Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

BioMARS: A Multi-Agent Robotic System for Autonomous Biological Experiments

T0 review · 4 major / 6 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 BioMARS shows that a three-agent LLM/VLM system can autonomously design, execute, and monitor cell culture protocols on a dual-arm robot, matching manual outcomes in viability, consistency, and morphology.

desk verdict A credible multi-agent robotics integration with real passaging data, but the RPE differentiation claim rests on interpolated surrogate scores, not measured cells. read the letter →

arxiv 2507.01485 v1 pith:2YDRVQUN submitted 2025-07-02 cs.RO cs.AIcs.MAq-bio.QM

Yibo Qiu , Zan Huang , Zhiyu Wang , Handi Liu , Yiling Qiao , Yifeng Hu , Shu'ang Sun , Hangke Peng
show 2 more authors
Ronald X Xu Mingzhai Sun
This is my paper
classification cs.ROcs.AIcs.MAq-bio.QM
keywords multi-agentroboticslaboratoryautomationlargelanguagemodelsvision-languageretrieval-augmentedgenerationcellcultureanomalydetectioniPSC-RPEdifferentiation
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

BioMARS is an attempt to close the loop between an LLM/VLM reading biology papers and a dual-arm robot doing the wet work. The paper claims that a three-agent hierarchy—a Biologist that designs protocols from retrieved literature, a Technician that turns them into robotic pseudo-code, and an Inspector that watches for errors—can run cell passaging and culture tasks with results no worse than a human technician. In side-by-side tests on HeLa, Y79, and DC2.4 cells, automated passaging matched manual viability and morphology while shaving hands-on time from about 60 minutes to 5–8 minutes per cell line. The same agent is then shown using historical data to tune seven iPSC-RPE differentiation parameters, outperforming conventional Bayesian optimization in a 20-iteration benchmark. If these results hold, general-purpose language-driven laboratory automation becomes a realistic near-term option rather than a vision.

What carries the argument

The load-bearing mechanism is the tri-agent decomposition with strict separation of cognitive roles. The Biologist Agent is an agentic retrieval-augmented generation pipeline whose sub-modules—Knowledge Checker, Workflow Generator, Workflow Checker—enforce domain validity and laboratory constraints before a protocol is produced; the Technician Agent compiles that protocol into validated pseudo-code against an 11-function robotic API via a CodeGenerator and a rule-based CodeChecker; the Inspector Agent closes the loop with two-stage perception, using a ViT keyframe library for fast geometric anomaly detection and a VLM for semantic confirmation. The paper's evaluation suggests each layer is needed: removing the Workflow Checker or Knowledge Checker drops baseline scores below 3/5, and adding CodeChecker raises instruction matching from 92.4% to 96.4%.

What would settle it

Run the identical 20-iteration comparison on physical iPSC-RPE cultures: for each parameter set proposed by DeepSeek-R1, GPT-4o, and Bayesian optimization, measure actual pigment score after differentiation. If the LLM-chosen sets do not beat the Bayesian sets on measured rather than interpolated pigment, the optimization claim fails.

Watch

Extended reading notes

Core claim

On its own terms, BioMARS's central discovery is that a hierarchical multi-agent LLM/VLM pipeline can carry a biological experiment from literature to liquid handling without human scripting of each step. The Biologist Agent's retrieval-augmented generation, with a Knowledge Checker filtering domain-inconsistent content and a Workflow Checker refining logical structure, turns free-text queries into executable protocols, with the best configuration scoring 5/5 on a 70-query benchmark. The Technician Agent's CodeGenerator plus CodeChecker pair translates those protocols into calls from an 11-function robot API, reaching 96.4% instruction-matching accuracy by inserting implicit operations and correcting parameter errors. The Inspector Agent adds a ViT keyframe library for fast detection and a VLM for zero-shot semantic validation, cutting false-positive anomaly alerts from 8.7% to 1.4%. In physical experiments, automated passaging was statistically indistinguishable from manual passaging in CCK-8 viability and live/dead staining while showing 12–18% lower variability in HeLa and Y79. In the optimization benchmark, DeepSeek-R1 with prior data reached a predicted pigment score of 0.5913 versus 0.4344 for GPT-4o and 0.3130 for Bayesian optimization.

Load-bearing premise

The optimization claim rests on an offline stand-in: the agent picks differentiation parameters by interpolating pigment scores from a published dataset, and if those interpolated scores diverge from what real cultures produce, the reported lead over Bayesian optimization will not transfer to actual cells.

Editorial extensions

If this is right

  • The same Biologist–Technician–Inspector split should extend to other cell types and protocol classes without rewriting the robot commands.
  • The 90% reduction in hands-on passaging time means a single operator can supervise several BioMARS stations, multiplying throughput per lab.
  • Cutting false-positive anomaly alerts from 8.7% to 1.4% lowers the chance of unnecessary halts, making extended unattended operation practical.
  • When historical data are sparse, LLMs with biological priors can start closer to promising regions than Bayesian optimization, which stays near baseline in the no-prior setting.
  • A modular 11-function API plus a web interface means adding a new instrument or protocol does not require new agent-level code.

Reading between the lines

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

  • Beyond the paper's claims, the size of the optimization margin is an open question: the comparison uses a KDTree-interpolated pigment-score oracle, not physical differentiation, so the 15–28 percentage-point lead is an upper bound on what a user should expect in the lab.
  • A natural next test would feed real assay readouts back into the Biologist Agent each round, converting the offline optimizer into a closed-loop experimenter and measuring whether the interpolation artifacts disappear.
  • The Inspector's dependence on manually refined bounding boxes and a finite error library suggests the system's robustness is bounded by coverage of known failure modes; measuring performance on novel, unlisted errors would stress-test the autonomy claim.
  • The 90% hands-on-time reduction supports the paper's labor-cost argument, but the economic benefit scales only if supervisory load does not grow with more parallel stations—something the current single-pipeline evaluation does not address.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper presents BioMARS, a hierarchical multi-agent system in which a Biologist Agent retrieves and synthesizes wet-lab protocols, a Technician Agent translates them into robotic pseudo-code, and an Inspector Agent performs VLM/ViT-based anomaly detection. The system is evaluated on protocol-generation quality across seven cell lines, instruction-translation accuracy, error-detection latency and F1, and real cell-passaging experiments on HeLa, Y79, and DC2.4 cells. The final section claims a context-aware optimization capability for iPSC-RPE differentiation, comparing LLM-driven parameter selection with Bayesian optimization.

Significance. If the execution claims hold, BioMARS is a substantial systems contribution: it couples LLM/VLM planning with a dual-arm robotic platform, reports concrete translation accuracy (96.4%), low-latency ViT-based anomaly detection (0.3066 s mean latency), and wet-lab passaging comparisons across three cell lines. The optimization claim, however, is not established as stated: the reported differentiation superiority rests entirely on a KDTree-interpolated surrogate over a public dataset, not on measured differentiation outcomes. The passaging results provide partial support for the robotic-execution claim, but the abstract's headline 'outperforming conventional strategies in differentiating retinal pigment epithelial cells' currently overstates what the evidence supports.

major comments (4)
  1. [Section 2.6] The central optimization claim is not established by the experiments reported. Parameter selection uses KDTree-based nearest-neighbor interpolation over the Kanda et al. dataset, and outcomes are reported as interpolated pigment scores rather than measured differentiation results. In a seven-dimensional parameter space with only a sparse set of historical conditions, a nearest-neighbor surrogate can be a poor oracle, and an optimizer may exploit interpolation artifacts or simply navigate toward known high-scoring neighbors. The manuscript reports no held-out validation, no cross-validation, no error bars on the interpolated scores, and no wet-lab confirmation of the recommended conditions. As written, the abstract's statement that BioMARS 'outperforming conventional strategies in differentiating retinal pigment epithelial cells' conflates surrogate predictions with biological outcomes.
  2. [Section 2.5] The claim that automated passaging matches manual performance is statistically underspecified. The text states that 'OD measurements showed no significant difference' without reporting the number of replicates n, the statistical test used, the test statistic, or the p-value. The 12-18% lower coefficient of variation for HeLa and Y79 is reported without confidence intervals or a dispersion test. The 'over 92% concordance' from live/dead staining is also undefined. These missing details are load-bearing for the abstract's claim that BioMARS matches or exceeds manual performance in viability, consistency, and morphological integrity.
  3. [Section 2.6] The reported optimization numbers are internally inconsistent. The text states that DeepSeek-R1 reached a final pigment score of 0.5913, surpassing GPT-4o (0.4344, +15.8%) and Bayesian optimization (0.3130, +28.5%). As percentages, these differences are +36.1% and +88.9%, respectively, not +15.8% and +28.5%; the numbers could make sense as percentage-point differences, but the '%' notation is misleading. In addition, the same paragraph says that by iteration 7 DeepSeek-R1 'achieved 0.6252 and continued steady improvement,' which contradicts a final score of 0.5913 if the score is to be maximized. The authors should clarify the iteration schedule, the direction of the objective, and report the full trajectory with errors.
  4. [Section 2.4] The Inspector evaluation is partly human-in-the-loop, but the manuscript presents it as an autonomous capability. The bounding boxes generated by the VLM are 'manually refined' before cropped subregions are processed by the ViT. This manual step is not quantified in terms of effort, selection criteria, or potential bias, and it undermines the claim of fully autonomous multimodal perception. The authors should either measure the impact of manual refinement on the reported F1/precision numbers or demonstrate that the pipeline operates without this intervention.
minor comments (6)
  1. [Figure 4] The caption of Fig. 4a says 'Workflow diagram of the Technician Agent,' but the figure illustrates the Inspector Agent; the caption should be corrected.
  2. [Equation (1)] The similarity function φsim is used without definition; the authors should specify the distance or similarity metric used for embedding comparison.
  3. [Section 2.5] The phrase 'over 92% concordance' between automated and manual groups should be replaced with a precisely defined metric, such as percentage of live cells per condition or a statistical agreement measure.
  4. [Section 2.6] The difference between the 'prior-informed' and 'no-prior' settings is described only in figure captions; the method should state explicitly how the 10 prior data points are incorporated in each setting.
  5. [Section 4.4] The α-quantile threshold for keyframe anomaly detection is mentioned in Eq. (1) but the value of α and the procedure for selecting it are not reported; this parameter should be documented for reproducibility.
  6. [Appendix B.1] The protocol-generation evaluation uses human expert scoring on a 5-point scale, but no inter-rater reliability, blinding, or number of independent raters is reported; adding these details would strengthen the comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: all load-bearing results are evaluated against external benchmarks or wet-lab assays, and the surrogate-based optimization comparison is not forced by construction.

full rationale

BioMARS's claimed derivation chain is self-contained against external evidence. Protocol synthesis is scored on a human 5-point rubric (Section 2.2); protocol-to-code translation is measured by instruction-matching accuracy against a fixed interface specification (Section 2.3); error detection is benchmarked with precision/recall on predefined error scenarios (Section 2.4); and the passaging claim is supported by CCK-8 viability, live/dead staining, morphology, and CV analysis comparing automated versus manual execution (Section 2.5). None of these outputs is defined in terms of a parameter the system fitted to that same output. The iPSC-RPE optimization section (Section 2.6) indeed evaluates parameter choices through KDTree nearest-neighbor interpolation over a public dataset rather than through wet-lab differentiation, so the abstract's wording 'outperforming conventional strategies in differentiating retinal pigment epithelial cells' is an overstatement of the evidence; Section 3 partially concedes this by noting limited adaptive parameter tuning. However, this is a validity and external-transfer concern, not circularity: all compared strategies are scored by the same external surrogate, the LLM proposals are not fitted to that surrogate's training labels, and the margin over Bayesian optimization is an empirical outcome of the benchmark rather than an identity or a renamed input. No self-citation chain is load-bearing; references [20], [26], and [29] are external benchmarks, datasets, or algorithms.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

BioMARS is an integration paper, so the ledger records hand-set system parameters and surrogate assumptions rather than fitted physical constants. The results depend on several human-chosen parameters (thresholds, keyframe library, retrieval counts) and on the fidelity of an interpolated optimization oracle.

free parameters (3)
  • α-quantile threshold for keyframe anomaly detection = not stated (adaptively chosen per task)
    Equation 1 compares frame embeddings against a reference library and selects thresholds by the α-quantile of descending scores; α is a hand-set sensitivity parameter whose value is not reported.
  • 23-action keyframe library = 23 actions from 11 control primitives
    The Inspector's ViT encodes a manually selected set of visually discriminative actions; this selection determines which anomalies are detectable and is not derived from data.
  • retrieval and RAG hyperparameters = 3 PDFs, 3 snippets, top-5 chunks
    Biologist Agent retrieval amounts are hard-coded choices in Section 2.2 that influence protocol quality; no ablation beyond module presence is reported.
assumptions (4)
  • ad hoc to paper KDTree nearest-neighbor interpolation over the Kanda et al. dataset is a valid surrogate for real iPSC-RPE differentiation pigment scores.
    Section 2.6 treats interpolation as the oracle for 20-iteration optimization; no validation against held-out wet-lab experiments is given.
  • domain assumption The human 5-point protocol scoring (following Boiko et al.) reflects biological executability.
    Section 2.2 relies on this scoring to claim DeepSeek-R1+WC+KC achieves scores of 5; scorers are not described and inter-rater agreement is not reported.
  • domain assumption Manual passaging in Section 2.5 is an expert-level baseline with comparable care to BioMARS execution.
    The 'no significant difference' and CV claims depend on the manual arm being a fair, stable comparator; replicate counts and operator skill are not stated.
  • ad hoc to paper Manual refinement of VLM bounding boxes (Section 2.4) does not systematically bias the Inspector evaluation.
    The pipeline shows 'bounding boxes generated by the VLM are manually refined' before grayscale embedding; human refits may encode the authors' expectations about error scenarios.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BioMARS: A Multi-Agent Robotic System for Autonomous Biological Experiments." pith.science (2026). https://pith.science/paper/2YDRVQUN

@misc{pith2026250701485,
  author       = {Pith},
  title        = {Pith review of: BioMARS: A Multi-Agent Robotic System for Autonomous Biological Experiments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2YDRVQUN}},
  note         = {Machine review of arXiv:2507.01485}
}
read the original abstract

Large language models (LLMs) and vision-language models (VLMs) have the potential to transform biological research by enabling autonomous experimentation. Yet, their application remains constrained by rigid protocol design, limited adaptability to dynamic lab conditions, inadequate error handling, and high operational complexity. Here we introduce BioMARS (Biological Multi-Agent Robotic System), an intelligent platform that integrates LLMs, VLMs, and modular robotics to autonomously design, plan, and execute biological experiments. BioMARS uses a hierarchical architecture: the Biologist Agent synthesizes protocols via retrieval-augmented generation; the Technician Agent translates them into executable robotic pseudo-code; and the Inspector Agent ensures procedural integrity through multimodal perception and anomaly detection. The system autonomously conducts cell passaging and culture tasks, matching or exceeding manual performance in viability, consistency, and morphological integrity. It also supports context-aware optimization, outperforming conventional strategies in differentiating retinal pigment epithelial cells. A web interface enables real-time human-AI collaboration, while a modular backend allows scalable integration with laboratory hardware. These results highlight the feasibility of generalizable, AI-driven laboratory automation and the transformative role of language-based reasoning in biological research.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. LabEvolver: Training-Free Experience Evolution for Safe and Grounded Wet-Lab Agents

    cs.RO 2026-07 conditional novelty 6.0 of 10

    A training-free dual-loop framework that distills completed lab-robot trials into reusable state-paired experience raises wet-lab pH-task efficiency by ~48% and ALFWorld Success@20 from 76.2% to 91.4%.

Reference graph

Works this paper leans on

91 extracted references · 74 canonical work pages · cited by 1 Pith paper

  1. [1]

    O’Donoghue, O. et al. Bioplanner: automatic evaluation of llms on protocol planning in biology. arXiv preprint arXiv:2310.10632(2023)

    arXiv 2023

  2. [2]

    Huang, K. et al. Crispr-gpt: An llm agent for automated design of gene-editing experiments. arXiv preprint arXiv:2404.18021(2024)

    arXiv 2024

  3. [3]

    & Fakhruldeen, H

    Cooper, A., Zhou, Z., Veeramani, S., Galeano, F. & Fakhruldeen, H. Lira: Localization, inspection, and reasoning module for autonomous workflows in self-driving labs (2025)

    2025

  4. [4]

    & Davies, J

    Holland, I. & Davies, J. A. Automation in the life science research laboratory. Frontiers in bioengineering and biotechnology8, 571777 (2020)

    2020

  5. [5]

    et al.Open-source personal pipetting robots with live-cell incubation and microscopy compatibility

    Dettinger, P. et al.Open-source personal pipetting robots with live-cell incubation and microscopy compatibility. Nature Communications 13, 2999 (2022)

    2022

  6. [6]

    Novak, R. et al. A robotic platform for fluidically-linked human body-on-chips experimentation. Nature biomedical engineering4, 407 (2020)

    2020

  7. [7]

    & Ozaki, H

    Taguchi, S., Suda, Y., Irie, K. & Ozaki, H. Automation of yeast spot assays using an affordable liquid handling robot. SLAS technology 28, 55–62 (2023)

    2023

  8. [8]

    Hamm, J. et al. A modular robotic platform for biological research: Cell cul- ture automation and remote experimentation. Advanced Intelligent Systems 6, 2300566 (2024)

    2024

Show all 91 references
  1. [9]

    Tristan, C. A. et al. Robotic high-throughput biomanufacturing and functional differentiation of human pluripotent stem cells. Stem Cell Reports16, 3076–3092 (2021)

    2021

  2. [10]

    Rebia—robotic enabled biological automation: 3d epithelial tissue production

    K¨ oniger, L.et al. Rebia—robotic enabled biological automation: 3d epithelial tissue production. Advanced Science11, 2406608 (2024)

    2024

  3. [11]

    & Natsume, T

    Yachie, N. & Natsume, T. Robotic crowd biology with maholo labdroids. Nature biotechnology 35, 310–312 (2017)

    2017

  4. [12]

    Ochiai, K. et al. A variable scheduling maintenance culture platform for mam- malian cells. SLAS TECHNOLOGY: Translating Life Sciences Innovation26, 209–217 (2021)

    2021

  5. [13]

    & Zhao, H

    Chao, R., Mishra, S., Si, T. & Zhao, H. Engineering biological systems using automated biofoundries. Metabolic Engineering 42, 98–108 (2017)

    2017

  6. [14]

    & Lee, P

    Miles, B. & Lee, P. L. Achieving reproducibility and closed-loop automation in biological experimentation with an iot-enabled lab of the future. SLAS TECHNOLOGY: Translating Life Sciences Innovation23, 432–439 (2018). 16

    2018

  7. [15]

    Automated sample preparation with sp 3 for low-input clinical proteomics

    M¨ uller, T.et al. Automated sample preparation with sp 3 for low-input clinical proteomics. Molecular systems biology16, e9111 (2020)

    2020

  8. [16]

    Vaswani, A. et al. Attention is all you need. Advances in neural information processing systems30 (2017)

    2017

  9. [17]

    Wang, C. et al. Biorag: A rag-llm framework for biological question reasoning. arXiv preprint arXiv:2408.01107(2024)

    2024 arXiv

  10. [18]

    Luu, R. K. & Buehler, M. J. Bioinspiredllm: Conversational large language model for the mechanics of biological and bio-inspired materials. Advanced Science11, 2306724 (2024)

    2024

  11. [19]

    Zhang, Q. et al. Scientific large language models: A survey on biological & chemical domains. ACM Computing Surveys57, 1–38 (2025)

    2025

  12. [20]

    A., MacKnight, R., Kline, B

    Boiko, D. A., MacKnight, R., Kline, B. & Gomes, G. Autonomous chemical research with large language models. Nature 624, 570–578 (2023)

    2023

  13. [21]

    Darvish, K. et al. Organa: a robotic assistant for automated chemistry experimentation and characterization. Matter 8 (2025)

    2025

  14. [22]

    & Elhoseiny, M

    Zhu, D., Chen, J., Shen, X., Li, X. & Elhoseiny, M. Minigpt-4: Enhancing vision- language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)

    2023 arXiv

  15. [23]

    Zhang, J., Huang, J., Jin, S. & Lu, S. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence(2024)

    2024

  16. [24]

    & Khoei, T

    Singh, A., Ehtesham, A., Kumar, S. & Khoei, T. T. Agentic retrieval-augmented generation: A survey on agentic rag. arXiv preprint arXiv:2501.09136(2025)

    2025 arXiv

  17. [25]

    M. Bran, A. et al. Augmenting large language models with chemistry tools. Nature Machine Intelligence6, 525–535 (2024)

    2024

  18. [26]

    Kanda, G. N. et al. Robotic search for optimal cell culture in regenerative medicine. Elife 11, e77007 (2022)

    2022

  19. [27]

    Jiang, P., Ergu, D., Liu, F., Cai, Y. & Ma, B. A review of yolo algorithm developments. Procedia computer science199, 1066–1073 (2022)

    2022

  20. [28]

    Han, K. et al. A survey on vision transformer. IEEE transactions on pattern analysis and machine intelligence45, 87–110 (2022)

    2022

  21. [29]

    H., Bentley, J

    Friedman, J. H., Bentley, J. L. & Finkel, R. A. An algorithm for finding best matches in logarithmic expected time. ACM Transactions on Mathematical Software (TOMS) 3, 209–226 (1977). 17 18 Appendix A BioMARS Algorithmic Framework A.1 Core Algorithm Implementation Algorithm 1 ...

    1977

  22. [30]

    Knowledge Acquisition: Slit ← LiteratureSearch(Q) ▷ Literature search using Google/Bing Prag ← RAG(Slit) ▷ Extract key paragraphs using RAG K ← KnowledgeChecker(Prag) ▷ Reorganize knowledge

  23. [31]

    W orkflow Generation: Wraw ← WorkflowGenerator(K, Q, E) ▷ Generate initial workflow Wenv ← WorkflowChecker(Wraw, E) ▷ Adapt to experimental environment

  24. [32]

    Code Generation: Cbase ← CodeGenerator(Wenv, F) ▷ Map to 11 basic actions API Cvalid ← CodeChecker(Cbase, F) ▷ Validate logic and interfaces

  25. [33]

    Execution with Real-time Anomaly Detection: for a ∈ Actions(Cvalid) do MR ← StartAction(a) ▷ Begin executing action a while IsActionRunning(a) do ▷ Monitor during execution evit ← ViTKeyPointDetection(CurrentState(a)) ▷ Keyframe visual detection if evit = warning then evlm ← V...

  26. [34]

    Replanning: Wnew ← BiologistPlanner(L, Wenv) ▷ Human-in-the-loop planning Goto Step 3

  27. [35]

    Each question uses the same format, with the only variation being the **cell type name**:

    Return: return R, L 19 Appendix B Biologist Agent B.1 10-Query Biological Task Set The following 10 standardized questions apply to all 7 cell lines (HeLa, HUVEC, HepG2, DC2.4, Y79, K562, CHO). Each question uses the same format, with the only variation being the **cell type name**:

  28. [36]

    How to resuscitate [Cell Type] cells in detail?

  29. [37]

    How to perform passaging of [Cell Type] cells in detail?

  30. [38]

    How to change the medium for [Cell Type] cells in detail?

  31. [39]

    How to freeze and store [Cell Type] cells in detail?

  32. [40]

    What is the recommended seeding density for the [Cell Type] cell line?

  33. [41]

    How to detect the metabolic activity of the [Cell Type] cell line in detail?

  34. [42]

    How to evaluate the apoptotic level of the [Cell Type] cell line in detail?

  35. [43]

    What is the cryopreservation solution formula for the [Cell Type] cell line?

  36. [44]

    How to detect the proliferation of the [Cell Type] cell line in detail?

  37. [45]

    How to culture 3D cell spheres using the [Cell Type] cell line in detail? B.2 Protocol Generation Evaluation Metrics Score 5 User Question: ”How to change the medium for HepG2 cells in detail?” Model: deepseek-r1+WC+KC Standard: Very detailed and biologically accurate cell cul...

  38. [46]

    Retrieve ContainerA from the incubator and place it on one of the two pipetting platforms

  39. [47]

    Aspirate 10ml of old medium using a sterile pipette

  40. [48]

    Add 10ml pre-warmed (37 °C) PBS, gently shake horizontally for 10 seconds, then aspirate completely

  41. [49]

    Add 10ml fresh pre-warmed (37 °C) culture medium (e.g., DMEM with 10% FBS), shake horizontally for 10 seconds to distribute

  42. [50]

    Reason: • Includes precise volumes (10ml PBS/10ml medium)

    Return ContainerA to the incubator (37 °C, 5% CO). Reason: • Includes precise volumes (10ml PBS/10ml medium). • Specifies reagent pre-warming (37°C) and explicit culture conditions (37°C/5% CO). • Detailed operational instructions (e.g., horizontal shaking). Score 4 User Quest...

  43. [51]

    Pre-warm 10ml PBS and 10ml fresh medium to 37 °C

  44. [52]

    Retrieve ContainerA, aspirate 10ml old medium using a serological pipette

  45. [53]

    Repeat PBS wash

    Add 10ml PBS, gently swirl for 10 seconds, aspirate. Repeat PBS wash

  46. [54]

    Add 10ml fresh medium, swirl in a figure-eight motion for 15 seconds

  47. [55]

    Reason: • Steps are operationally correct but omit critical medium composition details (e.g., ”DMEM + 10% FBS”)

    Return ContainerA to the incubator. Reason: • Steps are operationally correct but omit critical medium composition details (e.g., ”DMEM + 10% FBS”). • Volumes are specified, but the lack of reagent specifics reduces biological clarity. Score 3 User Question: ”How to freeze and...

  48. [57]

    Discard old culture dish

  49. [58]

    Obtain new culture dishes ContainerB and ContainerC

  50. [59]

    Place ContainerA on a pipetting platform

  51. [60]

    Add 5ml of liquid to ContainerA

  52. [61]

    Add 2ml of enzyme solution to ContainerA

  53. [62]

    Shake ContainerA to ensure even distribution of the enzyme solution

  54. [63]

    Centrifuge ContainerA in TubeA at 1000 x g for 5 minutes

  55. [64]

    Remove the liquid after centrifugation

  56. [65]

    Add 5ml of culture medium to ContainerA

  57. [66]

    Distribute the cell suspension into ContainerB and ContainerC in perfect aliquots

  58. [67]

    Shake ContainerB and ContainerC to ensure uniform dispersion of cells

  59. [68]

    Place ContainerB on a pipetting platform

  60. [69]

    Add 5ml of culture medium to ContainerB

  61. [70]

    Store ContainerB in the incubator for cryopreservation

  62. [71]

    Place ContainerC on a pipetting platform

  63. [72]

    Add 5ml of culture medium to ContainerC

  64. [73]

    Reason: • Specific details are lacking, and there are issues with parameters such as centrifugal force

    Store ContainerC in the incubator for cryopreservation. Reason: • Specific details are lacking, and there are issues with parameters such as centrifugal force. 21 Score 2 User Question: ”How to change the medium for HepG2 cells in detail?” Model: deepseek-r1 Standard: Extremel...

  65. [74]

    Remove ContainerA, aspirate old medium

  66. [75]

    Add 5ml PBS, swirl, aspirate

  67. [76]

    Add 3ml trypsin-EDTA, incubate 5 minutes

  68. [77]

    Neutralize with 6ml medium, transfer to TubeA

  69. [78]

    Centrifuge TubeA at 300xg for 5 minutes, resuspend in 10ml fresh medium

  70. [79]

    Reason: • Includes irrelevant steps for HepG2 medium change (e.g., trypsin digestion, cen- trifugation, subculturing)

    Transfer 5ml to ContainerB and ContainerC, incubate. Reason: • Includes irrelevant steps for HepG2 medium change (e.g., trypsin digestion, cen- trifugation, subculturing). • Logic is flawed (e.g., splitting into multiple containers). HepG2 medium change does not require cell p...

  71. [80]

    Take out ContainerA from the incubator

  72. [81]

    Obtain two new 10cm cell culture dishes named ContainerB and ContainerC from the petri dish storage rack

  73. [82]

    Add 10ml of culture medium to ContainerB and ContainerC

  74. [83]

    Discard the old culture dish ContainerA

  75. [84]

    Add 5ml of cell suspension to ContainerB and ContainerC in perfect aliquots

  76. [85]

    Shake the culture dishes to ensure even distribution of cells

  77. [86]

    Place the culture dishes on the pipetting platform

  78. [87]

    Add 2ml of enzyme and buffer solution to ContainerB and ContainerC

  79. [88]

    Shake the culture dishes again to ensure uniform dispersion of cells in the cell suspension

  80. [89]

    Centrifuge the culture dishes at a centrifugal force of 3000g for 5 minutes using TubeA

  81. [90]

    Remove the liquid used to wash the cells after washing

  82. [91]

    Resuspend the cells in ContainerB and ContainerC by adding 10ml of fresh culture medium

  83. [92]

    PBS", float volume, string container) Aspirate a specified solu- tion and add it to the designated container. liquid type: The type of the specified solution (default:

    Place the culture dishes back in the incubator for further growth. Reason: • It is completely unclear what it is doing. 22 Appendix C Technician Agent C.1 Functional Interface Specifications T able C1: Function List for Cell Culture Operations F unction Description Parameters ...

Pith tools

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