Pith. sign in

REVIEW 5 major objections 6 minor 22 references

Aleks: AI powered Multi Agent System for Autonomous Scientific Discovery via Data-Driven Approaches in Plant Science

T0 review · 5 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a three-agent AI system, given a raw question and dataset, can autonomously decide how to frame, model, and iterate a data-driven plant-science study, converging on interpretable predictions of grapevine red blotch in

desk verdict A genuinely useful multi-agent system demonstration for plant science, but the validation is too thin to support the autonomy and robustness claims as written. read the letter →

arxiv 2508.19383 v1 pith:AHAJ3X4K submitted 2025-08-26 cs.AI cs.SYeess.SY

Daoyuan Jin , Nick Gunner , Niko Carvajal Janke , Shivranjani Baruah , Kaitlin M. Gold , Yu Jiang This is my paper
classification cs.AIcs.SYeess.SY
keywords multi-agentsystemslargelanguagemodelsautonomousscientificdiscoveryplantdiseasedetectiongrapevineredblotchautomatedmachinelearningfeatureengineeringdomainknowledgeintegration
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

This paper tries to establish that a multi-agent system built on large language models can carry out a data-driven scientific investigation end to end, from a raw question and dataset to a final model and interpretation, with no human intervention in between. The authors build Aleks around three specialized agents—a domain scientist, a data analyst, and a machine learning engineer—who share a complete experiment memory and take turns proposing, implementing, and critiquing modeling steps. In a grapevine red blotch disease case study, Aleks repeatedly converged on interpretable models, selected features that match known epidemiology, and its best models transferred across years. Ablation runs show the domain agent and the full history are necessary for coherent, biologically grounded results, not optional polish. The paper's own audit found one silent coding error in generated metrics, which the authors present as the key open problem for trustworthy autonomous discovery.

What carries the argument

The load-bearing mechanism is an iterative three-agent loop: a domain-scientist agent, configured as a plant pathologist by loading summaries of ten papers into its semantic memory, critiques each modeling suggestion and result for biological relevance; a data-analyst agent, with access to the full experiment history, proposes problem formulations, preprocessing, and feature engineering; and a machine-learning-engineer agent writes executable Python, runs it, and repairs it from error messages, working within the auto-sklearn library. A shared memory stores iteration index, suggestions, results, and domain feedback, with different access privileges per agent—the analyst sees all history, the

What would settle it

Run Aleks on a dataset with known answers while an independent script recomputes every metric from the raw predictions and logs; the central claim of reliable full autonomy collapses if any run's reported accuracy, F1, or R² differs from the independently recomputed value, as happened once in the paper's Exp4 2024 run.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that full autonomy in data-driven scientific discovery is achievable: Aleks, once given a natural-language question and an unprocessed tabular dataset, determines whether the task is classification or regression, selects and derives features, trains models, evaluates results, and iterates until it decides a solution is satisfactory or the budget ends. In the grapevine red blotch case study, the system chose the correct label column, avoided data leakage in the full-history configuration, consistently identified prior-year infection counts, coordinates, and canopy traits as key predictors, and later proposed derived features such as a spatial inf

Load-bearing premise

The loop is trustworthy only if the code the machine-learning engineer writes measures performance without silent errors; the paper's own manual audit caught one run where training and testing metrics were mixed, so every autonomous run's metrics need independent verification before the autonomy claim is established.

Editorial extensions

If this is right

  • Given only a question and a dataset, Aleks can select the modeling framing, engineer and select features, train models, and iterate to a final report without human input, with each full run taking roughly two hours.
  • Removing the domain-scientist agent turns the system into a purely performance-driven feature optimizer that can chase uninformative features or stop early, showing domain feedback is what keeps exploration biologically grounded.
  • Restricting shared memory to the latest iteration makes feature selection inconsistent and allowed one data-leakage event, so complete experiment history is a core requirement, not an enhancement.
  • The best model for 2023 prediction, with only year-specific features shifted forward, achieved an R² of 0.79 on 2024 data, and the reverse transfer reached 0.72, indicating the modeling strategies generalize across years.
  • The paper flags that autonomous code generation can contain silent metric errors—the top 2024 result in one run was invalid—so trustworthy autonomy will require formal validation of agent-generated code.

Reading between the lines

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

  • Beyond the paper's claims, the documented training/test metric mix-up suggests a cheap architectural fix the paper does not implement: an independent auditor agent that recomputes all metrics from raw logs before the domain scientist signs off.
  • Beyond the paper's claims, the cross-year transfer result implies the durable output is the feature-engineering recipe, not the fitted model; a direct test would compare Aleks's recipe against a pure AutoML baseline on the same out-of-year data.
  • Beyond the paper's claims, if the autonomy claim generalizes beyond this tabular 20-iteration setting, the real bottleneck in AI-driven science shifts from doing analyses to verifying them, making formal methods for checking agent-generated code the rate-limiting step.
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

5 major / 6 minor

Summary. This manuscript introduces Aleks, a three-agent LLM system (domain scientist, data analyst, ML engineer) with shared memory, designed to autonomously answer data-driven plant-science questions. In a grapevine red blotch disease case study, Aleks is given a one-line question and a multi-year vineyard dataset; it autonomously chooses classification vs regression, feature engineering, model selection, and stopping criteria. Four experiments evaluate the full system, an ablation without the domain scientist, an ablation without full history, and an ablation with a leaderboard. The paper reports high accuracy/F1 for 2023 and R2 values for 2024, cross-year R2 values around 0.72–0.79, and claims full autonomy and biologically meaningful feature discovery. The authors also transparently document that the top Exp4 2024 result was invalidated by a train/test metric-conflation coding error found only by human re-verification.

Significance. If the performance evidence were robust, this would be a valuable demonstration that a multi-agent LLM system can autonomously frame a data-driven ML question, select features, train models, and iterate to a useful solution in plant science. The architecture choices—three specialized agents, shared memory with scoped access, episodic/semantic memory, and an iterative domain-knowledge feedback loop—are sensible and clearly described. The paper also includes a genuine case study with multiple experiments and an explicit discussion of a known limitation. However, the central quantitative claims are currently supported by fragile evidence: there is no baseline comparison, evaluation metrics are selected by the system after seeing the data, several experiments are single runs, and the only documented human audit of generated evaluation code invalidated the headline result. The paper's own limitations section acknowledges the need for formal validation of AI-generated code, but the current manuscript does not yet deliver that validation.

major comments (5)
  1. [Evaluation Methods and Results (Exp4)] The central quantitative claims rest on metrics computed by code that Aleks itself generates, and the paper explicitly reports that the Exp4 2024 original result (R²=0.9075) was invalid due to a coding error that conflated training and testing metrics, discovered only by human re-verification. The corrected second-best model (R²=0.7584) is comparable to other runs. The statement that 'all reported metrics were manually verified for consistency with the experiment logs' is an assertion, not evidence: no audit protocol, raw logs, or independent recomputation are provided. Because evaluation metrics are chosen autonomously and the final model is selected based on the same data, invalid evaluation code can corrupt both model selection and reported performance. This is load-bearing for the claims of 'robust performance' and reliable autonomy. I request release of the full logs and code, or at
  2. [Aleks Configuration, Results Table 2] No baseline is provided. The MLE agent is explicitly restricted to the auto-sklearn library, so the natural control is a plain auto-sklearn run on the same dataset with default settings and the same features. Without such a baseline, the reported F1/R² values in Table 2 cannot support the claim that the multi-agent system improves on, or even matches, a standard automated ML pipeline. The paper's contribution is the multi-agent autonomy and domain-knowledge integration, so a baseline is necessary to show that the agentic layer adds value (or at least does not degrade performance). At minimum, this should be added as an experiment.
  3. [Model Generalizability Across Years, Table 4] The cross-year validation is presented as evidence of robust generalization, but the 'best model across all experiments' is selected using the same dataset on which the original performance is reported, and the cross-year test is a single run with no uncertainty or confidence interval. The description of how the codebase is adapted ('shifting the corresponding features forward by one year') is insufficiently detailed: it is not clear how missing future features are handled, whether the same train/test separation is used, and whether any label leakage is possible when moving between years. A rigorous protocol with multiple runs, clear feature-year alignment, and per-run results is needed to support the generalizability claim.
  4. [Results Overall Performance, Evaluation Methods] The paper claims 'full autonomy ... without any human interventions through repetitive runs,' but the study involves several human inputs: the research question, the dataset, the selection of domain papers for the DS agent's semantic memory, the research budget (maximum iterations), and manual verification of all reported metrics. The Exp4 bug was also found by human re-verification. The term 'full autonomy' is therefore ambiguous and potentially overstated. The authors should define precisely what is meant by autonomy, list all human interventions that occurred in the reported experiments, and adjust the claims to match that definition. This is not merely a wording issue; it affects how readers interpret the central contribution.
  5. [Table 1, Ablation Comparisons] Exp2, Exp3, and Exp4 are each run once (Sets = 1), while Exp1 uses five repeated runs. The ablation conclusions—that the DS agent is critical, that full memory is important, and that the leaderboard adds no benefit—are based on single runs of each configuration. Given that the system is stochastic (LLM-powered), these differences could be due to run-to-run variability rather than the ablated component. The ablation experiments should be repeated at least as many times as Exp1, and results should be reported as distributions, not single points.
minor comments (6)
  1. [Deploying Aleks for Tissue Sampling Strategy Recommendation] The quoted human question contains a typo: 'redvine disease' should likely be 'red blotch disease' or 'red vein disease' as used elsewhere.
  2. [Table 2] The header 'T otal Runtime' has an extra space. Also, for regression rows, accuracy and weighted F1 columns are empty; this is understandable but should be explicitly indicated as 'not applicable' for clarity.
  3. [Table 3] The row marked 'Exp4 2024 (Original)' is an invalid result due to the documented coding error. Presenting it in the main table without a clear 'invalid' annotation is misleading; it should be moved to a supplementary or explicitly flagged as a failed run.
  4. [System Design and Implementation] Several typos appear: 'accessbility' (should be 'accessibility'), 'epidermiology' (should be 'epidemiology'), and 'T otal' in Table 2. A careful proofread is needed.
  5. [Aleks Configuration] Reproducibility would be improved by specifying the exact LLM version (DeepSeek Chat is vague), temperature and other sampling parameters, retry limits, auto-sklearn version, and the computing environment. These details are currently omitted.
  6. [Figures 5–9] The feature selection heatmaps are difficult to compare because the order and labeling of derived features is not clearly explained in the captions. A unified legend or a table listing all raw and derived features would help.

Circularity Check

1 steps flagged · score 5.0 of 10

Cross-year "prediction" is a same-year refit, but the core autonomous-loop claim is not circular.

  1. fitted input called prediction [Evaluation Methods, paragraph beginning "To address the interests of human scientists"; Results, "Model Generalizability Across Years" and Table 4]
    "the optimal feature set identified by Aleks for the 2023 subset was used to train a model for predicting GRBD infection in 2024, and vice versa. In this way, models were evaluated not only on the year for which they were originally developed but also across years... All other components, including data preprocessing, feature generation functions, and modeling operations, remained unchanged. Using this adapted codebase, the model achieved a similarly high correlation in 2024 (R2 = 0.79) with an RMSE of 0.65."

    The only element transferred from 2023 is the feature-engineering code; the model itself is retrained on the 2024 target-year labels before the 2024 R2 is computed. The paper's own words "train a model for predicting GRBD infection in 2024" and "modeling operations remained unchanged" show that the target-year labels enter the training set. Therefore Table 4's cross-year R2 values are same-year fits obtained by re-running the auto-sklearn pipeline on the year being scored, not out-of-sample predictions from a model trained on the other year. The claimed cross-year generalizability thus reduces, by construction, to a refit with transferred feature engineering.

full rationale

The paper's core contribution is an autonomous multi-agent loop: LLM agents propose problem formulations, select features, generate auto-sklearn code, debug via error messages, and stop on a budget. That process is not a formal derivation whose output is equivalent to its inputs; domain knowledge enters via human-selected paper summaries and auto-sklearn is a fixed configuration choice, not a fitted parameter renamed as a discovery. No load-bearing self-citation or imported uniqueness theorem is present. The one clear reduction-by-construction is the cross-year experiment: the 2024 model is trained on 2024 data after shifting the 2023 feature-engineering code forward, so the reported 2024 R2 is a same-year fit, not evidence of temporal generalization. This overstates the generalizability result but does not make the whole system circular. Separately, the Exp4 train/test metric-conflation bug is a verification concern rather than a circularity: it shows the self-generated evaluation code can be wrong, but it does not by itself make any claim equivalent to its inputs. Overall score 5 reflects one overclaimed 'prediction' while the central autonomy claim retains independent content.

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

The central claim rests on a configured software system and on assumptions about the dataset, the injected literature summaries, and the reliability of auto-sklearn and human verification. No code or data are released, and the system itself is an invented artifact without independent evidence beyond the paper's internal logs. The 20-iteration budget and 60% frequency threshold are hand-chosen analysis parameters.

free parameters (3)
  • Maximum iterations (research budget) = 20
    Set based on a small-scale preliminary test (Aleks Configuration section); determines when runs stop and which final model is selected.
  • Key-feature frequency threshold = 60%
    Used in Results to declare 'frequently used' features (Figure 4); a hand-chosen reporting threshold, not a fitted scientific parameter.
  • Number of domain papers for DS agent memory = 10
    Researchers selected ten papers for the domain scientist agent's semantic memory (Aleks Configuration); this shapes all domain feedback.
assumptions (4)
  • domain assumption The compiled 10 m x 10 m grid dataset contains sufficient signal to predict GRBD infection status from historical counts and canopy traits.
    Underlies the entire case study; if the features are uninformative, no system can succeed. Invoked in the case study section.
  • domain assumption LLM summaries of ten researcher-selected papers provide accurate, non-hallucinated domain knowledge for the DS agent.
    The DS agent's semantic memory is built from these summaries (Aleks Configuration); the paper acknowledges LLM factual accuracy limitations but does not audit the summaries.
  • domain assumption auto-sklearn yields reliable models for tabular data and its API is sufficient for all needed preprocessing and feature engineering.
    MLE agent is restricted to auto-sklearn (Aleks Configuration); this limits tool space but supports the assumption that generated code is simple enough to succeed.
  • domain assumption Human manual verification of reported metrics is complete and correct.
    The paper states 'all reported metrics were manually verified for consistency with the experiment logs'; this manual check is not auditable and missed a bug, by the authors' own admission, before re-verification.
invented entities (1)
  • Aleks (three-agent MAS with shared memory)
    purpose: Autonomous data-driven scientific discovery from a research question and dataset
    The system is the paper's contribution but no code or reproducible artifact is provided, so its behavior cannot be independently validated; the only evidence is the paper's own logs and reported metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Aleks: AI powered Multi Agent System for Autonomous Scientific Discovery via Data-Driven Approaches in Plant Science." pith.science (2026). https://pith.science/paper/AHAJ3X4K

@misc{pith2026250819383,
  author       = {Pith},
  title        = {Pith review of: Aleks: AI powered Multi Agent System for Autonomous Scientific Discovery via Data-Driven Approaches in Plant Science},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AHAJ3X4K}},
  note         = {Machine review of arXiv:2508.19383}
}
read the original abstract

Modern plant science increasingly relies on large, heterogeneous datasets, but challenges in experimental design, data preprocessing, and reproducibility hinder research throughput. Here we introduce Aleks, an AI-powered multi-agent system that integrates domain knowledge, data analysis, and machine learning within a structured framework to autonomously conduct data-driven scientific discovery. Once provided with a research question and dataset, Aleks iteratively formulated problems, explored alternative modeling strategies, and refined solutions across multiple cycles without human intervention. In a case study on grapevine red blotch disease, Aleks progressively identified biologically meaningful features and converged on interpretable models with robust performance. Ablation studies underscored the importance of domain knowledge and memory for coherent outcomes. This exploratory work highlights the promise of agentic AI as an autonomous collaborator for accelerating scientific discovery in plant sciences.

Figures

Figures reproduced from arXiv: 2508.19383 by the authors.

Figure 1
Figure 1. Conceptual framework of Aleks. Human researchers provide a scientific question and an associated dataset to Aleks. Within [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Aleks system design and implementation. In the present study, Aleks consists of three specialized agents: a plant pathologist, a [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Modeling choice change plot for Exp1, 2023 prediction, repetition 1 and 4. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Feature frequency of the final selected models from five repeated experiments for each year’s prediction in Exp1. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Feature selection map for five repeated runs in Exp1 for the 2023 prediction question. The heatmap illustrates the evolution of [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Feature selection map for five repeated runs in Exp1 for the 2024 prediction question. The heatmap illustrates the evolution of [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Feature selection maps in Exp2 for the 2023 (top) and 2024 (bottom) prediction questions, respectively. The heatmap illustrates [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Feature selection maps in Exp3 for the 2023 (top) and 2024 (bottom) prediction questions, respectively. The heatmap illustrates [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Feature selection maps in Exp4 for the 2023 (top) and 2024 (bottom) prediction questions, respectively. The heatmap illustrates [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 20 canonical work pages

  1. [1]

    PhenoAssistant: A Conversational Multi-Agent AI System for Automated Plant Phenotyping

    Feng Chen, Ilias Stogiannidis, Andrew Wood, Danilo Bueno, Dominic Williams, Fraser Macfarlane, Bruce Grieve, Darren Wells, Jonathan A Atkinson, Malcolm J Hawkesford, et al. Phenoassistant: A conversational multi-agent ai system for automated plant phenotyping. arXiv preprint arXiv:2504.19818, 2025

    work page Pith review arXiv 2025

  2. [2]

    Grapevine red blotch disease: A threat to the grape and wine industries

    Elizabeth Cieniewicz and Marc Fuchs. Grapevine red blotch disease: A threat to the grape and wine industries. Annual Review of Virology, 12, 2025

    work page 2025

  3. [3]

    Spatiotemporal spread of grapevine red blotch-associated virus in a california vineyard

    Elizabeth J Cieniewicz, Sarah J Pethybridge, Adrienne Gorny, Laurence V Madden, Heather McLane, Keith L Perry, and Marc Fuchs. Spatiotemporal spread of grapevine red blotch-associated virus in a california vineyard. Virus Research, 241:156–162, 2017

    work page 2017

  4. [4]

    Investigating the latency period of grapevine red blotch virus in a diseased cabernet franc vineyard experiencing secondary spread

    Madison Flasco, Elizabeth J Cieniewicz, Monica L Cooper, Heather McLane, and Marc Fuchs. Investigating the latency period of grapevine red blotch virus in a diseased cabernet franc vineyard experiencing secondary spread. American Journal of Enology and Viticulture, 75(1), 2024

    work page 2024

  5. [5]

    Distinct red blotch disease epidemiological dynamics in two nearby vineyards

    Madison T Flasco, Elizabeth J Cieniewicz, Sarah J Pethybridge, and Marc F Fuchs. Distinct red blotch disease epidemiological dynamics in two nearby vineyards. Viruses, 15(5):1184, 2023

    work page 2023

  6. [6]

    The three-cornered alfalfa hopper, spissistilus festinus, is a vector of grapevine red blotch virus in vineyards

    Madison T Flasco, Victoria Hoyle, Elizabeth J Cieniewicz, Greg Loeb, Heather McLane, Keith Perry, and Marc F Fuchs. The three-cornered alfalfa hopper, spissistilus festinus, is a vector of grapevine red blotch virus in vineyards. Viruses, 15(4):927, 2023

    work page 2023

  7. [7]

    A decade of grapevine red blotch disease epidemiology reveals zonal roguing as novel disease management

    MT Flasco, DW Heck, EJ Cieniewicz, ML Cooper, SJ Pethybridge, and MF Fuchs. A decade of grapevine red blotch disease epidemiology reveals zonal roguing as novel disease management. npj Viruses, 3(1):29, 2025

    work page 2025

  8. [8]

    Scalable early detection of grapevine viral infection with airborne imaging spectroscopy

    Fernando E Romero Galvan, Ryan Pavlick, Graham Trolley, Somil Aggarwal, Daniel Sousa, Charles Starr, Elisabeth Forrestel, Stephanie Bolton, Maria del Mar Alsina, Nick Dokoozlian, et al. Scalable early detection of grapevine viral infection with airborne imaging spectroscopy. Phytopathology®, 113(8):1439–1446, 2023

    work page 2023

Show all 22 references
  1. [9]

    Plant disease sensing: studying plant-pathogen interactions at scale

    Kaitlin M Gold. Plant disease sensing: studying plant-pathogen interactions at scale. Msystems, 6(6):e01228–21, 2021

    2021

  2. [10]

    Michael Jeger, Robert Beresford, Anna Berlin, Clive Bock, Adrian Fox, Kaitlin M Gold, Adrian C Newton, Antonio Vicent, and Xiangming Xu. Impact of novel methods and research approaches in plant pathology: Are individual advances sufficient to meet the wider challenges of disea...

    2024

  3. [11]

    Emerging themes and approaches in plant virus epidemiology

    Mike Jeger, Fred Hamelin, and Nik Cunniffe. Emerging themes and approaches in plant virus epidemiology. Phytopathology®, 113(9):1630–1646, 2023

    2023

  4. [12]

    Grapevine red blotch virus detection in the vineyard: Leveraging machine learning with vis/nir hyperspectral images for asymptomatic and symptomatic vines

    E Laroche-Pinel, K Singh, M Flasco, ML Cooper, M Fuchs, and L Brillante. Grapevine red blotch virus detection in the vineyard: Leveraging machine learning with vis/nir hyperspectral images for asymptomatic and symptomatic vines. Computers and Electronics in Agriculture, 234:11...

    2025

  5. [13]

    The cost of fungicide resistance evolution in multi-field plant epidemics

    Alexey Mikaberidze, Chaitanya S Gokhale, Maria Bargu´ es-Ribera, and Prateek Verma. The cost of fungicide resistance evolution in multi-field plant epidemics. PLOS Sustainability and Transformation, 4(6):e0000178, 2025

    2025

  6. [14]

    Toward cloud-native, machine learning base detection of crop disease with imaging spectroscopy

    Gloire Rubambiza, Fernando Romero Galvan, Ryan Pavlick, Hakim Weatherspoon, and Kaitlin M Gold. Toward cloud-native, machine learning base detection of crop disease with imaging spectroscopy. Journal of Geophysical Research: Biogeosciences, 128(6):e2022JG007342, 2023

    2023

  7. [15]

    Phenotyping grapevine red blotch virus and grapevine leafroll-associated viruses before and after symptom expression through machine-learning analysis of hyperspectral images

    Erica Sawyer, Eve Laroche-Pinel, Madison Flasco, Monica L Cooper, Benjamin Corrales, Marc Fuchs, and Luca Brillante. Phenotyping grapevine red blotch virus and grapevine leafroll-associated viruses before and after symptom expression through machine-learning analysis of hypers...

    2023

  8. [16]

    Agent laboratory: Using llm agents as research assistants

    Samuel Schmidgall, Yusheng Su, Ze Wang, Ximeng Sun, Jialian Wu, Xiaodong Yu, Jiang Liu, Michael Moor, Zicheng Liu, and Emad Barsoum. Agent laboratory: Using llm agents as research assistants. arXiv preprint arXiv:2501.04227, 2025

    2025 arXiv

  9. [17]

    Grapevine red blotch-associated virus, an emerging threat to the grapevine industry

    Mysore R Sudarshana, Keith L Perry, and Marc F Fuchs. Grapevine red blotch-associated virus, an emerging threat to the grapevine industry. Phytopathology, 105(7):1026–1032, 2015

    2015

  10. [18]

    The virtual lab of ai agents designs new sars-cov-2 nanobodies

    Kyle Swanson, Wesley Wu, Nash L Bulaong, John E Pak, and James Zou. The virtual lab of ai agents designs new sars-cov-2 nanobodies. Nature, pages 1–3, 2025

    2025

  11. [19]

    Automl-agent: A multi-agent llm framework for full-pipeline automl

    Patara Trirat, Wonyong Jeong, and Sung Ju Hwang. Automl-agent: A multi-agent llm framework for full-pipeline automl. arXiv preprint arXiv:2410.02958, 2024

    2024 arXiv

  12. [20]

    React: Synergizing reasoning and acting in language models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. International Conference on Learning Representations (ICLR), 2023

    2023

  13. [21]

    Previsual symptoms of xylella fastidiosa infection revealed in spectral plant-trait alterations

    TEJADA Pablo Jesus ZARCO, Carlos CAMINO, Pieter BECK, Rocio CALDERON, Alberto HORNERO, Rocio HERNANDEZ- CLEMENTE, Teja KATTENBORN, Miguel MONTES-BORREGO, Leonardo SUSCA, Massimiliano MORELLI, et al. Previsual symptoms of xylella fastidiosa infection revealed in spectral plant-...

  14. [22]

    Divergent abiotic spectral pathways unravel pathogen stress signals across species

    Pablo J Zarco-Tejada, Tom´ as Poblete, Carlos Camino, Victoria Gonz´ alez-Dugo, Rocio Calderon, Alberto Hornero, Roc ´ ıo Hern´ andez- Clemente, Miguel Rom´ an-´Ecija, Mar ´ ıa Pilar Velasco-Amo, Blanca B Landa, et al. Divergent abiotic spectral pathways unravel pathogen stres...

    2021

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.