desk verdict Solid engineering wrapped in an unsupported claim: the system is a real contribution, but the paper does not establish that it produced correct, human-surpassing theorems. read the letter →

arxiv 2607.14178 v2 pith:YXEMXN6K submitted 2026-07-15 cs.AI cs.MA

Yutong He , Daibo Li , Guohong Li , Jiahe Geng , Zhengyang Huang , Can Ren , Zekun Zhang , Yifan Liu
show 10 more authors
Shuchen Zhu Hengrui Zhang Boao Kong Ming Sun Shu Li Chenyi Li Jiang Hu Kun Yuan Zaiwen Wen Pingwen Zhang
This is my paper
classification cs.AIcs.MA
keywords autonomousscientificdiscoverymulti-agentsystemsappliedmathematicstheoremprovingknowledgecardsLLM-as-judgeverificationloopalgorithmdesign
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper aims to show that autonomous AI systems can move beyond empirical, benchmark-driven research into theory-driven discovery in applied mathematics. It presents ReasFlow, a multi-agent pipeline that designs algorithms, proves convergence theorems, runs experiments, and writes manuscripts from minimal prompts. The central claim is that with an internal verification loop and knowledge-card retrieval, the system can produce rigorous theoretical advances—including a decentralized stochastic optimization method claimed to match a lower complexity bound for the first time, and convergence rates that strictly improve on human-designed federated subspace methods. If true, this would extend AI-assisted discovery to fields where progress is defined by proofs rather than test scores.

What carries the argument

The central machinery is a pair of mechanisms: (i) knowledge cards—compact text files encoding proof techniques, heuristics, and cautionary patterns extracted from reference papers via automatically refined prompts—that give the agents domain-specific procedural know-how; and (ii) an internal verification loop in which a dedicated verifier agent audits each generated lemma and theorem for logical gaps, hidden assumptions, and overly loose bounds, rejecting proofs under a zero-tolerance policy until they pass. A MetaAgent orchestrates specialized sub-agents (survey, algorithm, prover, experiment, introduction, writing) around these mechanisms.

What would settle it

Have the five companion papers' central theorems formally checked in a proof assistant such as Lean, Coq, or Isabelle; finding one unprovable statement or a counterexample to a claimed convergence rate would falsify the claim of rigorous new theory. Alternatively, a panel of independent human experts could review the companion proofs for a logical gap.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that ReasFlow—an LLM-based multi-agent architecture with a human principal investigator and an AI research assistant—can autonomously generate complete, publishable research papers in applied mathematics. The system unifies literature survey, algorithm design, theorem proving, experimentation, and manuscript writing. Five case studies are reported, spanning federated learning, decentralized optimization, manifold optimization, and LLM training, with claims of first-time optimal complexity guarantees and improved convergence rates over human-designed methods. The two key mechanisms are an internal verification loop that audits logical coherence and tightness

Load-bearing premise

The entire claim that ReasFlow produces rigorous theoretical advances depends on LLM-based review being a trustworthy proxy for mathematical correctness; if those judges reward fluent but wrong proofs, the central claim collapses.

Editorial extensions

If this is right

  • If ReasFlow's claims hold, LLM-based agents can produce rigorous theoretical contributions in applied mathematics, not just empirical results.
  • The five case studies, if correct, would establish new state-of-the-art convergence guarantees for decentralized stochastic optimization, federated subspace learning, decentralized Muon optimizers, and retraction-free Stiefel manifold methods.
  • The knowledge-card mechanism offers a way to inject implicit procedural knowledge (proof heuristics, cautionary patterns) into LLMs without manual curation.
  • The internal verification loop demonstrates a scalable approach to auditing natural-language proofs for common reasoning errors, though it does not guarantee formal correctness.

Reading between the lines

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

  • The strongest load-bearing assumption is that LLM-as-judge scores reflect genuine mathematical correctness; since the companion papers' proofs are not reproduced or formally verified here, the headline results stand or fall on the reliability of those judges.
  • A natural testable extension is to run the five generated theorems through an interactive theorem prover (Lean, Coq, or Isabelle) or to have human experts independently verify the companion proofs.
  • If the approach generalizes, the same architecture—verification loop plus knowledge cards—could extend to other theory-driven fields beyond applied mathematics, such as theoretical physics or theoretical computer science.
  • The paper's evaluation protocol, which relies on LLM reviewers, could itself be a confound: the system may be optimized to produce outputs that look correct to LLM reviewers rather than being mathematically correct.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 3 minor

Summary. The paper introduces ReasFlow, a multi-agent LLM system intended to automate reasoning-centric research in applied mathematics. The system combines literature survey, algorithm design, theorem proving, experimentation, and manuscript writing, augmented by a knowledge-card mechanism and an automatic prompt-refinement loop. The authors claim that ReasFlow autonomously produced five complete research papers with rigorous theoretical and empirical content, including cases said to match or surpass the best human-derived convergence guarantees in decentralized optimization and federated learning. Evaluation is carried out primarily with LLM-based reviewers, supplemented by component-level benchmarks. The manuscript is detailed and openly released, and it explicitly acknowledges several limitations; however, the load-bearing evidence for the headline scientific claims—the theorem statements and proofs—is not present in the manuscript.

Significance. If the central claims were established, this would be a significant milestone for AI-assisted theoretical research. The system architecture is described in unusual detail, the repository is public, and several component evaluations (e.g., first-run script success in Table 3, LaTeX compilation checks in Table 7, citation verification in §B.2) are more objective than pure LLM scoring. These are genuine strengths. However, the main scientific assertions—that ReasFlow generated new theorems, matched a lower complexity bound, and strictly improved on human-designed rates—are not verifiable from the manuscript. No theorem is stated or proved here; the proofs are deferred to self-published companion papers; and the reported verification is LLM-based, with large inter-reviewer disagreement and explicit concessions that the verification loop does not guarantee formal correctness. The significance of the contribution therefore remains unestablished in this version.

major comments (4)
  1. [§1, §5.1 (Cases 2–4)] The headline claims are that ReasFlow 'can autonomously produce substantial theoretical advances' and that the generated papers include 'the first decentralized stochastic optimization method to attain this lower complexity bound' (Case 4), rates that 'strictly improve upon' human-designed federated subspace methods (Case 2), and a topology-independent rate for decentralized Muon (Case 3). None of the underlying theorems is stated or proved in this manuscript. The case summaries defer to companion manuscripts [16, 42, 59, 64, 24] that are co-authored by overlapping authors and hosted on the team's blog. A reader cannot verify the claimed lower-bound match, the strict improvement over [60], or the convergence rates in Cases 2–3. For an applied-mathematics claim, this is not a presentation issue; the central evidence is absent from the submitted work.
  2. [§5.2, §5.3.3, Tables 1, 4, 19; Algorithms 1–2] The reported verification is circular and unstable. The 90.7/100 proof score and the perfect Gemini-3-Pro scores come from ProofReviewAgent/PaperReviewAgent, LLM judges drawn from the same model families (GPT-5.x, Gemini-3-Pro) that generated the content. Algorithms 1–2 refine prompts until an LLM comparator Ω judges the card-assisted proof 'at least as good' as a baseline; this optimizes for LLM approval, not mathematical truth. Cross-backend disagreement is large: ReasFlow's own paper score ranges from 65.6 to 87.6 in Table 1, and its 'logical deduction and correctness' sub-score in Table 19 is 46/50 under GPT-5.1, 16/50 under GPT-5.4, and 50/50 under Gemini-3-Pro. Such instability cannot support claims of rigorous proof or of matching a lower complexity bound.
  3. [§4.3, §6] The manuscript itself concedes that natural-language proofs 'may still contain errors, and human evaluation remains necessary to ensure correctness' and that the verification loop 'does not guarantee formal correctness'; integration with Lean/Coq/Isabelle is listed as future work. The LemmaVerifierAgent's 'zero-tolerance' policy makes it a strict critic, but it is still an LLM with no access to a proof checker. Consequently, the abstract's characterization of 'rigorous theoretical and empirical content' and §5.1's 'rigorous convergence guarantees' are not supported by the reported verification pipeline. The internal verification loop can filter some flaw patterns, but it cannot certify the correctness of the claimed theorems.
  4. [§5.2, §5.3, Appendix B] The controlled full-paper benchmark is run on a single task ([16]) despite the paper's global claim that ReasFlow consistently achieves the highest scores across the five case studies. The other four cases are presented without the same controlled comparison, and Mode I cases involve direct human interaction with sub-agents rather than the end-to-end MetaAgent workflow. In addition, several evaluation thresholds and weights are not justified: the SurveyAgent acceptance threshold, the introduction's 0.60/0.75 refinement thresholds, the nine-dimension overall-score weights, and the 'medium' reasoning effort for ReasFlow in §5.3.3. No sensitivity analysis is reported, so the large margins in Tables 1 and 4 could be materially affected by these choices.
minor comments (3)
  1. [§1, Table 1] The introduction states that ReasFlow 'consistently received perfect scores from Gemini-3-Pro.' In the full-paper evaluation of Table 1, ReasFlow's Gemini-3-Pro score is 87.6, not 100. If the claim refers only to the ProofReviewAgent scores in Table 4, it should be worded accordingly.
  2. [Figure 1 caption] Typo: 'verfication' should be 'verification'.
  3. [References [16, 24, 42, 59, 64]] The companion papers are hosted as 'unpolished ReasFlow-generated manuscripts' on the team's blog. If these are to be used as evidence for the central claims, they should be archived in a stable, independent venue rather than a project blog, and their provenance should be clarified.

Circularity Check

3 steps flagged · score 7.0 of 10

Headline theoretical-advance claim is validated by a self-referential chain: LLM-judge scores from the generating model families, knowledge-card 'equivalence' defined by the LLM comparator's stopping rule, and five case-study theorems deferred to same-author companion papers on the team's blog.

  1. self definitional [Sec. 3.2, Algorithms 1–2 and Remark]
    "Ω : P × P → {true,false}: compares whether pk ∈ P is at least as good as p⋆ ∈ P. ... if Ω(p k, p⋆) = false then Update prompts k+1 = Υ(sk, R, Ck, pk, p⋆) ... else break. ... Remark. We empirically observe that after refinement of the skill prompts, the ProverAgent equipped with built-in knowledge cards achieves performance comparable to the agent with direct access to original reference papers."

    The claimed 'empirical observation' of equivalence is exactly the loop's termination condition: Algorithm 1/2 runs until the LLM comparator Ω approves p_k as 'at least as good' as the baseline p⋆. Thus 'knowledge-card-augmented proof generation is comparable to direct reference/expert guidance' is true by construction with respect to Ω, not by an independent measurement. Since §4.3 and §6 also concede that natural-language proofs 'may still contain errors' and the verification loop 'does not guarantee formal correctness', the comparator's approval is a same-model subjective judgment, not a disinterested check.

  2. self citation load bearing [Sec. 1 (Contributions); Sec. 5.1 Cases 1–5; refs [16,42,59,64,24]]
    "To validate ReasFlow, we used it to develop five theoretically significant algorithms and turn them into complete research papers ... To the best of our knowledge, [42] is the first decentralized stochastic optimization method to attain this lower complexity bound. ... The unpolished ReasFlow-generated manuscript, prior to human-expert polishing, is available at https://blog.reaslab.io/papers/[blog]MC-ADSGD.pdf."

    The paper's central evidence that ReasFlow can 'autonomously produce substantial theoretical advances' consists of five case-study results that are not stated or proved in this manuscript. Each case defers to a companion paper by overlapping ReasFlow authors, hosted on the team's blog (e.g., [42] by Sun & Yuan; [64] by Zhu, Huang, Xu, Li; [16] by He, Huang, Geng). The claims 'first to attain this lower complexity bound' and 'strictly improves on human-designed methods' therefore rest on self-citations whose correctness is not independently verified in this paper; the validation loop reduces to ReasFlow's output citing itself.

1 more flagged steps
  1. other [Sec. 5.2–5.3, Tables 1 and 4; Sec. 6]
    "A specialized LLM-based reviewer evaluated all outputs according to a carefully curated rubric. ... ReasFlow with GPT-5.4 achieves the highest average score of 90.7 ... when evaluated by Gemini-3-Pro, ReasFlow attains a perfect score of 100 across all three LLM backbones. ... although the internal verification loop detects many common reasoning errors, it does not guarantee formal correctness."

    Generation, internal verification, and final evaluation all use LLMs from the same model families (GPT-5.x, Gemini-3-Pro). Because the prompt/knowledge-card refinement loops in Sec. 3.2 are explicitly driven by an LLM comparator (Ω) rather than a formal proof checker, the reported 'prover score' and 'perfect scores' measure the LLM's approval of its own output, not mathematical correctness. The paper itself states that formal verification (Lean/Coq/Isabelle) is future work; using these self-generated scores as evidence for 'substantial theoretical advances' closes the evaluative circle.

full rationale

The headline assertion — that ReasFlow 'can autonomously produce substantial theoretical advances and, in several cases, surpass the best results previously achieved by human experts' — is supported by three interlocking moves that are circular. First, the knowledge-card ablation's conclusion ('comparable to direct access to original reference papers') is a definitional tautology: it is the stopping condition of Algorithms 1–2, where Ω is defined as comparing whether p_k is 'at least as good' as p⋆. Second, the five case-study theorems are neither stated nor proved here; each is deferred to companion papers by the same research group, hosted on the team's blog, making the claimed advances a self-citation chain. Third, the only in-manuscript verification of these theoretical results is LLM-as-judge scoring from the same model families that generated and internally verified the proofs, and the paper explicitly concedes that its loop 'does not guarantee formal correctness.' These moves are load-bearing: without them, no evidence remains for the central claim. Some non-circular components do exist — the WritingAgent's programmatic compile/template checks, the AlgorithmAgent's first-run code-success measurement, and the SurveyAgent's citation verification against a fixed FAISS database. But those do not certify mathematical correctness of the five advances. Score 7 reflects strong partial circularity: the system engineering may be real, but the claimed theoretical results are not independently established within this manuscript.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

This is a systems paper, so the ledger entries are design choices and evaluation assumptions rather than fit constants. No scientific constant is fitted, but the headline scores depend on hand-chosen rubric weights, refinement thresholds, reasoning-effort settings, and the unstated premise that LLM review is a valid correctness gate — a premise the authors themselves qualify in §4.3 and §6.

free parameters (5)
  • Introduction evaluation thresholds = 0.60 (auto_refine) / 0.75 (suggest_refine)
    Hand-chosen thresholds in the IntroductionAgent evaluation loop (§4.5) that decide when the agent stops refining; the reported introduction quality gains depend on them.
  • Overall-score rubric weights = 0.25·SA + 0.15·SB + 0.2·SC + 0.1·SD + 0.1·SE + 0.05·SF + 0.05·SG + 0.05·SH + 0.05·SI
    Hand-assigned weights in Appendix B.1.2 that define the Table 1 ranking; changing the weights can change the winner ordering, yet the 'best system' claim is stated categorically.
  • Reasoning effort level = medium
    All ReasFlow agents run at 'medium reasoning level' in §5.3.3 to reduce cost; all prover scores are reported for this single configuration.
  • CardGenerator/CardRetriever skill prompts = refined by Algorithms 1–2 until LLM comparator Ω accepts
    The prompts are optimized until an LLM comparator says the card-assisted proof is 'at least as good' as the baseline; the stopping criterion is an LLM judgment, not an external metric, so the refinement target is the LLM's own approval.
  • SurveyAgent acceptance threshold = unspecified
    The ReviewTool acceptance threshold in §4.1 that determines iterative revision is not stated; it affects the SurveyAgent quality scores in Table 2.
assumptions (5)
  • domain assumption LLM reviewer scores are a valid proxy for mathematical correctness and paper quality
    The entire evaluation (§5, Appendix B) treats PaperReviewAgent/ProofReviewAgent scores as ground truth; the paper itself notes 'human evaluation remains necessary' (§4.3) and the loop 'does not guarantee formal correctness' (§6).
  • domain assumption The 339-paper FAISS knowledge database is a reliable ground truth for citation verification
    SurveyAgent Content Accuracy (Appendix B.2) is verified against a database the authors constructed; the database itself is not independently audited.
  • domain assumption The benchmark comparison is fair (identical prompts, equivalent resources per system)
    §5.2 says systems received an identical prompt, but the baseline set mixes frontier web assistants, open systems on GPT-5.4, and an open 12B model (CycleResearcher-ML-12B); resource and model-scale differences are not controlled.
  • ad hoc to paper LLM comparator Ω (Algorithms 1–2) gives a reliable ordering of proof quality
    The self-improvement loops optimize until Ω accepts the proof; if Ω is satisfied by fluent but wrong proofs, the refinement criterion is vacuous.
  • standard math Standard optimization-theory background (smoothness, strong convexity, consensus-error analysis)
    The case-study claims (rates, lower bounds) are stated within the conventional framework of the cited literature (e.g., [4], [20], [39], [42]); background results are taken as given.
invented entities (1)
  • Knowledge cards
    purpose: Encapsulate procedural proof heuristics and declarative facts for retrieval by Prover/Algorithm/Experiment agents (§3.2)
    The benefit of knowledge cards is measured only within the system's own LLM-judged pipeline and against the authors' own reference-paper extraction; no falsifiable handle outside the paper's framework is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ReasFlow: Assisting Reasoning-Centric Scientific Discovery in Applied Mathematics via a Knowledge-Based Multi-Agent System." pith.science (2026). https://pith.science/paper/YXEMXN6K

@misc{pith2026260714178,
  author       = {Pith},
  title        = {Pith review of: ReasFlow: Assisting Reasoning-Centric Scientific Discovery in Applied Mathematics via a Knowledge-Based Multi-Agent System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YXEMXN6K}},
  note         = {Machine review of arXiv:2607.14178}
}
read the original abstract

Recent advances in Large Language Models have fueled autonomous AI agents capable of tackling complex scientific tasks, yet existing automated research systems remain predominantly focused on empirically driven domains with quantitative benchmarks, leaving theory-driven discovery, particularly in mathematically grounded disciplines requiring rigorous proofs and synthesis of domain knowledge, largely underexplored. Key challenges include the difficulty of verifying theoretical reasoning at scale, insufficient reasoning ability for autonomous frontier exploration, and a scarcity of procedural heuristics in the literature. We introduce ReasFlow, an end-to-end autonomous agent system for reasoning-centric scientific discovery that operationalizes a collaborative paradigm where the human expert acts as Principal Investigator while the agent executes rigorous derivations as a capable graduate student. ReasFlow incorporates (i) a robust internal verification loop that audits logical coherence and corrects fundamental errors prior to human inspection, and (ii) an automated knowledge retrieval and self-improvement mechanism that proactively surfaces both declarative facts and overlooked procedural heuristics, substantially reducing expert intervention. The system unifies literature synthesis, algorithm design, theorem proving, experimentation, and manuscript preparation in a single system. Deployed to autonomously generate five complete research papers with rigorous theoretical and empirical content from minimal prompts, ReasFlow consistently achieves the highest evaluation scores among state-of-the-art open-access baselines under a curated LLM-based review rubric. ReasFlow is publicly accessible via the ReasLab platform, providing a collaborative workspace for AI-assisted theoretical research. Github repo: https://github.com/reaslab/ReasFlow.git.

Figures

Figures reproduced from arXiv: 2607.14178 by the authors.

Figure 1
Figure 1. Convergence and final precision on non-IID MNIST ( [PITH_FULL_IMAGE:figures/full_fig_p089_1.png] view at source ↗
Figure 2
Figure 2. Sensitivity to the Dirichlet concentration parameter [PITH_FULL_IMAGE:figures/full_fig_p090_2.png] view at source ↗
Figure 3
Figure 3. Convergence of FedSLoP for different projection ranks r. 5.4 Projection-Rank Ablation The projection rank r is the central design parameter of FEDSLOP [PITH_FULL_IMAGE:figures/full_fig_p091_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Convergence of FedSLoP for different numbers of clients. 5.6 Summary Across all experiments, FEDSLOP achieves a strong balance between accuracy and com￾munication efficiency. On non-IID MNIST it remains very close to the full-parameter FEDAVGWITHMOMENTUM baseline, subs…
Figure 1
Figure 1. Figure 1: Relative error versus communication rounds for [PITH_FULL_IMAGE:figures/full_fig_p134_1.png]
Figure 2
Figure 2. Figure 2: Effect of the subspace dimension on convergence at high heterogeneity ( [PITH_FULL_IMAGE:figures/full_fig_p135_2.png]
Figure 3
Figure 3. Figure 3: Convergence behaviour at medium heterogeneity ( [PITH_FULL_IMAGE:figures/full_fig_p137_3.png]
Figure 4
Figure 4. Figure 4: Test accuracy versus training epoch on CIFAR-100 with ResNet-110. Full [PITH_FULL_IMAGE:figures/full_fig_p138_4.png]
Figure 1
Figure 1. Figure 1: Counterexample for linear speedup under transverse noise ( [PITH_FULL_IMAGE:figures/full_fig_p185_1.png]
Figure 2
Figure 2. Figure 2: CIFAR-100 performance of the three decentralized Muon variants on a 20- [PITH_FULL_IMAGE:figures/full_fig_p188_2.png]
Figure 3
Figure 3. Figure 3: Validation perplexity of decentralized Muon variants on GPT-2 / Wikitext [PITH_FULL_IMAGE:figures/full_fig_p189_3.png]
Figure 1
Figure 1. Figure 1: Synthetic decentralized PCA: robustness of RF-EXTRA with respect to graph topology and internal [PITH_FULL_IMAGE:figures/full_fig_p238_1.png]
Figure 2
Figure 2. Figure 2: Synthetic decentralized PCA on ER(0.6) versus communication quantities. Each method uses its best step size selected from {1,2,4,6,8}×{10−5 ,10−4 ,10−3 ,10−2}. Under this matched search space, RF-EXTRA, DESTINY, DPRGT, and REXTRA all select βˆ =0.08, while DPRGD select…
Figure 3
Figure 3. Figure 3: Decentralized PCA on the MNIST dataset versus communication quantities. RF-EXTRA, DESTINY, [PITH_FULL_IMAGE:figures/full_fig_p240_3.png]
Figure 4
Figure 4. Figure 4: Decentralized LRMC on the ring graph versus communication quantities. Only the stationarity [PITH_FULL_IMAGE:figures/full_fig_p241_4.png]
Figure 5
Figure 5. Figure 5: Decentralized LRMC on the ring graph versus communication quantities for representative RF-EXTRA [PITH_FULL_IMAGE:figures/full_fig_p241_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 5 linked inside Pith

  1. [16]

    Prox-pda: The proximal primal- dual algorithm for fast distributed nonconvex optimization and learning over networks

    Mingyi Hong, Davood Hajinezhad, and Ming-Min Zhao. Prox-pda: The proximal primal- dual algorithm for fast distributed nonconvex optimization and learning over networks. In International Conference on Machine Learning, pages 1529–1538. PMLR, 2017. 26 ReasFlow Paper Copilot / CSIAM Trans. Appl. Math.,1(2026), pp. 1-27

    2017

  2. [1]

    Dual averaging for distributed opti- mization: Convergence analysis and network scaling.IEEE Transactions on Automatic Control, 2011

    John Duchi, Alekh Agarwal, and Martin Wainwright. Dual averaging for distributed opti- mization: Convergence analysis and network scaling.IEEE Transactions on Automatic Control, 2011

    2011

  3. [2]

    On the convergence of decentralized gradient descent

    Kun Yuan, Qing Ling, and Wotao Yin. On the convergence of decentralized gradient descent. SIAM Journal on Optimization, 26(3):1835–1854, 2016

    2016

  4. [3]

    Alghunaim and Kun Yuan

    Sulaiman A. Alghunaim and Kun Yuan. A unified and refined convergence analysis for non-convex decentralized learning.IEEE Transactions on Signal Processing, 2022

    2022

  5. [4]

    Extra: An exact first-order algorithm for decentralized consensus optimization.SIAM Journal on Optimization, 25(2):944–966, 2015

    Wei Shi, Qing Ling, Gang Wu, and Wotao Yin. Extra: An exact first-order algorithm for decentralized consensus optimization.SIAM Journal on Optimization, 25(2):944–966, 2015

    2015

  6. [5]

    Kun Yuan, Bicheng Ying, Xiaochuan Zhao, and Ali H. Sayed. Exact diffusion for distributed optimization and learning — part i: Algorithm development.IEEE Transactions on Signal Processing, 2018

    2018

  7. [6]

    Orthogonal weight normalization: Solution to optimization over multiple dependent stiefel manifolds in deep neural networks

    Lei Huang, Xianglong Liu, Bo Lang, Adams Yu, Yongliang Wang, and Bo Li. Orthogonal weight normalization: Solution to optimization over multiple dependent stiefel manifolds in deep neural networks. InProceedings of the AAAI Conference on Artificial Intelligence, 2018

    2018

  8. [7]

    Riemannian approach to batch normalization

    Minhyung Cho and Jaehyung Lee. Riemannian approach to batch normalization. InAd- vances in Neural Information Processing Systems, 2017

    2017

Show all 47 references
  1. [8]

    Riemannian preconditioned lora for fine-tuning foundation models.arXiv preprint arXiv:2402.02347, 2024

    Furong Zhang and Mert Pilanci. Riemannian preconditioned lora for fine-tuning foundation models.arXiv preprint arXiv:2402.02347, 2024

    2024 arXiv

  2. [9]

    An updated set of basic linear algebra subprograms (blas).ACM Transactions on Mathematical Software, 28(2):135–151, 2002

    L Susan Blackford, Antoine Petitet, Roldan Pozo, Karin Remington, R Clint Whaley, James Demmel, Jack Dongarra, Iain Duff, Sven Hammarling, Greg Henry, et al. An updated set of basic linear algebra subprograms (blas).ACM Transactions on Mathematical Software, 28(2):135–151, 2002

    2002

  3. [10]

    Anastasia Koloskova, Tao Lin, and Sebastian U. Stich. An improved analysis of gradient tracking for decentralized machine learning. InAdvances in Neural Information Processing Systems, 2022

    2022

  4. [11]

    D2: Decentralized training over decentralized data

    Hanlin Tang, Xiangru Lian, Ming Yan, Ce Zhang, and Ji Liu. D2: Decentralized training over decentralized data. InProceedings of the International Conference on Machine Learning, 2018

    2018

  5. [12]

    S-diging: A stochastic gradient tracking algorithm for distributed optimization.IEEE Transactions on Emerging Topics in Computational Intelligence, 2020

    Huaqing Li, Lifeng Zheng, Zheng Wang, Yu Yan, Liping Feng, and Jing Guo. S-diging: A stochastic gradient tracking algorithm for distributed optimization.IEEE Transactions on Emerging Topics in Computational Intelligence, 2020

    2020

  6. [13]

    Distributed stochastic gradient tracking methods.Mathematical Programming, 2021

    Shi Pu and Angelia Nedi ´c. Distributed stochastic gradient tracking methods.Mathematical Programming, 2021

    2021

  7. [14]

    Yue Liu, Tao Lin, Anastasia Koloskova, and Sebastian U. Stich. Decentralized gradient track- ing with local steps.Optimization Methods and Software, 2023

    2023

  8. [15]

    Z. Li, W. Shi, and M. Yan. A decentralized proximal-gradient method with network inde- pendent step-sizes and separated convergence rates.IEEE Transactions on Signal Processing, 67(17):4494–4506, 2019

    2019

  9. [17]

    Alghunaim, and Xinmeng Huang

    Kun Yuan, Sulaiman A. Alghunaim, and Xinmeng Huang. Removing data heterogeneity influence enhances network topology dependence of decentralized sgd.Journal of Machine Learning Research, 2023

    2023

  10. [18]

    Next: In-network nonconvex optimization.IEEE Transactions on Signal and Information Processing over Networks, 2016

    Paolo Di Lorenzo and Gesualdo Scutari. Next: In-network nonconvex optimization.IEEE Transactions on Signal and Information Processing over Networks, 2016

    2016

  11. [19]

    Stich, and Martin Jaggi

    Anastasia Koloskova, Sebastian U. Stich, and Martin Jaggi. Decentralized stochastic op- timization and gossip algorithms with compressed communication. InProceedings of the International Conference on Machine Learning, 2019

    2019

  12. [20]

    Unbiased compression saves communication in distributed optimization: When and how much? InAdvances in Neural Information Pro- cessing Systems, 2023

    Yutong He, Xinmeng Huang, and Kun Yuan. Unbiased compression saves communication in distributed optimization: When and how much? InAdvances in Neural Information Pro- cessing Systems, 2023

    2023

  13. [21]

    Greedy low-rank gradi- ent compression for distributed learning with convergence guarantees.IEEE Transactions on Signal Processing, 2026

    Chuyan Chen, Yutong He, Pengrui Li, Weichen Jia, and Kun Yuan. Greedy low-rank gradi- ent compression for distributed learning with convergence guarantees.IEEE Transactions on Signal Processing, 2026

    2026

  14. [22]

    On biased compression for distributed learning.Journal of Machine Learning Research, 2023

    Aleksandr Beznosikov, Samuel Horv ´ath, Peter Richt ´arik, and Mher Safaryan. On biased compression for distributed learning.Journal of Machine Learning Research, 2023

    2023

  15. [23]

    Error compensated distributed sgd can be accelerated

    Xun Qian, Peter Richt ´arik, and Tong Zhang. Error compensated distributed sgd can be accelerated. InAdvances in Neural Information Processing Systems, 2021

    2021

  16. [24]

    Achieving geometric convergence for dis- tributed optimization over time-varying graphs.SIAM Journal on Optimization, 2016

    Angelia Nedi ´c, Alex Olshevsky, and Wei Shi. Achieving geometric convergence for dis- tributed optimization over time-varying graphs.SIAM Journal on Optimization, 2016

    2016

  17. [25]

    Understanding the influence of digraphs on decentralized optimization: Effective metrics, lower bound, and optimal algo- rithm.SIAM Journal on Optimization, 2025

    Liyuan Liang, Xinmeng Huang, Ran Xin, and Kun Yuan. Understanding the influence of digraphs on decentralized optimization: Effective metrics, lower bound, and optimal algo- rithm.SIAM Journal on Optimization, 2025

    2025

  18. [26]

    Alghunaim

    Sulaiman A. Alghunaim. Local exact-diffusion for decentralized optimization and learning. IEEE Transactions on Automatic Control, 2024

    2024

  19. [27]

    Exponential graph is provably efficient for decentralized deep training

    Bicheng Ying, Kun Yuan, Yiming Chen, Hanbin Hu, Pan Pan, and Wotao Yin. Exponential graph is provably efficient for decentralized deep training. InAdvances in Neural Information Processing Systems, 2021

    2021

  20. [28]

    Decentralized riemannian gradient descent on the stiefel manifold

    Shixiang Chen, Alfredo Garcia, Mingyi Hong, and Shahin Shahrampour. Decentralized riemannian gradient descent on the stiefel manifold. InInternational Conference on Machine Learning, pages 1594–1605. PMLR, 2021

    2021

  21. [29]

    Decentralized projected riemannian gradient method for smooth optimization on compact submanifolds embedded in the euclidean space.Nu- merische Mathematik, 2025

    Kangkang Deng and Jiang Hu. Decentralized projected riemannian gradient method for smooth optimization on compact submanifolds embedded in the euclidean space.Nu- merische Mathematik, 2025

    2025

  22. [30]

    A decentralized proximal gradient tracking algorithm for composite optimization on riemannian manifolds.Journal of Machine Learning Research, 2025

    Lei Wang, Le Bao, and Xin Liu. A decentralized proximal gradient tracking algorithm for composite optimization on riemannian manifolds.Journal of Machine Learning Research, 2025

    2025

  23. [31]

    Riemannian extra: Communication-efficient decentralized optimization over compact submanifolds with data heterogeneity.arXiv preprint arXiv:2505.15537, 2025

    Jiayuan Wu, Zhanwang Deng, Jiang Hu, Weijie Su, and Zaiwen Wen. Riemannian extra: Communication-efficient decentralized optimization over compact submanifolds with data heterogeneity.arXiv preprint arXiv:2505.15537, 2025

    2025 arXiv

  24. [32]

    Decentralized optimization on compact submanifolds by quantized riemannian gradient tracking.IEEE Transactions on Signal Processing, 2025

    Jun Chen, Lina Liu, Tianyi Zhu, Yong Liu, Guang Dai, Yunliang Jiang, and Ivor W Tsang. Decentralized optimization on compact submanifolds by quantized riemannian gradient tracking.IEEE Transactions on Signal Processing, 2025

    2025

  25. [33]

    Improving the communication in decentralized manifold op- timization through single-step consensus and compression.arXiv preprint arXiv:2407.08904, 2024

    Jiang Hu and Kangkang Deng. Improving the communication in decentralized manifold op- timization through single-step consensus and compression.arXiv preprint arXiv:2407.08904, 2024

    2024 arXiv

  26. [34]

    Decentralized projected riemannian stochastic recursive momentum method for nonconvex optimization

    Kangkang Deng and Jiang Hu. Decentralized projected riemannian stochastic recursive momentum method for nonconvex optimization. InProceedings of the AAAI Conference on ReasFlow Paper Copilot / CSIAM Trans. Appl. Math.,1(2026), pp. 1-27 27 Artificial Intelligence, 2025

    2026

  27. [35]

    Decentralized riemannian conjugate gradient method on the stiefel manifold

    Jun Chen, Haishan Ye, Mengmeng Wang, Tianxin Huang, Guang Dai, Ivor W Tsang, and Yong Liu. Decentralized riemannian conjugate gradient method on the stiefel manifold. arXiv preprint arXiv:2308.10547, 2023

    2023 arXiv

  28. [36]

    Decentralized riemannian natural gradient methods with kronecker product approximations.Journal of the Operations Research Society of China, 2025

    Jiang Hu, Kangkang Deng, and Quanzheng Li. Decentralized riemannian natural gradient methods with kronecker product approximations.Journal of the Operations Research Society of China, 2025

    2025

  29. [37]

    On the local linear rate of consensus on the stiefel manifold.IEEE Transactions on Automatic Control, 2023

    Shixiang Chen, Alfredo Garcia, Mingyi Hong, and Shahin Shahrampour. On the local linear rate of consensus on the stiefel manifold.IEEE Transactions on Automatic Control, 2023

    2023

  30. [38]

    Riemannian consensus for manifolds with bounded curvature.IEEE Transactions on Automatic Control, 2012

    Roberto Tron, Bijan Afsari, and Ren ´e Vidal. Riemannian consensus for manifolds with bounded curvature.IEEE Transactions on Automatic Control, 2012

    2012

  31. [39]

    Consensus optimization on manifolds.SIAM Journal on Control and Optimization, 2009

    Alain Sarlette and Rodolphe Sepulchre. Consensus optimization on manifolds.SIAM Journal on Control and Optimization, 2009

    2009

  32. [40]

    Global conver- gence of decentralized retraction-free optimization on the stiefel manifold.arXiv preprint arXiv:2405.11590, 2024

    Yifei Sun, Shixiang Chen, Alfredo Garcia, and Shahin Shahrampour. Global conver- gence of decentralized retraction-free optimization on the stiefel manifold.arXiv preprint arXiv:2405.11590, 2024

    2024 arXiv

  33. [41]

    Decentralized optimization over the stiefel manifold by an approxi- mate augmented lagrangian function.IEEE Transactions on Signal Processing, 2022

    Lei Wang and Xin Liu. Decentralized optimization over the stiefel manifold by an approxi- mate augmented lagrangian function.IEEE Transactions on Signal Processing, 2022

    2022

  34. [42]

    Fast and accurate optimization on the orthogonal manifold without retraction

    Pierre Ablin and Gabriel Peyr ´e. Fast and accurate optimization on the orthogonal manifold without retraction. InProceedings of The 25th International Conference on Artificial Intelligence and Statistics, pages 5636–5657. PMLR, 2022

    2022

  35. [43]

    Parallelizable algorithms for optimization problems with orthogonality constraints.SIAM Journal on Scientific Computing, 41(3):A1949–A1983, 2019

    Bin Gao, Xin Liu, and Ya-xiang Yuan. Parallelizable algorithms for optimization problems with orthogonality constraints.SIAM Journal on Scientific Computing, 41(3):A1949–A1983, 2019

    2019

  36. [44]

    Local linear conver- gence of infeasible optimization with orthogonal constraints.IEEE Control Systems Letters, 2024

    Yifei Sun, Shixiang Chen, Alfredo Garcia, and Shahin Shahrampour. Local linear conver- gence of infeasible optimization with orthogonal constraints.IEEE Control Systems Letters, 2024

    2024

  37. [45]

    Dissolving constraints for riemannian opti- mization.Mathematics of Operations Research, 2024

    Nachuan Xiao, Xin Liu, and Kim-Chuan Toh. Dissolving constraints for riemannian opti- mization.Mathematics of Operations Research, 2024

    2024

  38. [46]

    Horn and Charles R

    Roger A. Horn and Charles R. Johnson.Matrix Analysis. Cambridge University Press, 2012

    2012

  39. [47]

    Convergence analysis of extra in non-convex distributed optimization

    Lei Qin and Ye Pu. Convergence analysis of extra in non-convex distributed optimization. IEEE Control Systems Letters, 2025. 245 Table 23Phase3(WritingQuality)sub-dimensionscores. III1clarity, III2coherence(projectfocus), III3argumentation, III4 evidence use. System III1 III2 ...

    2025

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.