Pith. sign in

REVIEW 3 major objections 5 minor 22 references

The paper claims that reliability in LLM-driven data workflows is an architectural property, not a model property: role-separated agents with deterministic validators and audited handoffs detected and blocked a coordinate transformation err

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 16:52 UTC pith:JSJLGN3V

load-bearing objection A useful, honest practice report on architecting LLM agents for irreversible data publication; the ISS-004 incident is real evidence, but the reliability claim is weaker than the abstract suggests because the auditor is the same model and there's no denominator. the 3 major comments →

arxiv 2604.01647 v2 pith:JSJLGN3V submitted 2026-04-02 cs.AI

Exploring Robust Multi-Agent Workflows for Environmental Data Management

Boyuan Guan , Jason Liu , Yanzhao Wu , Kiavash Bahreini This is my paper
classification cs.AI
keywords multi-agent systemsLLM agentsFAIR data managementfail-open failureaudited handoffsknowledge externalizationdeterministic validationenvironmental data
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.

This practice-and-experience paper argues that LLM agents can be embedded in environmental FAIR (findable, accessible, interoperable, reusable) data publication workflows without accepting silent failures, provided reliability is designed into the system structure. The proposed system separates governance constraints, retrievable domain knowledge, and executable skills into three persistent, interlocking artifact tracks, and gives every agent-to-agent transition a deterministic validation gate and an audited handoff record. In the paper's flagship incident, a coordinate transformation mistake produced structurally valid but geographically impossible output for all 2,452 monitored stations; the boundary audit caught it in about 10 minutes, blocked publication, and required independent verification before release. The paper contrasts this multi-agent deployment with an earlier single-agent deployment that needed near-continuous review, and reports that a single operator completed the 8,557-file workflow in about two days. A sympathetic reader would care because the result suggests fail-safe boundaries are a realistic engineering target for imperfect probabilistic components.

Core claim

The central claim is that LLM pipelines fail open—emitting confident, well-structured but subtly wrong outputs that propagate into irreversible actions—and that this can be contained architecturally. The system tracks governance, domain knowledge, and skills separately, and treats every agent-to-agent transition as a trust boundary with deterministic validation and an audited four-phase handoff (prepare, validate, approve, commit). In incident ISS-004, a coordinate field mismatch produced format-valid GeoJSON with all 2,452 stations spatially implausible; the audit caught it in 10 minutes, blocked the commit, held user exposure at zero, and required independent verification before republish

What carries the argument

The audited handoff protocol is the central mechanism: at every agent-to-agent transition, the sending role packages artifacts with provenance, a deterministic validator (e.g., a coordinate range check) runs read-only, results are recorded in an immutable audit trail, and unapproved outputs are quarantined so they cannot become inputs to the next stage or trigger publication. The three-track knowledge architecture—behaviors, domain knowledge, and skills—supplies the standards the validators check against, makes operational rationale durable across personnel changes, and gates skill execution so that 'do not guess' rules are enforced rather than suggested. Together they convert a workflow who

Load-bearing premise

The central claim depends on the two deployments being comparable enough to attribute differences to architecture, and on the system's own audit trail faithfully recording every event; the paper states these were sequential deployments under different workloads, not a controlled A/B comparison.

What would settle it

Search the external publication platform for a file with no matching audit-recorded handoff, or replay the same compound-flooding workflow with the single-agent baseline under matched workload and boundary-only supervision; a single unvalidated published file, or a matched run where per-step review catches errors the multi-agent boundary misses, would break the claim.

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

If this is right

  • If the architecture is sound, LLM-driven FAIR publication can be operated by a single domain expert rather than a team supervising every step: the compound-flooding workload completed in about two days with boundary-only oversight.
  • Errors that are semantically wrong but format-valid can be caught before irreversible steps when deterministic validators encode domain knowledge—demonstrated by the blocked coordinate transformation affecting all 2,452 stations.
  • Operational knowledge externalized in three-track artifacts is transferable: 27 reuse instances were documented, including more than 10 across projects, so onboarding and project handoff need not start from scratch.
  • The architecture can be extended without refactoring: new platform roles were added through a standard tool-integration protocol, supporting incremental adoption.
  • The failure-mode narrative flips from multiplicative decay (each stage's success probability compounding) to multiplicative filtering (independent validation layers reducing error escape), giving a concrete design target for production reliability.

Where Pith is reading between the lines

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

  • The reported 4/4 versus 0/4 detection differential suggests that where validation is placed matters more than the model's raw capability; a fair test would run the same model in a producing role with identical boundary validators but no separated auditing role.
  • The boundary-handoff pattern is not specific to environmental publication: any workflow with irreversible side effects (dataset deletion, instrument commands, code releases) could adopt audited gates, though this extrapolates beyond the paper's evidence.
  • Because deterministic validators only cover checkable properties, the architecture's guarantee is bounded: errors that are semantic in ways no deterministic rule encodes will pass. A testable extension is adding learned semantic plausibility checks at boundaries and measuring how detection changes.
  • If the efficiency result transfers, onboarding a new operator could shrink from weeks to minutes-scale, since the three-track store carries rationale and procedure; this could be tested by a handoff experiment with a fresh operator.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper presents EnviSmart, a production multi-agent system for environmental FAIR data management that embeds LLM agents in a three-track knowledge architecture (behaviors, domain knowledge, skills) and a role-separated operating model with deterministic validators and audited handoffs at trust boundaries. It reports two sequential production deployments: a single-agent baseline (GIS Center Ecological Archive, 849 datasets) and a multi-agent deployment (SF2Bench, 2,452 monitoring-station datasets packaged into 8,557 files). The central claim is that the multi-agent approach improved both efficiency (weeks to about two days; single operator; artifact reuse) and reliability (ISS-004 coordinate error affecting all 2,452 stations detected in about 10 minutes and blocked before publication; 4/4 audit detections vs 0/4 self-detections). The paper is an explicit practice/experience report, disclaims controlled comparison, and labels its p^n and (1-q)^k models as illustrative.

Significance. If the results hold, EnviSmart offers a valuable architectural template for risk-managed integration of LLM agents into irreversible research data-publication workflows. Its concrete contributions are the three-track artifact store, least-privilege role separation, boundary-anchored validation, and an auditable incident chain (ISS-004 to ISS-005 to ISS-009) that demonstrates an error being contained before publication. Strengths include the explicit distinction between illustrative reliability models and auditable incident records, the reproducibility of the deterministic validator logic, and the paper's transparency about the absence of controlled comparison. The report is most plausibly read as an existence proof and operational lesson set; the comparative claims in the abstract and conclusion go beyond what the evidence can support.

major comments (3)
  1. [§4.3/§4.5, Table 2] The reliability half of the abstract's claim rests on a numerator without a denominator. Table 2 reports '0/4 self-detections vs. 4/4 audit detections' and §4.5 concludes that 'the producing role missed all four errors... while the same model in an auditing role... caught all four.' The paper does not report the total number of handoffs, validation calls, or error/near-miss events in the SF2Bench run, so 4/4 cannot be separated from selective reporting of the only incidents that produced issue IDs. Moreover, the detecting layer is not attributed: §4.3 only says publication was 'blocked at the boundary,' while Table 2 credits 'audit detections.' Since §4.3 states that preparing and auditing roles used the same model family (Claude Sonnet 4.5), nothing in the evidence rules out that ISS-004 was caught by a second LLM prompt rather than by a deterministic validator from §3.2. This attributi
  2. [Tables 1 and 2; abstract] The efficiency claim is based on two sequential deployments with different workloads, team sizes, and curation units. Case 1 processed 849 heterogeneous datasets (aerial photography, GIS layers, sensor time series, field surveys) over weeks with one developer and two domain collaborators; Case 2 processed 2,452 monitoring-station datasets into 8,557 files in about two days with one researcher. Table 2 explicitly disclaims a controlled A/B comparison, yet the abstract states without qualification that the multi-agent approach 'improved both efficiency' and the conclusion repeats 'improved scalability.' To make this claim defensible, the paper should either provide a normalized throughput measure (e.g., files or artifacts per person-day, or number of human decision points per completed dataset) or explicitly rephrase the claim as 'a larger workload was completed in less wall-clock time and
  3. [§3.2; §4.3] The phrase 'deterministic validators and audited handoffs restore fail-stop semantics' overstates the mechanism. The layered model in §3.2 explicitly includes an 'independent agent review' layer and models error pass-through as (1−q)^k, which is a stochastic filter, not a fail-stop gate. ISS-004 was a 'structurally valid but semantically wrong' coordinate transformation; the deterministic validator described in §3.2 (latitude/longitude ranges for Florida) would not necessarily catch clustered but in-range stations. Therefore the paper cannot claim that the boundary is fail-stop unless the detecting layer is shown to be deterministic. If some boundaries are deterministic and others probabilistic, the guarantee should be scoped accordingly.
minor comments (5)
  1. [§4.3; Table 2] Please define 'same-model audit differential' precisely. §4.3 says preparing and auditing roles used the same model family (Claude Sonnet 4.5 with extended thinking), but Table 2's 'same-model' wording suggests identical model instance/parameters. Clarify whether 'same' means same family, same version, or same inference configuration, and specify whether context differed between producing and auditing roles.
  2. [§3.2] The names 'Envita,' 'Stori,' and 'DIVA' appear without introduction. If these are agent or server role names, define them at first use; if 'Envita' is a typo for 'EnviSmart,' correct it.
  3. [§6] 'The control database uses standard SQL' is unclear. This appears to be a typo for 'core database' or 'central database'; otherwise explain what 'control database' refers to.
  4. [Table 1] The E3 row lists 'N/A (no MCP discovery)' for Case 1. Clarify whether MCP discovery was absent by design, not implemented, or simply not measured in the baseline deployment.
  5. [§4.2] The counts of '16 broken skill→behavior references and ∼20 missing knowledge→skill links' would benefit from a short methodological note on how engineering transcripts were coded to produce these numbers.

Circularity Check

0 steps flagged

No circular derivation found; the paper's illustrative formulas are explicit disclaimers and its self-citations are not load-bearing.

full rationale

The claimed derivation chain is an experience report, not a mathematical derivation. The only quantitative models, p^n and (1-q)^k, are explicitly marked as illustrative: 'As with the p^n model in Section 2, this formula is illustrative; our evaluation relies on auditable incident records.' No parameter is fitted to data and then renamed as a prediction. The central reliability evidence comes from the system's own audit trail, and Table 2 explicitly disclaims a controlled comparison: 'Cases reflect sequential deployments under different workloads, not a controlled A/B comparison; differences are attributed to architectural properties.' This is an evidentiary limitation—there is no independent ground truth or denominator for the 0/4 vs 4/4 same-model audit differential—but it is not a circular reduction: the paper does not define reliability in terms of the very incidents it uses to prove reliability, nor does any equation make the conclusion identical to an input. The self-citations, [Poudel et al. 2025] and [Zheng et al. 2025], supply background and the SF2Bench dataset, respectively, but the architectural claims are not justified solely by those citations. The paper also disclaims strong guarantees: 'Note that this paper does not claim provable guarantees or fully autonomous FAIR data management.' Accordingly, no step in the paper reduces to its own inputs by construction, and the appropriate circularity score is 0.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The central claims depend mainly on design assumptions and the fidelity of the self-reported audit trail, not on fitted free parameters. The two illustrative probability models (p, q) are explicitly non-evaluative. No new physical or ontological entities are introduced. The load-bearing axioms are comparability of the two deployments and trustworthiness of the system's own logs.

free parameters (2)
  • p (per-stage success probability in p^n model)
    Introduced in Section 2 as an illustrative model of error propagation; not fitted to data, not used in evaluation. Hand-chosen to motivate architecture.
  • q (per-layer error-catch probability in (1-q)^k model)
    Introduced in Section 3.2 as an illustrative model of error filtering; explicitly said to be illustrative; not fitted or measured.
axioms (4)
  • domain assumption Stage failures are independent in the p^n reliability model.
    Section 2: 'If each stage succeeds with probability p, an n-stage workflow yields p^n end-to-end reliability.' Independence of errors across stages is assumed, not demonstrated.
  • domain assumption Validation layers catch errors independently in the (1-q)^k model.
    Section 3.2: 'If each layer catches errors with probability q, pass-through after k layers is (1-q)^k.' The paper notes the formula is illustrative, but the independence assumption is load-bearing for the architectural motivation.
  • domain assumption The audit trail accurately records handoffs, validation outcomes, and incident metrics.
    Section 4.3 and ISS-004 rely on the system's internal logs for detection latency, user exposure, and blocked commits. No external audit or released logs verify these records.
  • domain assumption Differences between the GIS single-agent baseline and SF2Bench multi-agent deployment are attributable to architectural properties.
    Table 2 states 'Cases reflect sequential deployments under different workloads, not a controlled A/B comparison; differences are attributed to architectural properties.' This attribution is assumed, not established.

pith-pipeline@v1.3.0-alltime-deepseek · 9786 in / 7675 out tokens · 80588 ms · 2026-08-02T16:52:57.215969+00:00 · methodology

0 comments
read the original abstract

Embedding LLM-driven agents into environmental FAIR data management is compelling - they can externalize operational knowledge and scale curation across heterogeneous data and evolving conventions. However, replacing deterministic components with probabilistic workflows changes the failure mode: LLM pipelines may generate plausible but incorrect outputs that pass superficial checks and propagate into irreversible actions such as DOI minting and public release. We introduce EnviSmart, a production data management system deployed on campus-wide storage infrastructure for environmental research. EnviSmart treats reliability as an architectural property through two mechanisms: a three-track knowledge architecture that externalizes behaviors (governance constraints), domain knowledge (retrievable context), and skills (tool-using procedures) as persistent, interlocking artifacts; and a role-separated multi-agent design where deterministic validators and audited handoffs restore fail-stop semantics at trust boundaries before irreversible steps. We compare two production deployments. The University's GIS Center Ecological Archive (849 curated datasets) serves as a single-agent baseline. SF2Bench, a compound flooding benchmark comprising 2,452 monitoring stations and 8,557 published files spanning 39 years, validates the multi-agent workflow. The multi-agent approach improved both efficiency - completed by a single operator in two days with repeated artifact reuse across deployments - and reliability: audited handoffs detected and blocked a coordinate transformation error affecting all 2,452 stations before publication. A representative incident (ISS-004) demonstrated boundary-based containment with 10-minute detection latency, zero user exposure, and 80-minute resolution. This paper has been accepted at PEARC 2026.

Figures

Figures reproduced from arXiv: 2604.01647 by Boyuan Guan, Jason Liu, Kiavash Bahreini, Yanzhao Wu.

Figure 1
Figure 1. Figure 1: The “fail-open” problem under composition. LLM agents can succeed in bounded, reversible tasks (left), but reliability degrades across end-to-end publication workflows with irreversible steps (center). Core LLM limitations compound under composition (right), yielding multiplicative reliability decay 𝑝 𝑛 (bottom). At production scale, silent failures create irreversible contamination and trust collapse. as … view at source ↗
Figure 2
Figure 2. Figure 2: System design overview. Top: Three-Track artifact store externalizing governance (Track 1), semantic context (Track 2), and executable procedures (Track 3) that interlock at execution time. Bottom: multi-agent operating model with role-separated agents (data preparation, publishing, platform operations) connected by audited handoffs (H1/H2) with deterministic validators before state-changing publication. p… view at source ↗
Figure 3
Figure 3. Figure 3: Three-Track artifact graph and an enforceable “do-not-guess” gate. Left: the full artifact graph with typed nodes for Behaviors (Track 1), Knowledge (Track 2), and Skills (Track 3) used during production. Right: zoomed interlock between a metadata￾creation workflow (Track 3) and an authoritative mapping source behavior (Track 1). The behavior enforces that when filenames do not align, the workflow must obt… view at source ↗
Figure 4
Figure 4. Figure 4: Audited handoff protocol. Each agent-to-agent transition follows four phases: prepare (package outputs with provenance), validate (deterministic gates), approve (record and escalate on failure), and commit (apply state changes). Failures block downstream execution. Inset: production incident ISS-004, from boundary detection to verified resolution. long-horizon governance remains under-measured [Wei et al. … view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

22 extracted references · 4 canonical work pages · 1 internal anchor

  1. [5]

    Sayash Kapoor, Benedikt Stroebl, Zachary S

    ACON: Optimizing Context Compression for Long-horizon LLM Agents.arXiv preprint arXiv:2510.00615(2025). Sayash Kapoor, Benedikt Stroebl, Zachary S. Siber, Michele Catasta, and Arvind Narayanan

    Pith/arXiv arXiv 2025

  2. [6]

    arXiv:2407.01502 https://arxiv.org/abs/2407.01502 Gary King

    AI Agents That Matter.Transactions on Machine Learning Research(2025). arXiv:2407.01502 https://arxiv.org/abs/2407.01502 Gary King

    Pith/arXiv arXiv 2025

  3. [10]

    doi:10.1162/tacl_a_00638 Microsoft Research

    Lost in the Middle: How Language Models Use Long Contexts.Transactions of the Association for Computational Linguistics(2023). doi:10.1162/tacl_a_00638 Microsoft Research

    doi:10.1162/tacl_a_00638 2023

  4. [11]

    https://modelcontextprotocol.io/specification/2025-11-25

    Model Context Protocol: Specification (Version 2025-11-25). https://modelcontextprotocol.io/specification/2025-11-25. Accessed 2026-01-31. OpenAI

    2025

  5. [12]

    Melissa Z

    GPT-4 Technical Report.arXiv preprint arXiv:2303.08774(2023). Melissa Z. Pan, Negar Arabzadeh, Riccardo Cogo, Yuxuan Zhu, Alexander Xiong, et al

    Pith/arXiv arXiv 2023

  6. [13]

    Pratik Poudel, Boyuan Guan, Andres Lopez, Nicole Sanchez, Kiavash Bahreini, Wencong Cui, Zhaohui Fu, Hamed Najafi, Leonardo Bobadilla, and Jason Liu

    Measuring Agents in Production.arXiv preprint arXiv:2512.04123(2025). Pratik Poudel, Boyuan Guan, Andres Lopez, Nicole Sanchez, Kiavash Bahreini, Wencong Cui, Zhaohui Fu, Hamed Najafi, Leonardo Bobadilla, and Jason Liu

    Pith/arXiv arXiv 2025

  7. [14]

    InPractice and Experience in Advanced Research Computing (PEARC ’25)(Columbus, OH, USA)(PEARC ’25)

    A Scalable Framework for Heterogeneous Environmental Data Management Using Smart Data Pipeline. InPractice and Experience in Advanced Research Computing (PEARC ’25)(Columbus, OH, USA)(PEARC ’25). Association for Computing Machinery, New York, NY, USA, 9 pages. doi:10.1145/3708035.3736017 Scott Rose, Oliver Borchert, Stu Mitchell, and Sean Connelly. 2020.Z...

    arXiv 2020

  8. [16]

    doi:10.1016/j.future.2024.03.037 D

    Validity constraints for data analysis workflows.Future Generation Computer Systems157 (2024), 82–97. doi:10.1016/j.future.2024.03.037 D. Sculley, Gary Holt, Daniel Golovin, Eugene Davydov, Todd Phillips, Dietmar Ebner, Vinay Chaudhary, Michael Young, Jean-François Crespo, and Dan Dennison

    work page doi:10.1016/j.future.2024.03.037 2024

  9. [18]

    Tianxin Wei, Ting-Wei Li, Zhining Liu, Xuying Ning, Ze Yang, Jiaru Zou, et al

    Failure Modes in LLM Systems: A System-Level Taxonomy for Reliable AI Applications.arXiv preprint arXiv:2511.19933(2025). Tianxin Wei, Ting-Wei Li, Zhining Liu, Xuying Ning, Ze Yang, Jiaru Zou, et al

    arXiv 2025

  10. [19]

    arXiv:2601.12538 [cs.AI] https://arxiv.org/abs/2601.12538 Derek Weitzel et al

    Agentic Reasoning for Large Language Models. arXiv:2601.12538 [cs.AI] https://arxiv.org/abs/2601.12538 Derek Weitzel et al

    Pith/arXiv arXiv

  11. [20]

    doi:10.1145/3626203.3670557 Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, et al

    ACM. doi:10.1145/3626203.3670557 Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, et al

    arXiv

  12. [21]

    Xu Zheng, Chaohao Lin, Sipeng Chen, Zhuomin Chen, Jimeng Shi, Wei Cheng, Jayantha Obeysekera, Jason Liu, and Dongsheng Luo

    AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation.arXiv preprint arXiv:2308.08155(2023). Xu Zheng, Chaohao Lin, Sipeng Chen, Zhuomin Chen, Jimeng Shi, Wei Cheng, Jayantha Obeysekera, Jason Liu, and Dongsheng Luo

    Pith/arXiv arXiv 2023

  13. [22]

    Uncovering Insights of Compound Flooding with Data-Driven AI

    SF2Bench: Evaluating Data-Driven Models for Compound Flood Forecasting in South Florida.arXiv preprint arXiv:2506.04281(2025). doi:10.48550/arXiv.2506.04281 Manuscript submitted to ACM

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2506.04281 2025

  14. [1975]

    IEEE63, 9 (1975), 1278–1308

    The Protection of Information in Computer Systems.Proc. IEEE63, 9 (1975), 1278–1308. doi:10.1109/ PROC.1975.9939 Florian Schintke, Khalid Belhajjame, Ninon De Mecquenem, David Frantz, Vanessa Emanuela Guarino, Marcus Hilbrich, Fabian Lehmann, Paolo Missier, Rebecca Sattler, Jan Arne Sparka, Daniel T. Speckhard, Hermann Stolte, Anh Duc Vu, and Ulf Leser

    arXiv 1975

  15. [2007]

    doi:10.1177/0049124107306660 LangChain, Inc

    An Introduction to the Dataverse Network as an Infrastructure for Data Sharing.Sociological Methods & Research36, 2 (2007), 173–199. doi:10.1177/0049124107306660 LangChain, Inc

    work page doi:10.1177/0049124107306660 2007

  16. [2011]

    doi:10.1109/MIC.2011.64 Haoyu Han, Yu Wang, Harry Shomer, Kai Guo, Jiayuan Ding, Yongjia Lei, et al

    Globus Online: Accelerating and Democratizing Science through Cloud-Based Services.IEEE Internet Computing15, 3 (2011), 70–73. doi:10.1109/MIC.2011.64 Haoyu Han, Yu Wang, Harry Shomer, Kai Guo, Jiayuan Ding, Yongjia Lei, et al

    work page doi:10.1109/mic.2011.64 2011

  17. [2016]

    Manuscript submitted to ACM 12 Guan et al

    Concrete Problems in AI Safety.arXiv preprint arXiv:1606.06565(2016). Manuscript submitted to ACM 12 Guan et al. Ian Foster

    Pith/arXiv arXiv 2016

  18. [2020]

    doi:10.1080/10447318.2020.1741118 Vaishali Vinay

    Human-Centered Artificial Intelligence: Reliable, Safe & Trustworthy.International Journal of Human-Computer Interaction36, 6 (2020), 495–504. doi:10.1080/10447318.2020.1741118 Vaishali Vinay

    arXiv 2020

  19. [2023]

    Minki Kang, Wei-Ning Chen, Dongge Han, Huseyin A

    MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework.arXiv preprint arXiv:2308.00352(2023). Minki Kang, Wei-Ning Chen, Dongge Han, Huseyin A. Inan, Lukas Wutschitz, Yanzhi Chen, Robert Sim, and Saravan Rajmohan

    Pith/arXiv arXiv 2023

  20. [2024]

    arXiv:2407.16833 [cs.CL] https://arxiv.org/abs/2407.16833 Jiaqi Liu, Yaofeng Su, Peng Xia, Siwei Han, Zeyu Zheng, Cihang Xie, Mingyu Ding, and Huaxiu Yao

    Retrieval Augmented Generation or Long-Context LLMs? A Comprehensive Study and Hybrid Approach. arXiv:2407.16833 [cs.CL] https://arxiv.org/abs/2407.16833 Jiaqi Liu, Yaofeng Su, Peng Xia, Siwei Han, Zeyu Zheng, Cihang Xie, Mingyu Ding, and Huaxiu Yao

    Pith/arXiv arXiv

  21. [2025]

    Sirui Hong, Mingchen Zhuge, Jiaqi Chen, Xiawu Zheng, Yuheng Cheng, Ceyao Zhang, Jinlin Wang, et al

    Retrieval-Augmented Generation with Graphs (GraphRAG).arXiv preprint arXiv:2501.00309(2025). Sirui Hong, Mingchen Zhuge, Jiaqi Chen, Xiawu Zheng, Yuheng Cheng, Ceyao Zhang, Jinlin Wang, et al

    Pith/arXiv arXiv 2025

  22. [2026]

    Nelson F

    SimpleMem: Efficient Lifelong Memory for LLM Agents.arXiv preprint arXiv:2601.02553(2026). Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang

    Pith/arXiv arXiv 2026