Pith. sign in

REVIEW 3 major objections 5 minor 44 references

Argus: A General-Purpose Agentic Runtime for Long-Horizon Reasoning

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Argus, a fixed-weight agentic runtime with role-separated, verification-gated state, reaches about 78% on SWE-Bench Pro versus 59% for Direct Copilot at 1.41× aggregate tokens.

desk verdict A credible, unusually candid systems paper worth refereeing, but its two headline numbers (78% vs 59%, 21% savings) are softer than they look because the baseline traces were not retained and the self-evolution evidence is observational. read the letter →

arxiv 2608.05144 v1 pith:5ZAXOPKB submitted 2026-08-05 cs.AI

Boxiu Li , Zimo Wen , Yijia Fan , Junxiang Lei , Sufeng Guo , Jiaao Wu , Ruize Tang , Mukai Li
show 18 more authors
Yifei Shen Xiaoyu Chen Wanbo Zhang Runjing Gu Yifei Gao Yuheng Wu Xuyao Huang Zelong Zhao Jiachen Zhang Shibo Hu Hangxi Guo Yilin Chen Yuzhe Zhang Fan Yang Chuan Wen Xian Zhang Xuanhe Zhou Zhijie Deng
This is my paper
classification cs.AI
keywords agenticruntimelong-horizonreasoningverification-gatedadmissionroleseparationself-evolutionSWE-BenchPropersistentstateautonomousresearch
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

Long-horizon reasoning, this paper argues, is not a race to push farther along a fixed route: an agent should persist while evidence supports its approach and pivot when measurements expose a misspecified objective, a hidden constraint, or a failed route. The paper introduces Argus, a persistent runtime in which Manager, Planner, Engineer, and Reviewer roles execute bounded missions over durable project state, and no memory, skill, verifier, routing decision, or rejected route becomes reusable without role-owned review and, where it exists, task-native verification. The underlying model weights stay fixed; self-evolution happens in persistent state and control policy, not in parameters. Across seven GPT-5.5 benchmark arenas the paper reports about 78% on SWE-Bench Pro versus 59% for Direct Copilot at 1.41× aggregate tokens, plus mature waves that use 21% fewer solve-input tokens and 15% less active time per task than startup waves. A sympathetic reader would take the central claim to be that objective revision can be made evidence-backed and auditable, so pivoting is distinguishable from rationalized failure and verified progress can accumulate across missions.

What carries the argument

The load-bearing machinery is the bounded-mission loop over durable state with role-separated ownership. A Manager commits the campaign and moves it between Stages, a Planner authors bounded tasks, an Engineer executes against real artifacts, and a Reviewer returns structured done, continue, or blocked verdicts; the mission trajectory $\tau_t$ records states, actions, artifacts, measurements, and review outcomes, while a curated checkpoint and an append-only event tape carry cross-session continuity. Reusable state $H_t=\{Memory,Skills,Tools,Verifiers,Routing\}$ changes only through verification-gated admission: a generated candidate is checked against artifacts and task-native evidence by the responsible role and then committed by its authorized owner. The named ManagerAdmit operator is the analytical projection of these distributed updates that makes a material refinement of the contract $K_t$ explicit rather than an implicit consequence of replanning.

What would settle it

Re-run the 731 SWE-Bench Pro tasks with Direct Copilot under the identical task order, environment, and verifier conditions while retaining per-phase token and time traces; the headline comparison collapses if the baseline reaches about 78% or if the token ratio falls near 1.0. Separately, replay the same task order against a frozen runtime state; the self-evolution claim is falsified if the mature-wave token and time savings persist unchanged in that replay.

Watch

Extended reading notes

Core claim

The central claim is that one fixed-weight model, wrapped in a verification-gated persistent runtime, can combine breadth, recovery, and accumulation that are usually traded off: it repairs software, optimizes kernels and training runs, assists research, synthesizes mathematical data, carries proof campaigns, produces manuscripts, and drives hardware and chemistry verticals without changing model weights. The runtime makes objective revision explicit through the report-level contract $K_t=(\iota,o_t,c_t,v_t)$, separating stable user intent $\iota$ from operational objective, constraints, and verification criteria; a material contract change passes through a ManagerAdmit operator that requires evidence, a recorded verdict, and user or Manager authority. On the SWE-Bench Pro arena, the paper's headline quantitative assertion is approximately 78% accuracy versus 59% for Direct Copilot at 1.41× aggregate tokens, with the longitudinal observation that mature waves use 21% fewer solve-input tokens and 15% less active time per task than startup waves. The report presents these longitudinal differences as observational evidence of state accumulation rather than a controlled causal ablation, and it tracks review behavior as part of the same gate: 466 of 731 tasks invoke an independent Reviewer, 43 receive revision requests, 34 later pass the official verifier, and 22 complete the strict review loop.

Load-bearing premise

The load-bearing premise is that the benchmark protocol measures the runtime fairly: the 19-point accuracy gap and the 1.41× token ratio assume Direct Copilot ran under the same task ordering, environment, and verifier conditions, while the startup-to-mature savings assume the task sequence itself did not become easier; the paper states that per-wave Copilot traces were not retained and that no matched frozen-state replay is available.

Editorial extensions

If this is right

  • If the headline SWE-Bench Pro numbers hold, a fixed-weight runtime can beat a direct agentic loop on long software tasks by spending 1.41× the tokens, so verification overhead is an investment rather than a tax.
  • If the longitudinal trend generalizes, a deployed runtime should become cheaper per task as it accumulates reviewed skills and repository knowledge, with no retraining of the model.
  • If the Reviewer recovery numbers transfer, independent review is a working error-correction channel: 34 of 43 revision-requested tasks later passed the official verifier, and 22 completed the strict revision-to-done loop.
  • If the process-to-capability account is correct, retained typed trajectories are a training resource that final artifacts omit, because they carry objectives, actions, measurements, review verdicts, and state updates.
  • If the verticals are representative, the same orchestration can span software, proofs, manuscripts, chip synthesis, and materials research, because task-native verifiers, not domain-specific scaffolds, supply the gate.

Reading between the lines

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

  • A matched frozen-state replay of the same SWE-Bench Pro task order would likely shrink the startup-to-mature savings, because the current comparison mixes state accumulation with task-order and composition changes; the paper itself names this as the missing control.
  • Randomizing Reviewer routing would test whether independent review causes the observed recovery or merely gets assigned to recoverable tasks; the current adaptive routing confounds selection with treatment.
  • The process-data dominance proposition implies that future post-training should consume typed trajectories rather than final artifacts, but only after context-budget compression; raw event tapes are too large and contradictory to use directly.
  • The chip and materials verticals suggest that a large part of the runtime's near-term value is scope certification: both campaigns' most durable outputs are explicit statements of what was and was not verified, not just the artifacts.
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

3 major / 5 minor

Summary. The paper presents Argus, a persistent agentic runtime with four roles (Manager, Planner, Engineer, Reviewer) and verification-gated admission of memories, skills, procedures, verifiers, routing decisions, and rejected routes into durable project state, while model parameters remain fixed. The central claims are: (i) across seven GPT-5.5 benchmark arenas Argus reaches about 78% on SWE-Bench Pro versus 59% for Direct Copilot at 1.41x aggregate tokens; (ii) in a 731-task SWE-Bench Pro run, mature waves use 21% fewer solve-input tokens and 15% less active workflow time per task than startup waves, with 34 verifier recoveries and 22 strict review-loop rescues; and (iii) the runtime can carry complete research programs, evidenced by a mathematical campaign, six paper-production pipelines, an ACE-2 chip synthesis with static timing closure, and a materials-generation campaign that simplifies a published method. The formal contribution includes a contract-refinement model, a role-state machine, and a process-to-capability theory with equations for reuse value and review precision. The paper is unusually candid about the limitations of its longitudinal and case-study evidence, including the absence of retained Direct Copilot traces and the lack of a matched frozen-state replay.

Significance. If the headline results hold, the runtime design is a meaningful contribution: separating intent from operational contract, gating all persistent state updates by evidence and role ownership, and retaining rejected routes as first-class state are sensible and well-motivated mechanisms for long-horizon agentic reasoning. The evaluation includes several strengths: the endpoint benchmarks are measured against external verifiers (SWE-Bench Pro acceptance tests, MLE-Bench Kaggle scores, BPB, MOFChecker), the ACE-2 and materials verticals use verifiers the runtime does not own, the RWKV6 kernel was externally reviewed and merged upstream, and the paper explicitly labels the self-evolution result as observational rather than causal. The formalization of process-data dominance and reuse value is also useful framing. However, the two most prominent quantitative claims—the 78% versus 59% comparison and the 21%/15% mature-stage efficiency gain—are not currently auditable or causally isolated from the released artifacts, so their weight in the abstract and contribution list exceeds what the evidence supports.

major comments (3)
  1. [§5.2, Eq. (17)] The headline comparison 'about 78% vs 59% for Direct Copilot at 1.41× aggregate Tokens' cannot be reconstructed from the released artifacts because the paper states that raw Copilot Token totals and per-Wave resource traces were not retained. The 1.41× ratio in Eq. (17) and the 19-point gap in Table 4 therefore rest on an unverifiable denominator. Please release the retained aggregate Copilot token count and any per-task or per-wave breakdown, or explicitly mark the comparison as an unreproducible internal measurement in the abstract and in contribution C6.
  2. [§5.3, §6.3, Figure 4] The longitudinal claim that mature Waves use 21% fewer solve-input tokens and 15% less active time is explicitly observational, and the authors correctly state that no matched frozen-state replay is available. However, the data in Table 10 and Figure 4 show strong non-monotonicity (W13–18 has the lowest tokens but the highest active time; W23–24 rebounds), so the startup-versus-mature difference may be driven by task composition and difficulty rather than accumulated state. Please add a matched subset analysis (for example, tasks from shared repositories or difficulty strata across windows) or soften the claim in the abstract and C3 to a correlational observation.
  3. [§5.6, Eq. (13), Table 13] Equation (13) defines reuse value G_L as a counterfactual comparison against a frozen state on matched future tasks, but the measurement in Section 6.3 and Appendix D substitutes the raw startup-versus-mature window difference (0.62M Tokens and 1.27 min per task). That substitution is not a valid estimate of G_L under the paper's own definition, because the windows are not matched and no frozen-state counterfactual exists. Please either relabel the reported quantity as a descriptive window difference or provide a matched estimate before using G_L in the theory-to-measurement mapping.
minor comments (5)
  1. [Section 1] The passage beginning 'C19/C25/C24/C22/C1/C20/...' after the author list appears to be a corrupted rendering; the published version should contain the intended text or be removed.
  2. [§5.2] The term 'Direct Copilot' is used as the baseline, but its protocol is not defined: please state whether it is the same GPT-5.5/xhigh model without the Argus runtime, what task ordering it followed, and whether it used the same verifier and environment reset conditions.
  3. [§6.9, Table 7] The sentence 'the K=8 difference survives a paired test (p=0.0486)' does not name the test used; please specify the statistical procedure and whether it accounts for multiple comparisons.
  4. [Table 10] The wave table omits Waves 13 and 15 without an explicit note; please state in the caption that these are the two incomplete Waves excluded from grouped means.
  5. [§6.7] The discrepancy between the compositional-matching pipeline's completed submission Stage and its stale BLOCKED assurance snapshot is an important systems observation; consider moving it from a passing remark to a dedicated discussion point, since it bears directly on the paper's claim that verification state and pipeline state are synchronized.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: headline benchmark numbers are anchored to external verifiers and the self-evolution comparison is explicitly observational.

full rationale

The load-bearing quantitative claims do not reduce to their own inputs by construction. The SWE-Bench Pro accuracy and token ratio (Section 6.1, Eq. 17) are measured against the official SWE-Bench Pro acceptance tests; the token ratio uses recorded Argus totals and a Direct Copilot total, and the disclosure that Copilot raw traces were not retained (Section 5.2) is an auditability limitation, not a definitional equivalence. The seven-arena results are anchored to external or task-native evaluators (acceptance tests, BPB, verifier runs, MLE-Bench grader, MOFChecker), and the ACE-2 chip claims close against OpenSTA static timing rather than against Argus's own Reviewer. The longitudinal comparison, 'mature Waves use 21% fewer solve input Tokens and 15% less active workflow time per task than startup,' is explicitly labeled observational: the paper states the startup window 'begins with less project-specific state,' that 'no matched frozen-state replay is available,' and Section 8 calls it 'observational rather than a causal learning ablation.' The process-theory equations (Eqs. 4, 13, 14) are definitions, and Appendix D presents their numerical substitutions as interpretations rather than as derivations of any benchmark outcome. The sole self-citation, 'Argus Team (2026),' supplies terminology and formal definitions from the project website and is not used to justify any empirical result. No step exhibits the pattern of Eq. X = Eq. Y by construction or of a fitted parameter being renamed as a prediction.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The paper does not introduce new physical or mathematical entities. It introduces a software architecture (roles, persistent state, ManagerAdmit) and several theoretical quantities (rho_I, eta_r/a/v, G_L, Y_PRI) that are defined as protocol-specific metrics rather than invented entities with falsifiable handles. The only effectively free choices are the window definitions and the attribution labels, which are post-hoc descriptive decisions, not fitted parameters. The ledger lists the domain assumptions about benchmark validity, trace fidelity, and causal interpretation that the central claims depend on.

assumptions (5)
  • domain assumption The benchmark metrics (SWE-Bench Pro verifier, nanochat BPB, SOL score, AARRI task set, Math-Reasoning gap) are valid task-native measures of the claims made.
    Section 5.2 and Table 3 define each benchmark; the paper assumes the external verifiers are correct and that the reported reference baselines (59% Direct Copilot, AARRI paper best 68.3%, etc.) are trustworthy. This is standard in agent evaluation, but it is a load-bearing assumption for the capability claims.
  • domain assumption The role-based runtime, including Manager, Planner, Engineer, and Reviewer behavior, is implemented as described and the traces correctly attribute outcomes to these roles.
    Section 4 and Algorithm 1 describe the loop. The empirical claims about Review routing, blocked tasks, and stage rollbacks depend on the logging and attribution being reliable. The paper admits limited trace release and that some review snapshots are model-generated.
  • domain assumption The longitudinal comparison of startup vs mature windows assumes that task identity, repository mix, difficulty, and execution latency are not the sole driver of the observed token/time reduction.
    Section 5.3 explicitly states 'It is not, by itself, a causal estimate of the value of that state' and Section 8 lists missing matched frozen-state runs. The claim that the 21%/15% reduction is 'consistent with runtime self-evolution' rests on this assumption, which is acknowledged but not controlled.
  • domain assumption The probability model of Reviewer precision (Equation 15) assumes the Bayes rule with independent sensitivity and false-acceptance, and that the observed recovery rates approximate these.
    Section 3.9 derives Pr(C=1|A=1) using a standard Bayes decomposition; the paper does not measure alpha and beta directly, only the recovery funnel (43 revisions, 34 verifier passes, 22 strict rescues). The theoretical model is not fitted but also not directly validated.
  • domain assumption External verticals (ACE-2 static timing, MOFChecker, upstream kernel PR #1045) are correctly executed and the external tools are trusted.
    Sections 6.8 and 6.9 present these as external checkers. The paper states the chip certificate enumerates its exclusions and the MOF results are scored by the same validator the selection optimizes against, a acknowledged limitation. The upstream PR was reviewed by a human maintainer, which is external evidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Argus: A General-Purpose Agentic Runtime for Long-Horizon Reasoning." pith.science (2026). https://pith.science/paper/5ZAXOPKB

@misc{pith2026260805144,
  author       = {Pith},
  title        = {Pith review of: Argus: A General-Purpose Agentic Runtime for Long-Horizon Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5ZAXOPKB}},
  note         = {Machine review of arXiv:2608.05144}
}
read the original abstract

Long-horizon reasoning requires an agentic runtime that can persist when evidence supports its current approach and pivot when measurements reveal failure, hidden constraints, or a misspecified objective. We present Argus, a persistent, self-evolving runtime in which Manager, Planner, Engineer, and Reviewer execute bounded missions over durable project state. Argus separates stable user intent from operational objectives, constraints, and verification criteria, and admits memories, skills, procedures, verifiers, routing decisions, and rejected routes only after role-owned review and, when available, task-native verification. Model weights remain fixed; self-evolution occurs through persistent runtime state and control policy, with autonomous execution between operator-owned escalation points. Across seven GPT-5.5 benchmark arenas, Argus achieves about 78% on SWE-Bench Pro versus 59% for Direct Copilot while using 1.41 times the aggregate tokens. After verification-gated self-evolution, mature SWE-Bench waves use 21% fewer solve-input tokens and 15% less active workflow time per task than startup waves, while recording 34 verifier recoveries and 22 strict review-loop rescues. Argus also reaches 76.8% on AARRI-Bench and a 28.0-point gap on mathematical data synthesis, with competitive GPU-kernel and language-model-training results. Beyond benchmarks, an optimized RWKV6 kernel was merged upstream; a multi-day mathematics campaign retained falsified routes and proof-backed frontier updates; and six paper pipelines completed 254 missions with 16 stage rollbacks. These results show that a fixed-weight, self-evolving harness can revise, recover, and accumulate verified approaches while producing structured trajectories for future supervised and reinforcement learning.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 9 canonical work pages

  1. [1]

    arXiv preprint arXiv:2210.03629 , year =

    Yao, Shunyu and others , title =. arXiv preprint arXiv:2210.03629 , year =. 2210.03629 , archivePrefix=

    arXiv

  2. [2]

    arXiv preprint arXiv:2303.11366 , year =

    Shinn, Noah and others , title =. arXiv preprint arXiv:2303.11366 , year =. 2303.11366 , archivePrefix=

    arXiv

  3. [3]

    arXiv preprint arXiv:2305.16291 , year =

    Wang, Guanzhi and others , title =. arXiv preprint arXiv:2305.16291 , year =. 2305.16291 , archivePrefix=

    arXiv

  4. [4]

    Toolformer: Language Models Can Teach Themselves to Use Tools , journal =

    Schick, Timo and Dwivedi-Yu, Jane and Dess. Toolformer: Language Models Can Teach Themselves to Use Tools , journal =. 2023 , eprint =

    work page 2023

  5. [5]

    and Cai, Carrie J

    Park, Joon Sung and O'Brien, Joseph C. and Cai, Carrie J. and others , title =. arXiv preprint arXiv:2304.03442 , year =. 2304.03442 , archivePrefix=

    arXiv

  6. [6]

    arXiv preprint arXiv:2310.08560 , year =

    Packer, Charles and Wooders, Sarah and Lin, Kevin and others , title =. arXiv preprint arXiv:2310.08560 , year =. 2310.08560 , archivePrefix=

    arXiv

  7. [7]

    arXiv preprint arXiv:2409.07429 , year =

    Wang, Zora Zhiruo and Mao, Jiayuan and Fried, Daniel and Neubig, Graham , title =. arXiv preprint arXiv:2409.07429 , year =. 2409.07429 , archivePrefix=

    arXiv

  8. [8]

    arXiv preprint arXiv:2305.20050 , year =

    Lightman, Hunter and Kosaraju, Vineet and Burda, Yura and Edwards, Harri and Baker, Bowen and Lee, Teddy and Leike, Jan and Schulman, John and Sutskever, Ilya and Cobbe, Karl , title =. arXiv preprint arXiv:2305.20050 , year =. 2305.20050 , archivePrefix=

    arXiv

Show all 44 references
  1. [9]

    arXiv preprint arXiv:2408.03314 , year =

    Snell, Charlie and Lee, Jaehoon and Xu, Kelvin and Kumar, Aviral , title =. arXiv preprint arXiv:2408.03314 , year =. 2408.03314 , archivePrefix=

    arXiv

  2. [10]

    The Annals of Mathematical Statistics , volume =

    Blackwell, David , title =. The Annals of Mathematical Statistics , volume =. 1953 , doi =

    1953

  3. [11]

    arXiv preprint arXiv:2404.07738 , year =

    Baek, Jinheon and Jauhar, Sujay Kumar and Cucerzan, Silviu and Hwang, Sung Ju , title =. arXiv preprint arXiv:2404.07738 , year =. 2404.07738 , archivePrefix=

    arXiv

  4. [12]

    , title =

    Ghafarollahi, Alireza and Buehler, Markus J. , title =. arXiv preprint arXiv:2409.05556 , year =. 2409.05556 , archivePrefix=

    arXiv

  5. [13]

    International Conference on Learning Representations , year =

    Weng, Yixuan and Zhu, Minjun and Bao, Guangsheng and Zhang, Hongbo and Wang, Jindong and Zhang, Yue and Yang, Linyi , title =. International Conference on Learning Representations , year =

  6. [14]

    arXiv preprint arXiv:2502.18864 , year =

    Gottweis, Juraj and Weng, Wei-Hung and Daryin, Alexander and others , title =. arXiv preprint arXiv:2502.18864 , year =. 2502.18864 , archivePrefix=

    arXiv

  7. [15]

    arXiv preprint arXiv:2503.18102 , year =

    Schmidgall, Samuel and Moor, Michael , title =. arXiv preprint arXiv:2503.18102 , year =. 2503.18102 , archivePrefix=

    arXiv

  8. [16]

    arXiv preprint arXiv:2605.31468 , year =

    Qian, Weitong and Xu, Beicheng and Xie, Zhongao and others , title =. arXiv preprint arXiv:2605.31468 , year =. 2605.31468 , archivePrefix=

    arXiv

  9. [17]

    arXiv preprint arXiv:2606.31651 , year =

    Tang, Qiong and Sun, Tianxiang and Hu, Xiangkun and others , title =. arXiv preprint arXiv:2606.31651 , year =. 2606.31651 , archivePrefix=

    arXiv

  10. [18]

    arXiv preprint arXiv:2407.03502 , year =

    Mitra, Arindam and Del Corro, Luciano and Zheng, Guoqing and Mahajan, Shweti and Rouhana, Dany and Codas, Andres and others , title =. arXiv preprint arXiv:2407.03502 , year =. 2407.03502 , archivePrefix=

    arXiv

  11. [19]

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

    Chen, Zehui and Liu, Kuikun and Wang, Qiuchen and Zhang, Wenwei and Liu, Jiangning and Lin, Dahua and Chen, Kai and Zhao, Feng , title =. Findings of the Association for Computational Linguistics: ACL 2024 , pages =. 2024 , doi =

    2024

  12. [20]

    and Yang, Yuqing , title =

    Luo, Xufang and Zhang, Yuge and He, Zhiyuan and Wang, Zilong and Zhao, Siyun and Li, Dongsheng and Qiu, Luna K. and Yang, Yuqing , title =. arXiv preprint arXiv:2508.03680 , year =. 2508.03680 , archivePrefix=

    arXiv

  13. [21]

    arXiv preprint arXiv:2502.12110 , year =

    Xu, Wujiang and Liang, Zujie and Mei, Kai and others , title =. arXiv preprint arXiv:2502.12110 , year =. 2502.12110 , archivePrefix=

    arXiv

  14. [22]

    arXiv preprint arXiv:2405.15793 , year =

    Yang, John and others , title =. arXiv preprint arXiv:2405.15793 , year =. 2405.15793 , archivePrefix=

    arXiv

  15. [23]

    arXiv preprint arXiv:2407.16741 , year =

    Wang, Xingyao and Li, Boxuan and Song, Yufan and others , title =. arXiv preprint arXiv:2407.16741 , year =. 2407.16741 , archivePrefix=

    arXiv

  16. [24]

    and Yang, John and Wettig, Alexander and others , title =

    Jimenez, Carlos E. and Yang, John and Wettig, Alexander and others , title =. arXiv preprint arXiv:2310.06770 , year =. 2310.06770 , archivePrefix=

    arXiv

  17. [25]

    arXiv preprint arXiv:2509.16941 , year =

    Xiang Deng and Jeff Da and Edwin Pan and Yannis Yiming He and Charles Ide and Kanak Garg and Niklas Lauffer and Andrew Park and Nitin Pasari and Chetan Rane and Karmini Sampath and Maya Krishnan and Srivatsa Kundurthy and Sean Hendryx and Zifan Wang and Vijay Bharadwaj and Jef...

    arXiv

  18. [26]

    arXiv preprint arXiv:2308.03688 , year =

    Liu, Xiao and Yu, Hao and Zhang, Hanchen and others , title =. arXiv preprint arXiv:2308.03688 , year =. 2308.03688 , archivePrefix=

    arXiv

  19. [27]

    arXiv preprint arXiv:2401.13178 , year =

    Ma, Chang and Zhang, Junlei and Zhu, Zhihao and others , title =. arXiv preprint arXiv:2401.13178 , year =. 2401.13178 , archivePrefix=

    arXiv

  20. [28]

    arXiv preprint arXiv:2503.14499 , year =

    Kwa, Thomas and West, Ben and Becker, Joel and others , title =. arXiv preprint arXiv:2503.14499 , year =. 2503.14499 , archivePrefix=

    arXiv

  21. [29]

    arXiv preprint arXiv:2408.06292 , year =

    Lu, Chris and others , title =. arXiv preprint arXiv:2408.06292 , year =. 2408.06292 , archivePrefix=

    arXiv

  22. [30]

    arXiv preprint arXiv:2501.04227 , year =

    Schmidgall, Samuel and Su, Yusheng and Wang, Ze and others , title =. arXiv preprint arXiv:2501.04227 , year =. 2501.04227 , archivePrefix=

    arXiv

  23. [31]

    arXiv preprint arXiv:2502.13138 , year =

    Jiang, Zhengyao and Schmidt, Dominik and Srikanth, Dhruv and others , title =. arXiv preprint arXiv:2502.13138 , year =. 2502.13138 , archivePrefix=

    arXiv

  24. [32]

    arXiv preprint arXiv:2504.08066 , year =

    Yamada, Yutaro and Lange, Robert Tjarko and Lu, Cong and others , title =. arXiv preprint arXiv:2504.08066 , year =. 2504.08066 , archivePrefix=

    arXiv

  25. [33]

    Nature , volume =

    Romera-Paredes, Bernardino and Barekatain, Mohammadamin and Novikov, Alexander and others , title =. Nature , volume =. 2024 , doi =

    2024

  26. [34]

    arXiv preprint arXiv:2506.13131 , year =

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

    arXiv

  27. [35]

    arXiv preprint arXiv:2606.11926 , year =

    Jin, Jiajie and Hu, Yuyang and Qiu, Kai and Dai, Qi and Luo, Chong and others , title =. arXiv preprint arXiv:2606.11926 , year =. 2606.11926 , archivePrefix=

    arXiv

  28. [36]

    arXiv preprint arXiv:2410.07095 , year =

    Chan, Jun Shern and Chowdhury, Neil and Jaffe, Oliver and others , title =. arXiv preprint arXiv:2410.07095 , year =. 2410.07095 , archivePrefix=

    arXiv

  29. [37]

    arXiv preprint arXiv:2411.15114 , year =

    Wijk, Hjalmar and Lin, Tao and Becker, Joel and others , title =. arXiv preprint arXiv:2411.15114 , year =. 2411.15114 , archivePrefix=

    arXiv

  30. [38]

    arXiv preprint arXiv:2606.07462 , year =

    Wang, Jiayu and Lv, Weijiang and Fu, Bowen and others , title =. arXiv preprint arXiv:2606.07462 , year =. 2606.07462 , archivePrefix=

    arXiv

  31. [39]

    arXiv preprint arXiv:2602.06855 , year =

    Lupidi, Alisia and Gauri, Bhavul and Foster, Thomas Simon and others , title =. arXiv preprint arXiv:2602.06855 , year =. 2602.06855 , archivePrefix=

  32. [40]

    arXiv preprint arXiv:2603.19173 , year =

    Lin, Edward and Modi, Sahil and Hari, Siva Kumar Sastry and others , title =. arXiv preprint arXiv:2603.19173 , year =. 2603.19173 , archivePrefix=

  33. [41]

    2025 , howpublished =

    Karpathy, Andrej , title =. 2025 , howpublished =

    2025

  34. [42]

    2026 , howpublished =

    Karpathy, Andrej , title =. 2026 , howpublished =

    2026

  35. [43]

    2024 , howpublished =

    Jordan, Keller and contributors , title =. 2024 , howpublished =

    2024

  36. [44]

    Argus: A Self-Evolving Research Agent , year =

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.