Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

SWIRL turns multi-agent reinforcement learning into a sequence of single-agent RL tasks, updating one agent at a time while freezing the rest.

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 ↗

A multi-agent RL workflow that interleaves single-agent updates, applied to mobile GUI control, achieves SOTA zero-shot performance and a +14.8 MATH500 gain.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Useful empirical recipe for training GUI agent pairs, but the theoretical guarantees are borrowed and don't cover the actual GRPO objective; referee if the numbers reproduce. the 4 major comments →

arxiv 2508.20018 v1 pith:YUGAO2NT submitted 2025-08-27 cs.AI cs.CLcs.CVcs.MA

SWIRL: A Staged Workflow for Interleaved Reinforcement Learning in Mobile GUI Control

Quanfeng Lu , Zhantao Ma , Shuai Zhong , Jin Wang , Dahai Yu , Michael K. Ng , Ping Luo This is my paper
classification cs.AI cs.CLcs.CVcs.MA
keywords multi-agent reinforcement learninginterleaved optimizationmobile GUI controllarge vision-language modelszero-shot generalizationalternating direction methodGRPO
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 reading

This paper sets out to establish that coordinated multi-agent behavior can be trained without joint multi-agent reinforcement learning: SWIRL freezes all agents except one, runs several single-agent updates on that one, then switches to the next agent, repeating in rounds. The paper argues that this round-level interleaving makes training stable, gives provable monotonic improvement of the joint return across rounds, and lets standard single-agent RL toolchains be reused unchanged. In mobile GUI control, the method splits the task between a Navigator that turns instructions and screens into low-level plans and an Interactor that grounds those plans into atomic actions, and it reports state-of-the-art zero-shot performance on high- and low-level GUI benchmarks using only 3,500 training examples, with O(1) actor-memory usage. The same alternating recipe is applied to multi-agent mathematical reasoning, where it improves MATH500 by 14.8 points over the MARFT baseline. Why this matters: if true, expensive joint MARL is unnecessary for coordination, and multi-agent training of large vision-language models becomes memory-light, stable, and compatible with existing RL pipelines.

Core claim

The paper's discovery is that coordinated multi-agent behavior emerges from strictly alternating single-agent optimization. It defines a round-level interleaved update rule in which a rolling baseline freezes all agents except the one being updated, that agent takes several micro-steps maximizing the surrogate L^i - C D_max_KL, and then the next agent is updated. Each micro-step has a lower bound on the joint return, each round is monotonically improving, and the return sequence converges. The practical instantiation trains a Navigator and an Interactor with a GRPO-style objective combined with online reweighting, and the paper reports that this two-agent system outperforms single-agent and

What carries the argument

Round-level interleaved updating with a rolling baseline: at micro-step (k, i, j), the joint policy is (tau^{-i}_k, pi^i_{k,j}), with agents earlier in the order already updated and agents later still at round k. The per-step objective F = L^i_{Pi}(tau^{-i}, pi^i) - C_{k,i,j} D^{max}_{KL}(pi^i_{k,j}, pi^i) turns every update into a single-agent trust-region problem; because F equals zero at the baseline policy, any maximizer has nonnegative F, and summing over micro-steps gives monotonic round improvement. The practical SWIRL pipeline approximates this surrogate with the GRPO objective, using clipped importance ratios and a KL anchor to a reference policy, and adds online reweighting that di

Load-bearing premise

The implemented GRPO variant (clipped importance ratios plus a KL anchor to a reference policy) is assumed to preserve the ascent direction of the theoretical surrogate L - D_max_KL, so the monotonic improvement and convergence guarantees apply to the algorithm actually trained; the paper asserts this without proof.

What would settle it

Run SWIRL's exact Stage-2 training on a fixed benchmark and record the average reward after every round; if measured return ever decreases across a round (J(pi_{k+1}) < J(pi_k)), the practical monotonic-improvement claim fails. Alternatively, train the same two agents with simultaneous GRPO updates under identical data and hyperparameters; if simultaneous training matches or exceeds SWIRL's final score, the central claim that interleaved single-agent updates are necessary for stable coordination is falsified.

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

If this is right

  • If SWIRL's central claim is correct, multi-agent LVLM training no longer requires keeping all agents resident on one device; only the currently updated agent needs to be loaded, so actor memory stays O(1) in the number of agents.
  • Standard single-agent RL frameworks, such as GRPO, can drive cooperative multi-agent training, removing the need for bespoke MARL infrastructure.
  • Zero-shot mobile GUI control can be achieved from only 3,500 curated examples by separating planning from execution, beating baselines trained by SFT and RFT on larger or more specialized data.
  • The same alternating recipe transfers to mathematical reasoning, improving MATH500 by 14.8 points over MARFT and generalizing to CMATH and GSM8K.
  • The theoretical guarantees imply that per-round returns should be non-decreasing, so training progress does not require coordinated joint gradient updates and should scale with more rounds or more agents.

Where Pith is reading between the lines

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

  • The monotonic-improvement guarantee is proven for the exact surrogate, not for the clipped GRPO objective used in practice; a natural corollary to test is that actual per-round returns under the implemented SWIRL training are non-decreasing, and the paper does not yet show that curve for every round.
  • The ablation showing that online reweighting is necessary for stable improvement suggests part of the reported stability may come from sample filtering rather than interleaving itself; disentangling these two contributions would clarify what carries the result.
  • If the interleaving principle generalizes beyond two agents and beyond LVLMs, any MARL setting with heterogeneous actors and fixed update budgets could use the same round-level decomposition, with the active model resident and the others remote.
  • The paper's parallel-update ablation hints that the strict sequential order may be relaxable; if confirmed, a Jacobian-style simultaneous update could give efficiency gains without losing the monotonic-round property.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 5 minor

Summary. The paper introduces SWIRL, a staged workflow for interleaved reinforcement learning in multi-agent systems, and applies it to mobile GUI control and mathematical reasoning. The core idea is to replace joint multi-agent policy optimization with round-level alternation: at each stage one agent is trained while its partners are frozen, reducing each update to a single-agent RL problem. The authors claim theoretical guarantees (a per-micro-step lower bound, monotonic improvement across rounds, and convergence of returns) for an idealized interleaved-update algorithm, and then present a practical GRPO-based implementation with a Navigator/Interactor architecture for GUI tasks and a Teacher/Student architecture for math. Empirically, the paper reports state-of-the-art zero-shot GUI results using only 3,500 training samples and a +14.8 improvement over MARFT on MATH500, along with ablations on interleaving, online reweighting, and update order.

Significance. If the claims were fully established, SWIRL would be a practically valuable contribution: it reduces the memory footprint of multi-agent training to O(1) in the number of agents, reuses standard single-agent RL toolchains, and reports strong empirical results on challenging GUI benchmarks. The empirical study is fairly extensive, with multiple benchmarks, ablations, and a transfer experiment to math. However, the paper’s central theoretical selling point is not currently supported for the algorithm that is actually trained. The proofs apply to an exact-argmax scheme with a max-KL penalty, while the implementation optimizes a clipped GRPO objective with a KL anchor; no argument is given that the implemented objective preserves the ascent direction of the theoretical surrogate. Because the practical method is the one evaluated, the theoretical guarantees overclaim what is demonstrated. The computational-resource argument and the empirical results are interesting, but the theory–practice gap is load-bearing and needs to be resolved before the claims as stated are credible.

major comments (4)
  1. [Sec. 3.1–3.3, Eq. (2)] The theoretical results (Proposition 1, Theorem 1, Corollary 1) are proved for Algorithm 1, where each micro-step is an exact maximizer of F_{k,i,j}(π^i) = L^i − C_{k,i,j} Dmax_KL(π^i_{k,j}, π^i). The implemented Algorithms 2/3 instead maximize the GRPO-style objective in Eq. (2), which uses clipped importance ratios and a KL anchor to a reference policy. Section 3.2 merely asserts that this "preserves the ascent direction" of the theoretical target without proof or a quantitative bound. Clipping saturates outside [1−ε, 1+ε], the KL anchor is to π_ref rather than to the current iterate π^i_{k,j} appearing in Dmax_KL, and the coefficient λ is not set to C_{k,i,j} = 4γε/(1−γ)^2. An update can therefore increase Eq. (2) while decreasing the true surrogate L^i − C Dmax_KL, breaking the F ≥ 0 condition that Theorem 1 requires. The same gap affects Corollary 1. I request either a proof that Eq
  2. [Tables 2–5 and Sec. 4.1] All reported results are single numbers with no error bars, no multiple seeds, and no significance tests. This matters because several headline comparisons are close (e.g., Table 5: SWIRL 83.5 vs. MARFT 83.0 on CMATH; Table 4: Type 84.62 vs. GUI-R1-3B 83.58). Since the paper stresses stability and consistent improvements, reporting mean±std over at least 3 seeds, or a significance test, is necessary to support the claims. Without this, the +14.8 MATH500 gain is the only large margin that is easy to distinguish from noise.
  3. [Appendix C.1 and Sec. 3.3 (online reweighting)] The two-stage data selection uses hand-set thresholds: Stage 1 keeps 1,500 samples with average reward in [0.3, 0.4]; Stage 2 keeps 2,000 samples with average reward below 0.6 and variance above 0.175; online reweighting additionally discards samples whose mean reward is outside (0.1, 1) (GUI) or (0.2, 0.8) (math). These thresholds are free hyperparameters and no sensitivity analysis is given. The ablation in Appendix D.1 shows interleaving helps relative to continued warm-up training, which is useful, but the practical contribution of online reweighting could be confounded with the filtering thresholds. Please provide a sensitivity study or at least a clear statement of how robust the reported results are to these thresholds.
  4. [Appendix A.2 / Theorem 1] The proof of Proposition 1 is a direct application of Lemma 6 of Zhong et al. (2024), and Theorem 1 is then derived by summing per-micro-step inequalities. This is mathematically clean but derivative: the stepwise safety bound and monotonic improvement are not new results beyond the cited HAPPO analysis. Given that the theoretical framework does not cover the implemented GRPO objective, the contribution of Sections 3.1–3.2 is more a re-statement of existing theory than a new guarantee. The paper would be stronger if the theory were either extended to the actual objective or clearly positioned as an idealized motivation.
minor comments (5)
  1. [Sec. 3.2, Eq. (2)] The token-level indicator I^{(j)}_{k,ℓ} is defined as '1 iff r_{k,ℓ} ∼ π_{θ_j}', but tokens are not independent samples from a single policy; they are drawn from the joint autoregressive process. The notation should be clarified to describe which tokens are generated by which agent module, and how this indicator interacts with the GRPO advantage normalization.
  2. [Sec. 4.1 and Appendix C.1] The paper does not state the discount factor γ anywhere, although it is central to the theoretical bound. Please report it, and if the implementation uses undiscounted returns (γ = 1), discuss how the theory applies.
  3. [Appendix D.3] Figure 4 is referenced in the text ('Fig. 4') but no figure appears in the manuscript body. Please include the figure or remove the reference.
  4. [Sec. 4.4 / Appendix C.3] The math experiments bypass warm-up and directly run Stage 2 on 7,500 MATH samples with 10 rounds and 1 epoch per round. This is different from the GUI setup (warm-up + 20 rounds × 2 epochs), which is fine, but the paper should state whether the same hyperparameters were used for the Teacher/Student and how the difference affects the interpretation of the transferability claim.
  5. [Sec. 3.3 / Algorithm 2] Algorithm 2 is very sparse; it does not specify the stopping criteria for the inner RL calls, how `RL` relates to Algorithm 3, or how the online reweighting step interacts with the two agents' update budgets. A more explicit pseudocode would improve reproducibility.

Circularity Check

0 steps flagged

No significant circularity: the central claims are anchored in external prior theory and held-out empirical benchmarks; the theory/implementation gap is a correctness concern, not a circular reduction.

full rationale

Walking the paper's derivation chain, the theoretical core is not circular. Algorithm 1 defines each micro-step as the exact maximizer of F_{k,i,j}(π^i) = L^i − C_{k,i,j} Dmax_KL(π^i_{k,j}, π^i), and Proposition 1 is explicitly imported from Lemma 6 of Zhong et al. (2024), an external prior work, not from the authors' own results. Theorem 1 then follows algebraically from the baseline-zero property F_{k,i,j}(π^i_{k,j}) = 0, and Corollary 1 is a standard bounded-monotone-sequence argument. No fitted parameter is renamed as a prediction: the reward weights (α, β, λ1, λ2) and reweighting thresholds (0.1 < Rx < 1, 0.2 < Rx < 0.8) are hand-set hyperparameters, not derived from the evaluation benchmarks, and the reported metrics are zero-shot on held-out datasets. The paper does contain a real validity gap: Section 3.2 asserts that the GRPO-style objective in Eq. (2) 'preserves the ascent direction of the theoretical target L − Dmax_KL' without proof, and the clip and KL anchor are not the exact max-KL penalty used in Theorem 1. This means the monotonic-improvement guarantee does not demonstrably cover the implemented Alg. 2/3, but that is a theory-implementation mismatch, not an equivalence-by-construction or a self-referential fit. The only self-citations (e.g., GUI Odyssey as an evaluation dataset) are not load-bearing for the derivation and do not make the argument circular.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 0 invented entities

The central claim depends on free hyperparameters for rewards, data filtering, and reweighting, plus two unproven assumptions: that the practical GRPO surrogate inherits the theoretical monotonic improvement property, and that the designed reward is sufficient. No new physical or conceptual entities are introduced.

free parameters (4)
  • Reward weights alpha, beta, lambda1, lambda2 = alpha=0.1, beta=0.9, lambda1=0.2, lambda2=0.8
    Hand-chosen weights for format and accuracy rewards (Sec 3.2). No sensitivity analysis.
  • Online reweighting thresholds = GUI: 0.1 < Rx < 1; math: 0.2 < Rx < 0.8
    Hand-set thresholds for keeping samples in the training batch (Sec 4.1, App C.3).
  • Training data selection ranges = Stage 1: reward in [0.3, 0.4]; Stage 2: reward < 0.6 and variance > 0.175
    Ad hoc criteria for curating the 3,500 training samples (App C.1).
  • GRPO clipping epsilon and KL anchor coefficient = not stated in main text
    Hyperparameters of the practical objective in Eq (2); values missing from the provided text.
axioms (3)
  • standard math Lemma 6 from (Zhong et al., 2024) (HAPPO trust-region bound)
    The proof of Proposition 1 directly applies this lemma. The paper does not reproduce the proof, so the theoretical guarantee stands or falls with HAPPO.
  • domain assumption The GRPO clipping and KL anchor preserve the ascent direction of the theoretical surrogate
    Asserted in Sec 3.2 without proof. This links the monotonic improvement theorem to the practical algorithm.
  • domain assumption The reward function Racc = lambda1*Ract + lambda2*Rinfo is a sufficient and well-scaled learning signal for both agents
    Used to compute advantages for Navigator and Interactor (Sec 3.2). No validation that this decomposition captures task success.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of SWIRL: A Staged Workflow for Interleaved Reinforcement Learning in Mobile GUI Control." pith.science (2026). https://pith.science/paper/YUGAO2NT

@misc{pith2026250820018,
  author       = {Pith},
  title        = {Pith review of: SWIRL: A Staged Workflow for Interleaved Reinforcement Learning in Mobile GUI Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YUGAO2NT}},
  note         = {Machine review of arXiv:2508.20018}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The rapid advancement of large vision language models (LVLMs) and agent systems has heightened interest in mobile GUI agents that can reliably translate natural language into interface operations. Existing single-agent approaches, however, remain limited by structural constraints. Although multi-agent systems naturally decouple different competencies, recent progress in multi-agent reinforcement learning (MARL) has often been hindered by inefficiency and remains incompatible with current LVLM architectures. To address these challenges, we introduce SWIRL, a staged workflow for interleaved reinforcement learning designed for multi-agent systems. SWIRL reformulates MARL into a sequence of single-agent reinforcement learning tasks, updating one agent at a time while keeping the others fixed. This formulation enables stable training and promotes efficient coordination across agents. Theoretically, we provide a stepwise safety bound, a cross-round monotonic improvement theorem, and convergence guarantees on return, ensuring robust and principled optimization. In application to mobile GUI control, SWIRL instantiates a Navigator that converts language and screen context into structured plans, and an Interactor that grounds these plans into executable atomic actions. Extensive experiments demonstrate superior performance on both high-level and low-level GUI benchmarks. Beyond GUI tasks, SWIRL also demonstrates strong capability in multi-agent mathematical reasoning, underscoring its potential as a general framework for developing efficient and robust multi-agent systems.

Figures

Figures reproduced from arXiv: 2508.20018 by Dahai Yu, Jin Wang, Michael K. Ng, Ping Luo, Quanfeng Lu, Shuai Zhong, Zhantao Ma.

Figure 1
Figure 1. Figure 1: Our multi-agent inference pipelines. Given a [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Alternating training paradigm. (a) HAPPO (Zhong et al., 2024): step-level single￾agent sequential updates; (b) A2PO (Wang et al., 2023) & MARFT (Liao et al., 2025): en￾hancing sequential updates with preceding-agent off-policy correction for greater efficiency; (c) SWIRL: round-level alternation with inner solves. We introduce SWIRL, a staged workflow for in￾terleaved reinforcement learning. SWIRL de￾compo… view at source ↗
Figure 3
Figure 3. Figure 3: Our multi-agent training pipeline. (a) SWIRL decomposes multi-agent learning into two [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: SWIRL training dynamics showing steady performance gains and stability across rounds. [PITH_FULL_IMAGE:figures/full_fig_p022_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Effect of the training schedule on SWIRL performance with varying numbers of rounds [PITH_FULL_IMAGE:figures/full_fig_p023_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Filtered sample rates with online reweighting in SWIRL. [PITH_FULL_IMAGE:figures/full_fig_p023_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Ablation results on SWIRL zero-shot performance for AndroidControl-High and [PITH_FULL_IMAGE:figures/full_fig_p024_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Prompt design for the dual-agent framework in the GUI domain. [PITH_FULL_IMAGE:figures/full_fig_p024_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Prompt design for the dual-agent framework in the mathematics domain. [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Example of GUI agents collaboratively performing a mobile GUI control task (Part 1 of [PITH_FULL_IMAGE:figures/full_fig_p026_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Example of GUI agents collaboratively performing a mobile GUI control task (Part 2 of [PITH_FULL_IMAGE:figures/full_fig_p027_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Example of GUI agents collaboratively performing a mobile GUI control task (Part 3 of [PITH_FULL_IMAGE:figures/full_fig_p028_12.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Trust Region On-Policy Distillation

    cs.LG 2026-05 unverdicted novelty 5.0

    TrOPD stabilizes on-policy distillation for LLMs with trust-region learning, outlier estimation, and off-policy guidance, outperforming prior OPD methods on reasoning and code benchmarks.

Reference graph

Works this paper leans on

75 extracted references · 9 canonical work pages · cited by 1 Pith paper · 3 internal anchors

  1. [1]

    Chain-of-thought reasoning in the wild is not always faithful

    Iv \'a n Arcuschin, Jett Janiak, Robert Krzyzanowski, Senthooran Rajamanoharan, Neel Nanda, and Arthur Conmy. Chain-of-thought reasoning in the wild is not always faithful. arXiv preprint arXiv:2503.08679, 2025

    Pith/arXiv arXiv 2025

  2. [2]

    Qwen2.5-vl technical report

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-vl technical report. a...

    Pith/arXiv arXiv 2025

  3. [3]

    Distributed optimization and statistical learning via the alternating direction method of multipliers

    Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato, Jonathan Eckstein, et al. Distributed optimization and statistical learning via the alternating direction method of multipliers. Foundations and Trends in Machine learning , 3 0 (1): 0 1--122, 2011

    2011

  4. [4]

    Multi-agent reinforcement learning: A review of challenges and applications

    Lorenzo Canese, Gian Carlo Cardarilli, Luca Di Nunzio, Rocco Fazzolari, Daniele Giardino, Marco Re, and Sergio Span \`o . Multi-agent reinforcement learning: A review of challenges and applications. Applied Sciences, 11 0 (11): 0 4948, 2021

    2021

  5. [5]

    Amex: Android multi-annotation expo dataset for mobile gui agents

    Yuxiang Chai, Siyuan Huang, Yazhe Niu, Han Xiao, Liang Liu, Dingyu Zhang, Peng Gao, Shuai Ren, and Hongsheng Li. Amex: Android multi-annotation expo dataset for mobile gui agents. arXiv preprint arXiv:2407.17490, 2024

    Pith/arXiv arXiv 2024

  6. [6]

    Guicourse: From general vision language models to versatile gui agents

    Wentong Chen, Junbo Cui, Jinyi Hu, Yujia Qin, Junjie Fang, Yue Zhao, Chongyi Wang, Jun Liu, Guirong Chen, Yupeng Huo, et al. Guicourse: From general vision language models to versatile gui agents. arXiv preprint arXiv:2406.11317, 2024

    Pith/arXiv arXiv 2024

  7. [7]

    Multi-agent deep reinforcement learning for large-scale traffic signal control

    Tianshu Chu, Jie Wang, Lara Codec \`a , and Zhaojian Li. Multi-agent deep reinforcement learning for large-scale traffic signal control. IEEE transactions on intelligent transportation systems, 21 0 (3): 0 1086--1095, 2019

    work page 2019

  8. [8]

    Sft memorizes, rl generalizes: A comparative study of foundation model post-training

    Tianzhe Chu, Yuexiang Zhai, Jihan Yang, Shengbang Tong, Saining Xie, Dale Schuurmans, Quoc V Le, Sergey Levine, and Yi Ma. Sft memorizes, rl generalizes: A comparative study of foundation model post-training. arXiv preprint arXiv:2501.17161, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    Training verifiers to solve math word problems

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021

    Pith/arXiv arXiv 2021

  10. [10]

    Process reinforcement through implicit rewards

    Ganqu Cui, Lifan Yuan, Zefan Wang, Hanbin Wang, Wendi Li, Bingxiang He, Yuchen Fan, Tianyu Yu, Qixin Xu, Weize Chen, et al. Process reinforcement through implicit rewards. arXiv preprint arXiv:2502.01456, 2025

    Pith/arXiv arXiv 2025

  11. [11]

    Is independent learning all you need in the starcraft multi-agent challenge? arXiv preprint arXiv:2011.09533, 2020

    Christian Schroeder De Witt, Tarun Gupta, Denys Makoviichuk, Viktor Makoviychuk, Philip HS Torr, Mingfei Sun, and Shimon Whiteson. Is independent learning all you need in the starcraft multi-agent challenge? arXiv preprint arXiv:2011.09533, 2020

    Pith/arXiv arXiv 2011

  12. [12]

    Improving factuality and reasoning in language models through multiagent debate

    Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. Improving factuality and reasoning in language models through multiagent debate. In Forty-first International Conference on Machine Learning, 2023

    2023

  13. [13]

    Plan-and-act: Improving planning of agents for long-horizon tasks

    Lutfi Eren Erdogan, Nicholas Lee, Sehoon Kim, Suhong Moon, Hiroki Furuta, Gopala Anumanchipalli, Kurt Keutzer, and Amir Gholami. Plan-and-act: Improving planning of agents for long-horizon tasks. arXiv preprint arXiv:2503.09572, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    On alternating direction methods of multipliers: a historical perspective

    Roland Glowinski. On alternating direction methods of multipliers: a historical perspective. Modeling, simulation and optimization for science and technology, pp.\ 59--82, 2014

    work page 2014

  15. [15]

    Towards Efficient Multi-Agent Learning Systems

    Kailash Gogineni, Peng Wei, Tian Lan, and Guru Venkataramani. Towards efficient multi-agent learning systems. arXiv preprint arXiv:2305.13411, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  16. [16]

    Dong Guo, Faming Wu, Feida Zhu, Fuxing Leng, Guang Shi, Haobin Chen, Haoqi Fan, Jian Wang, Jianyu Jiang, Jiawei Wang, et al. Seed1. 5-vl technical report. arXiv preprint arXiv:2505.07062, 2025

    Pith/arXiv arXiv 2025

  17. [17]

    Llm multi-agent systems: Challenges and open problems

    Shanshan Han, Qifan Zhang, Yuhang Yao, Weizhao Jin, and Zhaozhuo Xu. Llm multi-agent systems: Challenges and open problems. arXiv preprint arXiv:2402.03578, 2024

    Pith/arXiv arXiv 2024

  18. [18]

    Measuring mathematical problem solving with the math dataset

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. arXiv preprint arXiv:2103.03874, 2021

    Pith/arXiv arXiv 2021

  19. [19]

    Openrlhf: An easy-to-use, scalable and high-performance rlhf framework

    Jian Hu, Xibin Wu, Zilin Zhu, Weixun Wang, Dehao Zhang, Yu Cao, et al. Openrlhf: An easy-to-use, scalable and high-performance rlhf framework. arXiv preprint arXiv:2405.11143, 2024

    Pith/arXiv arXiv 2024

  20. [20]

    Owl: Optimized workforce learning for general multi-agent assistance in real-world task automation

    Mengkang Hu, Yuhang Zhou, Wendong Fan, Yuzhou Nie, Bowei Xia, Tao Sun, Ziyu Ye, Zhaoxuan Jin, Yingru Li, Qiguang Chen, et al. Owl: Optimized workforce learning for general multi-agent assistance in real-world task automation. arXiv preprint arXiv:2505.23885, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Keming Lu, et al. Qwen2. 5-coder technical report. arXiv preprint arXiv:2409.12186, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Omniact: A dataset and benchmark for enabling multimodal generalist autonomous agents for desktop and web

    Raghav Kapoor, Yash Parag Butala, Melisa Russak, Jing Yu Koh, Kiran Kamble, Waseem AlShikh, and Ruslan Salakhutdinov. Omniact: A dataset and benchmark for enabling multimodal generalist autonomous agents for desktop and web. In European Conference on Computer Vision, pp.\ 161--178. Springer, 2024

    work page 2024

  23. [23]

    Os-harm: A benchmark for measuring safety of computer use agents

    Thomas Kuntz, Agatha Duzan, Hao Zhao, Francesco Croce, Zico Kolter, Nicolas Flammarion, and Maksym Andriushchenko. Os-harm: A benchmark for measuring safety of computer use agents. arXiv preprint arXiv:2506.14866, 2025

    arXiv 2025

  24. [24]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023

    2023

  25. [25]

    Reguide: Data efficient gui grounding via spatial reasoning and search

    Hyunseok Lee, Jeonghoon Kim, Beomjun Kim, Jihoon Tack, Chansong Jo, Jaehong Lee, Cheonbok Park, Sookyo In, Jinwoo Shin, and Kang Min Yoo. Reguide: Data efficient gui grounding via spatial reasoning and search. arXiv preprint arXiv:2505.15259, 2025

    Pith/arXiv arXiv 2025

  26. [26]

    Towards better chain-of-thought: A reflection on effectiveness and faithfulness

    Jiachun Li, Pengfei Cao, Yubo Chen, Jiexin Xu, Huaijun Li, Xiaojian Jiang, Kang Liu, and Jun Zhao. Towards better chain-of-thought: A reflection on effectiveness and faithfulness. arXiv preprint arXiv:2405.18915, 2024 a

    Pith/arXiv arXiv 2024

  27. [27]

    On the effects of data scale on computer control agents

    Wei Li, William Bishop, Alice Li, Chris Rawles, Folawiyo Campbell-Ajala, Divya Tyamagundlu, and Oriana Riva. On the effects of data scale on computer control agents. arXiv preprint arXiv:2406.03679, 2024 b

    Pith/arXiv arXiv 2024

  28. [28]

    Marft: Multi-agent reinforcement fine-tuning

    Junwei Liao, Muning Wen, Jun Wang, and Weinan Zhang. Marft: Multi-agent reinforcement fine-tuning. arXiv preprint arXiv:2504.16129, 2025

    Pith/arXiv arXiv 2025

  29. [29]

    Infigui-r1: Advancing multimodal gui agents from reactive actors to deliberative reasoners

    Yuhang Liu, Pengxiang Li, Congkai Xie, Xavier Hu, Xiaotian Han, Shengyu Zhang, Hongxia Yang, and Fei Wu. Infigui-r1: Advancing multimodal gui agents from reactive actors to deliberative reasoners. arXiv preprint arXiv:2504.14239, 2025

    Pith/arXiv arXiv 2025

  30. [30]

    Gui odyssey: A comprehensive dataset for cross-app gui navigation on mobile devices

    Quanfeng Lu, Wenqi Shao, Zitao Liu, Fanqing Meng, Boxuan Li, Botong Chen, Siyuan Huang, Kaipeng Zhang, Yu Qiao, and Ping Luo. Gui odyssey: A comprehensive dataset for cross-app gui navigation on mobile devices. arXiv preprint arXiv:2406.08451, 2024

    Pith/arXiv arXiv 2024

  31. [31]

    Ui-r1: Enhancing efficient action prediction of gui agents by reinforcement learning

    Zhengxi Lu, Yuxiang Chai, Yaxuan Guo, Xi Yin, Liang Liu, Hao Wang, Han Xiao, Shuai Ren, Guanjing Xiong, and Hongsheng Li. Ui-r1: Enhancing efficient action prediction of gui agents by reinforcement learning. arXiv preprint arXiv:2503.21620, 2025

    Pith/arXiv arXiv 2025

  32. [32]

    Gui-r1: A generalist r1-style vision-language action model for gui agents

    Run Luo, Lu Wang, Wanwei He, and Xiaobo Xia. Gui-r1: A generalist r1-style vision-language action model for gui agents. arXiv preprint arXiv:2504.10458, 2025

    Pith/arXiv arXiv 2025

  33. [33]

    Mm-eureka: Exploring the frontiers of multimodal reasoning with rule-based reinforcement learning

    Fanqing Meng, Lingxiao Du, Zongkai Liu, Zhixiang Zhou, Quanfeng Lu, Daocheng Fu, Tiancheng Han, Botian Shi, Wenhai Wang, Junjun He, Kaipeng Zhang, Ping Luo, Yu Qiao, Qiaosheng Zhang, and Wenqi Shao. Mm-eureka: Exploring the frontiers of multimodal reasoning with rule-based reinforcement learning. arXiv preprint arXiv:2503.07365, 2025

    Pith/arXiv arXiv 2025

  34. [34]

    Building a stable planner: An extended finite state machine based planning module for mobile gui agent

    Fanglin Mo, Junzhe Chen, Haoxuan Zhu, and Xuming Hu. Building a stable planner: An extended finite state machine based planning module for mobile gui agent. arXiv preprint arXiv:2505.14141, 2025

    Pith/arXiv arXiv 2025

  35. [35]

    Screenagent: A vision language model-driven computer control agent

    Runliang Niu, Jindong Li, Shiqi Wang, Yali Fu, Xiyu Hu, Xueyuan Leng, He Kong, Yi Chang, and Qi Wang. Screenagent: A vision language model-driven computer control agent. arXiv preprint arXiv:2402.07945, 2024

    Pith/arXiv arXiv 2024

  36. [36]

    Introducing gpt-5

    OpenAI . Introducing gpt-5. https://openai.com/index/introducing-gpt-5/, 2025. Accessed: August 2025

    work page 2025

  37. [37]

    Ui-tars: Pioneering automated gui interaction with native agents

    Yujia Qin, Yining Ye, Junjie Fang, Haoming Wang, Shihao Liang, Shizuo Tian, Junda Zhang, Jiahao Li, Yunxin Li, Shijue Huang, et al. Ui-tars: Pioneering automated gui interaction with native agents. arXiv preprint arXiv:2501.12326, 2025

    Pith/arXiv arXiv 2025

  38. [38]

    Monotonic value function factorisation for deep multi-agent reinforcement learning

    Tabish Rashid, Mikayel Samvelyan, Christian Schroeder De Witt, Gregory Farquhar, Jakob Foerster, and Shimon Whiteson. Monotonic value function factorisation for deep multi-agent reinforcement learning. Journal of Machine Learning Research, 21 0 (178): 0 1--51, 2020

    2020

  39. [39]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R \"a dle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024

    Pith/arXiv arXiv 2024

  40. [40]

    Android in the wild: A large-scale dataset for android device control

    Christopher Rawles, Alice Li, Daniel Rodriguez, Oriana Riva, and Timothy Lillicrap. Android in the wild: A large-scale dataset for android device control. arXiv preprint arXiv:2307.10088, 2023

    Pith/arXiv arXiv 2023

  41. [41]

    Trust region policy optimization

    John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region policy optimization. In International conference on machine learning, pp.\ 1889--1897. PMLR, 2015

    2015

  42. [42]

    Proximal policy optimization algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017

    Pith/arXiv arXiv 2017

  43. [43]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024

    Pith/arXiv arXiv 2024

  44. [44]

    Hybridflow: A flexible and efficient rlhf framework

    Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. Hybridflow: A flexible and efficient rlhf framework. In Proceedings of the Twentieth European Conference on Computer Systems, pp.\ 1279--1297, 2025

    2025

  45. [45]

    Towards trustworthy gui agents: A survey

    Yucheng Shi, Wenhao Yu, Wenlin Yao, Wenhu Chen, and Ninghao Liu. Towards trustworthy gui agents: A survey. arXiv preprint arXiv:2503.23434, 2025

    arXiv 2025

  46. [46]

    Teaching models to balance resisting and accepting persuasion

    Elias Stengel-Eskin, Peter Hase, and Mohit Bansal. Teaching models to balance resisting and accepting persuasion. arXiv preprint arXiv:2410.14596, 2024

    Pith/arXiv arXiv 2024

  47. [47]

    Multiagent finetuning: Self improvement with diverse reasoning chains

    Vighnesh Subramaniam, Yilun Du, Joshua B Tenenbaum, Antonio Torralba, Shuang Li, and Igor Mordatch. Multiagent finetuning: Self improvement with diverse reasoning chains. arXiv preprint arXiv:2501.05707, 2025

    Pith/arXiv arXiv 2025

  48. [48]

    Value-decomposition networks for cooperative multi-agent learning

    Peter Sunehag, Guy Lever, Audrunas Gruslys, Wojciech Marian Czarnecki, Vinicius Zambaldi, Max Jaderberg, Marc Lanctot, Nicolas Sonnerat, Joel Z Leibo, Karl Tuyls, et al. Value-decomposition networks for cooperative multi-agent learning. arXiv preprint arXiv:1706.05296, 2017

    Pith/arXiv arXiv 2017

  49. [49]

    Gui-g ^2 : Gaussian reward modeling for gui grounding

    Fei Tang, Zhangxuan Gu, Zhengxi Lu, Xuyang Liu, Shuheng Shen, Changhua Meng, Wen Wang, Wenqi Zhang, Yongliang Shen, Weiming Lu, et al. Gui-g ^2 : Gaussian reward modeling for gui grounding. arXiv preprint arXiv:2507.15846, 2025

    Pith/arXiv arXiv 2025

  50. [50]

    Multi-agent collaboration mechanisms: A survey of llms

    Khanh-Tung Tran, Dung Dao, Minh-Duong Nguyen, Quoc-Viet Pham, Barry O'Sullivan, and Hoang D Nguyen. Multi-agent collaboration mechanisms: A survey of llms. arXiv preprint arXiv:2501.06322, 2025

    Pith/arXiv arXiv 2025

  51. [51]

    Language models don't always say what they think: Unfaithful explanations in chain-of-thought prompting

    Miles Turpin, Julian Michael, Ethan Perez, and Samuel Bowman. Language models don't always say what they think: Unfaithful explanations in chain-of-thought prompting. Advances in Neural Information Processing Systems, 36: 0 74952--74965, 2023

    2023

  52. [52]

    Gui agents with foundation models: A comprehensive survey

    Shuai Wang, Weiwen Liu, Jingxuan Chen, Yuqi Zhou, Weinan Gan, Xingshan Zeng, Yuhan Che, Shuai Yu, Xinlong Hao, Kun Shao, et al. Gui agents with foundation models: A comprehensive survey. arXiv preprint arXiv:2411.04890, 2024

    Pith/arXiv arXiv 2024

  53. [53]

    Model-based multi-agent reinforcement learning: Recent progress and prospects

    Xihuai Wang, Zhicheng Zhang, and Weinan Zhang. Model-based multi-agent reinforcement learning: Recent progress and prospects. arXiv preprint arXiv:2203.10603, 2022

    Pith/arXiv arXiv 2022

  54. [54]

    Order Matters: Agent-by-agent Policy Optimization

    Xihuai Wang, Zheng Tian, Ziyu Wan, Ying Wen, Jun Wang, and Weinan Zhang. Order matters: Agent-by-agent policy optimization. arXiv preprint arXiv:2302.06205, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  55. [55]

    Mp-gui: Modality perception with mllms for gui understanding

    Ziwei Wang, Weizhi Chen, Leyang Yang, Sheng Zhou, Shengchu Zhao, Hanbei Zhan, Jiongchao Jin, Liangcheng Li, Zirui Shao, and Jiajun Bu. Mp-gui: Modality perception with mllms for gui understanding. In Proceedings of the Computer Vision and Pattern Recognition Conference, pp.\ 29711--29721, 2025

    work page 2025

  56. [56]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35: 0 24824--24837, 2022

    2022

  57. [57]

    Cmath: Can your language model pass chinese elementary school math test? arXiv preprint arXiv:2306.16636, 2023

    Tianwen Wei, Jian Luan, Wei Liu, Shuang Dong, and Bin Wang. Cmath: Can your language model pass chinese elementary school math test? arXiv preprint arXiv:2306.16636, 2023

    Pith/arXiv arXiv 2023

  58. [58]

    Autogen: Enabling next-gen llm applications via multi-agent conversations

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, et al. Autogen: Enabling next-gen llm applications via multi-agent conversations. In First Conference on Language Modeling, 2024 a

    2024

  59. [59]

    Os-atlas: A foundation action model for generalist gui agents

    Zhiyong Wu, Zhenyu Wu, Fangzhi Xu, Yian Wang, Qiushi Sun, Chengyou Jia, Kanzhi Cheng, Zichen Ding, Liheng Chen, Paul Pu Liang, et al. Os-atlas: A foundation action model for generalist gui agents. arXiv preprint arXiv:2410.23218, 2024 b

    Pith/arXiv arXiv 2024

  60. [60]

    Towards system 2 reasoning in llms: Learning how to think with meta chain-of-thought

    Violet Xiang, Charlie Snell, Kanishk Gandhi, Alon Albalak, Anikait Singh, Chase Blagden, Duy Phung, Rafael Rafailov, Nathan Lile, Dakota Mahan, et al. Towards system 2 reasoning in llms: Learning how to think with meta chain-of-thought. arXiv preprint arXiv:2501.04682, 2025

    Pith/arXiv arXiv 2025

  61. [61]

    Tradingagents: Multi-agents llm financial trading framework

    Yijia Xiao, Edward Sun, Di Luo, and Wei Wang. Tradingagents: Multi-agents llm financial trading framework. arXiv preprint arXiv:2412.20138, 2024

    Pith/arXiv arXiv 2024

  62. [62]

    Aguvis: Unified pure vision agents for autonomous gui interaction

    Yiheng Xu, Zekun Wang, Junli Wang, Dunjie Lu, Tianbao Xie, Amrita Saha, Doyen Sahoo, Tao Yu, and Caiming Xiong. Aguvis: Unified pure vision agents for autonomous gui interaction. arXiv preprint arXiv:2412.04454, 2024

    Pith/arXiv arXiv 2024

  63. [63]

    A survey of admm variants for distributed optimization: Problems, algorithms and features

    Yu Yang, Xiaohong Guan, Qing-Shan Jia, Liang Yu, Bolun Xu, and Costas J Spanos. A survey of admm variants for distributed optimization: Problems, algorithms and features. arXiv preprint arXiv:2208.03700, 2022

    Pith/arXiv arXiv 2022

  64. [64]

    Enhancing visual grounding for gui agents via self-evolutionary reinforcement learning

    Xinbin Yuan, Jian Zhang, Kaixin Li, Zhuoxuan Cai, Lujian Yao, Jie Chen, Enguang Wang, Qibin Hou, Jinwei Chen, Peng-Tao Jiang, et al. Enhancing visual grounding for gui agents via self-evolutionary reinforcement learning. arXiv preprint arXiv:2505.12370, 2025

    Pith/arXiv arXiv 2025

  65. [65]

    Android in the zoo: Chain-of-action-thought for gui agents

    Jiwen Zhang, Jihao Wu, Yihua Teng, Minghui Liao, Nuo Xu, Xiao Xiao, Zhongyu Wei, and Duyu Tang. Android in the zoo: Chain-of-action-thought for gui agents. arXiv preprint arXiv:2403.02713, 2024

    Pith/arXiv arXiv 2024

  66. [66]

    Does chain-of-thought reasoning help mobile gui agent? an empirical study

    Li Zhang, Longxi Gao, and Mengwei Xu. Does chain-of-thought reasoning help mobile gui agent? an empirical study. arXiv preprint arXiv:2503.16788, 2025

    Pith/arXiv arXiv 2025

  67. [67]

    Longagent: scaling language models to 128k context through multi-agent collaboration

    Jun Zhao, Can Zu, Hao Xu, Yi Lu, Wei He, Yiwen Ding, Tao Gui, Qi Zhang, and Xuanjing Huang. Longagent: scaling language models to 128k context through multi-agent collaboration. arXiv preprint arXiv:2402.11550, 2024 a

    Pith/arXiv arXiv 2024

  68. [68]

    Sirius: Self-improving multi-agent systems via bootstrapped reasoning

    Wanjia Zhao, Mert Yuksekgonul, Shirley Wu, and James Zou. Sirius: Self-improving multi-agent systems via bootstrapped reasoning. arXiv preprint arXiv:2502.04780, 2025

    Pith/arXiv arXiv 2025

  69. [69]

    An Electoral Approach to Diversify LLM-based Multi-Agent Collective Decision-Making

    Xiutian Zhao, Ke Wang, and Wei Peng. An electoral approach to diversify llm-based multi-agent collective decision-making. arXiv preprint arXiv:2410.15168, 2024 b

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  70. [70]

    Heterogeneous-agent reinforcement learning

    Yifan Zhong, Jakub Grudzien Kuba, Xidong Feng, Siyi Hu, Jiaming Ji, and Yaodong Yang. Heterogeneous-agent reinforcement learning. Journal of Machine Learning Research, 25 0 (32): 0 1--67, 2024

    work page 2024

  71. [71]

    Gui-g1: Understanding r1-zero-like training for visual grounding in gui agents

    Yuqi Zhou, Sunhao Dai, Shuai Wang, Kaiwen Zhou, Qinglin Jia, and Jun Xu. Gui-g1: Understanding r1-zero-like training for visual grounding in gui agents. arXiv preprint arXiv:2505.15810, 2025

    Pith/arXiv arXiv 2025

  72. [72]

    Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, et al. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprint arXiv:2504.10479, 2025

    Pith/arXiv arXiv 2025

  73. [73]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...

  74. [74]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...

  75. [75]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.