Pith. sign in

REVIEW 2 major objections 3 minor 42 references

The paper claims that a schema-bound MCP protocol surface gives LLM-driven scientific instruments a bounded, testable, and reproducible control layer.

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

2026-08-01 19:15 UTC pith:GE7KAOUP

load-bearing objection A clean, honestly scoped software-engineering contribution to agentic microscopy; the central claim holds for the protocol layer, but the one load-bearing seam — the host-process bridge to real vendor software — is structurally tested only against a simulator stub. the 2 major comments →

arxiv 2607.17012 v1 pith:GE7KAOUP submitted 2026-07-19 cs.SE cond-mat.mtrl-sciphysics.ins-det

Schema-Bound LLM Control of Scientific Instrumentation through Model Context Protocol Skills

Roberto dos Reis , Vinayak P. Dravid This is my paper
classification cs.SE cond-mat.mtrl-sciphysics.ins-det
keywords Model Context Protocolagentic instrumentationschema-bound tool typingbounded executionlarge language modelslive processingskill abstractionelectron microscopy
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that connecting LLMs to scientific instruments is a protocol-design problem rather than an instrument-engineering problem, and proposes four coupled abstractions to solve it: schema-bound tool validation with physical bounds, a vendor-neutral adapter contract, a persistent live-job lifecycle, and prompt-registered skills that compose tools into reusable protocols. The reference implementation, validated against a physics-plausible simulator, passes 120 deterministic tests and shows small open-weight local models can drive the surface without cloud dependency. If the approach holds on real hardware, agent traces become analyzable, regressions become deterministic, and the layer above the adapter becomes portable across vendors.

Core claim

The central claim is that the four abstractions together define 'a bounded, testable, and reproducible foundation for closed-loop agentic instrumentation research.' Schema-bound validation rejects out-of-range or malformed arguments before any adapter call, the vendor-neutral abstract base class keeps the backend swappable, the live-job lifecycle promotes long-running analyses to typed tools, and skills registered as MCP prompts make multi-step procedures discoverable and adapter-portable. All validation is software-only, against a simulator that shares the same schema layer as live adapters, so the same checks run without instrument occupancy.

What carries the argument

The carrying mechanism is the layered MCP protocol surface: a schema-bound tool surface that validates every request against physical bounds before dispatch; a MicroscopeAdapter abstract base class that declares a 24-family capability vocabulary and vendor-neutral method signatures; a four-transition live-job state machine (start, status, result, stop); and skills implemented as MCP prompts that unroll into ordered tool sequences. These convert instrument actions into typed, replayable objects that an agent can reason about before a physical experiment is attempted.

Load-bearing premise

The load-bearing premise is that the versioned JSON bridge contract works on real vendor acquisition hosts with the assumed in-process/socket/COM host-binding categories and reliable event-loop pumping.

What would settle it

Attempt a live acquire_tem_image round trip through the bridge plugin in each host-binding category the paper classifies (in-process Python, socket-based service, COM/.NET). If the hello handshake, the 100 ms event-loop pump, or the JSON image payload fails on any one of them, the host-process crossing claim collapses.

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

If this is right

  • Rejected calls are recorded as structured errors naming the field, so audit traces reflect the agent's actual output rather than its rationalization.
  • Adding a new instrument requires subclassing the adapter and declaring capabilities; the 30 typed tools, six skills, and five live-job types remain unchanged.
  • Skills are adapter-portable: the same declarative protocol runs on any adapter that declares the required capabilities.
  • The hardware-independent suite runs deterministically without microscope occupancy, enabling continuous integration for agentic instrumentation.
  • Small open-weight models from multiple providers can drive the schema-bound surface on local hardware, satisfying local-first governance.

Where Pith is reading between the lines

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

  • If the bridge contract holds on live hosts, the schema-bound pattern likely generalizes beyond microscopy to other instrument classes where host-process binding and bounded execution recur—synchrotron beamlines, clean-room tools, and field-deployed sensors.
  • The simulator-plus-schema combination offers an off-the-shelf testbed for LLM tool-calling research: agent behavior can be evaluated against a physically plausible instrument without hardware access.
  • The deterministic CI probe that introspects the live protocol surface suggests a general pattern: for any MCP-based system, project status can become a measured quantity rather than a documented claim.
  • The reproducible three-test failures under one model indicate multi-call sequencing reliability is currently a model property, not a protocol property; future protocol designs might add explicit sequencing constraints to reduce dependence on model nondeterminism.

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

2 major / 3 minor

Summary. The manuscript describes a protocol-level method for attaching local LLM agents to scientific instruments over MCP. Four components are proposed: a Pydantic v2 schema-bound tool surface that rejects out-of-bounds arguments before dispatch; a vendor-neutral MicroscopeAdapter abstract base class plus versioned JSON bridge for host-process-bound vendor APIs; a persistent live-processing job lifecycle exposed as typed tools; and a skill abstraction registered as MCP prompts. A reference implementation, nuance-mcp, is validated in software against a physics-plausible simulator sharing the same adapter contract: 120 deterministic tests pass, 15 local-LLM integration tests pass 12–15/15, and a single-run probe of five small open-weight models passes 15/15. The authors state explicitly that live-instrument validation is out of scope.

Significance. If the claims hold, the paper contributes a useful architectural layer: schema-bound tool typing with auditable rejections, a swappable backend contract, and a reproducible regression environment that operates without instrument time. The open-source release with a deterministic suite and a project-status probe that introspects the live server is a concrete strength. The work is also careful to distinguish bounded execution from safety and to report LLM nondeterminism honestly. The main caveat is that the 'foundation' claim is established only for the authors' own simulator; the vendor-facing host-process boundary remains structurally specified rather than exercised.

major comments (2)
  1. [§2.2 / Appendix A / Appendix C] The host-process adapter pattern is a core component of the claimed method, but the only exercised bridge is the simulator-side stub. Appendix A specifies the JSON contract; Listing 1 leaves vendor_pump_events() and vendor_dispatch() as placeholders; Table 7 is a categorization from public docs; and Table 2's 13 bridge-protocol tests run against the reference simulator/plugin, not a vendor host. The paper's own Limitations section correctly states live-instrument validation is out of scope, but the abstract and title still say 'connecting local LLMs to scientific instruments' and 'host-process adapter pattern'. This untested boundary is load-bearing for those claims. Please either narrow the claim to a software-only protocol foundation and mark the host-process crossing as an untested extension, or add a minimal real-host integration test (e.g., a scriptable out-of-process vendor stub wi
  2. [§3.1 / §2.4] Section 3.1's 'adapter-replaceable backend' and Section 2.4's 'adapter-portable skills' are supported only by tests against the authors' own simulator, which is a peer implementation of the same contract. These tests demonstrate internal consistency of the protocol surface, not transferability to a vendor adapter. The Limitations paragraph acknowledges this in general terms, but the word 'foundation' in the abstract and conclusion may be read as stronger than the evidence. A dedicated threats-to-validity paragraph, or at least an explicit statement that portability is a design property not yet verified by an independent adapter implementation, would tighten the claim.
minor comments (3)
  1. [§3.3] The single-run probe is reported as 'all five passed 15/15' while Section 3.2 documents qwen2.5:7b failing 3/15 reproducibly. The discrepancy is consistent with sampling nondeterminism, but a brief explanatory sentence would help readers avoid confusion.
  2. [Appendix C, Listing 2] _bridge_call is a blocking socket call inside an async method; consider wrapping it in an executor or documenting the threading model, especially since Appendix A promises the host event loop is pumped every 100 ms.
  3. [Table 3] The MCP-Bench row is all 'n/a' and adds little to the comparison; consider removing it or adding a sentence on why it is included.

Circularity Check

0 steps flagged

No significant circularity; central claims are demonstrated by implementation tests, and the software-only scope is stated explicitly.

full rationale

The paper makes no fitted-parameter predictions and claims no first-principles derivation. Its central contribution is a software protocol, and the validation is a conformance test: the reference simulator implements the same MicroscopeAdapter contract, so the 120 deterministic tests check that the schema layer, dispatch path, live-job lifecycle, and bridge JSON contract behave as specified. A conformance test can fail, so this is not a tautology. The only self-citation is [23], an invention disclosure used to state that live-instrument validation is intentionally out of scope and reported separately; it does not supply evidence for any claim asserted here. The line 'on-hardware performance cannot be assessed from this manuscript alone' is an explicit external-validity limitation, not a circular step. The bridge generalization in Table 7 is a categorization from public documentation, and the authors acknowledge the vendor boundary is untested; this is an unsupported transfer claim, properly classified as correctness risk rather than circularity. No equation is fit to data and renamed as a prediction, no uniqueness theorem from prior work is used to force a choice, and no ansatz is smuggled in via self-citation. Therefore the derivation chain is self-contained for what it actually claims.

Axiom & Free-Parameter Ledger

1 free parameters · 3 axioms · 0 invented entities

The central software claims rest on design choices (physical bounds, capability vocabulary, tool set) and on the simulator being a faithful proxy. There are no newly postulated physical entities. The bridge contract is an invented protocol, but it is a software artifact with structural tests, not a physical entity.

free parameters (1)
  • Schema physical bounds (alpha_deg, dwell-time, exposure ranges) = alpha in [-80, 80] deg; dwell-time in [0.5, 1e4] us; exposure in [1e-3, 60] s
    These bounds are chosen by hand as representative physical limits. They are not derived from a standard or calibration, and if they are wrong the schema could reject safe requests or accept unsafe ones. They are used in the simulator and validation but are not fitted to data.
axioms (3)
  • domain assumption The physics-plausible simulator's behavior is representative enough of real instrument behavior for the protocol validation to transfer.
    The entire validation suite runs against the simulator; if the simulator diverges from real hardware, the deterministic tests do not establish real-instrument correctness. The authors acknowledge this is a limitation (Section 4).
  • domain assumption The vendor host-binding categories (in-process Python, socket-based service, COM/.NET automation) cover the target instruments and behave as described.
    Table 7 categorizes Gatan, JEOL, Hitachi, and Thermo Fisher platforms. The bridge pattern and adapter guidance depend on this categorization, but no live vendor validation is provided.
  • domain assumption Local LLMs will honor MCP prompt-registered skills and schema-bound tool calls well enough to execute multi-step sequences.
    The 15 local-LLM integration tests show 12-15/15 pass; the three reproducible failures are due to the model re-ordering steps, showing this assumption is not always true. The skill abstraction relies on the model following the prompt.

pith-pipeline@v1.3.0-alltime-deepseek · 17731 in / 9509 out tokens · 87810 ms · 2026-08-01T19:15:03.861660+00:00 · methodology

0 comments
read the original abstract

Large language models (LLMs) can plan tool-mediated scientific work, but scientific instruments remain difficult to connect to such agents: vendor APIs may load only inside acquisition host processes, facilities may prohibit cloud-hosted agents, and natural-language interfaces can emit physically unreasonable arguments. We present a method for connecting local LLMs to scientific instruments through the Model Context Protocol (MCP). It combines: (1) a schema-bound tool surface that validates requests against physical bounds before adapter dispatch; (2) a vendor-neutral, host-process adapter pattern separating language-side reasoning from instrument-side execution; (3) a persistent lifecycle for long-running live-processing jobs; and (4) MCP-prompt-registered skills that compose typed tools into reusable multi-step protocols. Our open-source reference server exposes 30 typed tools, 5 live-job types, and 6 skills through a physics-plausible simulator implementing the same protocol surface. Validation is software-only: all 120 hardware-independent tests pass deterministically, while 15 local-LLM integration tests pass 12-15 of 15 across runs because of model nondeterminism. A preliminary single-run probe across five open tool-calling LLMs indicates that the schema-bound interface can be driven locally by small open-weight models without cloud dependency; it is not a benchmark and has no confidence intervals. The method provides a bounded, testable, and reproducible foundation for closed-loop agentic instrumentation research.

Figures

Figures reproduced from arXiv: 2607.17012 by Roberto dos Reis, Vinayak P. Dravid.

Figure 1
Figure 1. Figure 1: MCP instrumentation control architecture. The Model Context Protocol (MCP) connects a local LLM agent to a scientific instrument through a layered protocol surface that runs entirely on the instrument workstation. Three protocol layers sit above the backend: a skills layer of declarative MCP prompts (Section 2.4); a typed tool surface of 30 schema-validated tools, within which Pydantic v2 validation (Secti… view at source ↗
Figure 2
Figure 2. Figure 2: Schema-bound validation flow. A tool invocation proceeds through three discrete stages. The LLM agent emits a JSON tool call; the FastMCP server constructs the Pydantic v2 model instance and checks all physical bounds, types, and range constraints. Invalid requests are rejected at the boundary with a structured ValidationError naming the offending field; the adapter is never reached. Valid requests dispatc… view at source ↗
Figure 3
Figure 3. Figure 3: Vendor-neutral adapter contract. The MCP tool surface dispatches through the schema layer to the MicroscopeAdapter abstract base class. The ABC declares a capability vocabulary and 30 abstract methods with vendor-neutral signatures. Concrete implementations (SimulatorAdapter, GatanAdapter, JEOLAdapter, ThermoFisherAdapter, etc..) subclass the ABC and declare only the capabilities they support; undeclared c… view at source ↗
Figure 4
Figure 4. Figure 4: Persistent live-processing job lifecycle. Five live-job types (radial_profile, difference, fft_map, filtered_view, maximum_spot_mapping) share a common state machine with four tran￾sitions: start (IDLE→RUNNING), status (RUNNING→RUNNING, returns progress), result (RUNNING→COMPLETED, returns data), stop (RUNNING→STOPPED). State persists either server-side (simulator mode) or inside the host process (bridge m… view at source ↗
Figure 5
Figure 5. Figure 5: Declarative multi-step protocols (“skills”). Skills are registered as MCP prompts and are therefore discoverable via prompts/list. Each skill composes typed tools from the tool layer into an ordered sequence. The example shows the eels_survey skill unrolling into: (1) capability check (get_capabilities), (2) microscope-state check (get_microscope_state), (3) ZLP reference acquisition, (4) core-loss spectru… view at source ↗
Figure 6
Figure 6. Figure 6: Deployment and network model. The local workstation hosts the LLM stack (Ollama), the LangChain ReAct agent, and the FastMCP server with its schema layer and simulator. When a vendor adapter requires host-process binding, the bridge speaks a versioned JSON contract (nuance-mcp-bridge/1.0) over ZeroMQ on loopback (tcp://127.0.0.1:5555). The layered safety model distinguishes four levels: L1 schema validatio… view at source ↗
Figure 7
Figure 7. Figure 7: Validation behaviour of the v0.2 release against the reference simulator. (a) Pass/fail counts of the v0.2 automated suite by category. The hardware-independent, bridge￾protocol, and live-job lifecycle blocks are fully deterministic and pass in their entirety. The local-LLM (Ollama) block runs under qwen2.5:7b and fails three of fifteen tests reproducibly; the failures reflect model non-determinism in mult… 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

42 extracted references · 6 linked inside Pith

  1. [1]

    Autonomous chemical research with large language models.Nature, 624(7992):570–578, 2023

    Daniil A Boiko, Robert MacKnight, Ben Kline, and Gabe Gomes. Autonomous chemical research with large language models.Nature, 624(7992):570–578, 2023

    2023

  2. [2]

    Bran, Sam Cox, Oliver Schilter, Carlo Baldassari, Andrew D White, and Philippe Schwaller

    Andres M. Bran, Sam Cox, Oliver Schilter, Carlo Baldassari, Andrew D White, and Philippe Schwaller. Augmenting large language models with chemistry tools.Nature machine intelligence, 6(5):525–535, 2024

    2024

  3. [3]

    Self-driving laboratories for chemistry and materials science.Chemical Reviews, 124(16):9633–9732, 2024

    Gary Tom, Stefan P Schmid, Sterling G Baird, Yang Cao, Kourosh Darvish, Han Hao, Stanley Lo, Sergio Pablo-García, Ella M Rajaonson, Marta Skreta, et al. Self-driving laboratories for chemistry and materials science.Chemical Reviews, 124(16):9633–9732, 2024

    2024

  4. [4]

    Gonzalez

    Shishir G Patil, Huanzhi Mao, Fanjia Yan, Charlie Cheng-Jie Ji, Vishnu Suresh, Ion Stoica, and Joseph E. Gonzalez. The berkeley function calling leaderboard (BFCL): From tool use to agentic evaluation of large language models. In Aarti Singh, Maryam Fazel, Daniel Hsu, Simon Lacoste-Julien, Felix Berkenkamp, Tegan Maharaj, Kiri Wagstaff, and Jerry Zhu, edi...

    2025

  5. [5]

    Automated and autonomous experiments in electron and scanning probe microscopy.ACS nano, 15(8):12604–12627, 2021

    Sergei V Kalinin, Maxim Ziatdinov, Jacob Hinkle, Stephen Jesse, Ayana Ghosh, Kyle P Kelley, Andrew R Lupini, Bobby G Sumpter, and Rama K Vasudevan. Automated and autonomous experiments in electron and scanning probe microscopy.ACS nano, 15(8):12604–12627, 2021

    2021

  6. [6]

    Atomai: a deep learning framework for analysis of image and spectroscopy data in (scanning) transmission electron microscopy and beyond.arXiv preprint arXiv:2105.07485, 2021

    Maxim Ziatdinov, Ayana Ghosh, Tommy Wong, and Sergei V Kalinin. Atomai: a deep learning framework for analysis of image and spectroscopy data in (scanning) transmission electron microscopy and beyond.arXiv preprint arXiv:2105.07485, 2021. 18

    Pith/arXiv arXiv 2021

  7. [7]

    Autonomous scanning probe microscopy with hypothesis learning: Exploring the physics of domain switching in ferroelectric materials.Patterns, 4(3), 2023

    Yongtao Liu, Anna N Morozovska, Eugene A Eliseev, Kyle P Kelley, Rama Vasudevan, Maxim Ziatdinov, and Sergei V Kalinin. Autonomous scanning probe microscopy with hypothesis learning: Exploring the physics of domain switching in ferroelectric materials.Patterns, 4(3), 2023

    2023

  8. [8]

    Machine learning-based reward-driven tuning of scanning probe microscopy: Toward fully automated microscopy.ACS nano, 19(21):19659–19669, 2025

    Yu Liu, Roger Proksch, Jason Bemis, Utkarsh Pratiush, Astita Dubey, Mahshid Ahmadi, Reece Emery, Philip D Rack, Yu-Chen Liu, Jan-Chi Yang, et al. Machine learning-based reward-driven tuning of scanning probe microscopy: Toward fully automated microscopy.ACS nano, 19(21):19659–19669, 2025

    2025

  9. [9]

    Scientific exploration with expert knowledge (seek) in autonomous scanning probe microscopy with active learning.Digital Discovery, 4(1):252–263, 2025

    Utkarsh Pratiush, Hiroshi Funakubo, Rama Vasudevan, Sergei V Kalinin, and Yongtao Liu. Scientific exploration with expert knowledge (seek) in autonomous scanning probe microscopy with active learning.Digital Discovery, 4(1):252–263, 2025

    2025

  10. [10]

    Evaluating large language model agents for automation of atomic force microscopy.Nature Communications, 16(1):9104, 2025

    Indrajeet Mandal, Jitendra Soni, Mohd Zaki, Morten M Smedskjaer, Katrin Wondraczek, Lothar Wondraczek, Nitya Nand Gosvami, and NM Anoop Krishnan. Evaluating large language model agents for automation of atomic force microscopy.Nature Communications, 16(1):9104, 2025

    2025

  11. [11]

    Tem agent: enhancing transmission electron microscopy (tem) with modern ai tools

    Morgan K Wall, Alexander J Pattison, Edward S Barnard, Stephanie M Ribet, and Peter Ercius. Tem agent: enhancing transmission electron microscopy (tem) with modern ai tools. arXiv preprint arXiv:2511.08819, 2025

    arXiv 2025

  12. [12]

    Bridging electron microscopy and materials analysis with an autonomous agentic platform.Science Advances, 12(14):eaed0583, 2026

    Guangyao Chen, Wenhao Yuan, and Fengqi You. Bridging electron microscopy and materials analysis with an autonomous agentic platform.Science Advances, 12(14):eaed0583, 2026

    2026

  13. [13]

    Thinking microscopes: agentic ai and the future of electron microscopy.npj Computational Materials, 12(1):149, 2026

    Vida Jamali, Amirali Aghazadeh, and Josh Kacher. Thinking microscopes: agentic ai and the future of electron microscopy.npj Computational Materials, 12(1):149, 2026

    2026

  14. [14]

    Accelerating scientific discovery with co-scientist.Nature, pages 1–3, 2026

    Juraj Gottweis, Wei-Hung Weng, Alexander Daryin, Tao Tu, Petar Sirkovic, Artiom Myaskovsky, Grzegorz Glowaty, Felix Weissenberger, Alessio Orlandi, Dan Popovici, et al. Accelerating scientific discovery with co-scientist.Nature, pages 1–3, 2026

    2026

  15. [15]

    A multi-agent system for automating scientific discovery.Nature, pages 1–3, 2026

    Ali Essam Ghareeb, Benjamin Chang, Ludovico Mitchener, Angela Yiu, Caralyn J Szostkiewicz, Dmytro Shved, Gavin J Gyimesi, Jon M Laurent, Samantha M Wright, Muhammed T Razzak, et al. A multi-agent system for automating scientific discovery.Nature, pages 1–3, 2026

    2026

  16. [16]

    An ai system to help scientists write expert-level empirical software.Nature, pages 1–3, 2026

    Eser Aygün, Anastasiya Belyaeva, Gheorghe Comanici, Marc Coram, Hao Cui, Jake Garrison, Renee Johnston, Anton Kast, Cory Y McLean, Peter Norgaard, et al. An ai system to help scientists write expert-level empirical software.Nature, pages 1–3, 2026

    2026

  17. [17]

    Model Context Protocol: An open standard for connecting AI assistants to external data sources and tools.https://modelcontextprotocol.io, 2024

    Anthropic. Model Context Protocol: An open standard for connecting AI assistants to external data sources and tools.https://modelcontextprotocol.io, 2024. Accessed March 2026

    2024

  18. [18]

    Model context protocol (mcp): Landscape, security threats, and future research directions.ACM Transactions on Software Engineering and Methodology, 2025

    Xinyi Hou, Yanjie Zhao, Shenao Wang, and Haoyu Wang. Model context protocol (mcp): Landscape, security threats, and future research directions.ACM Transactions on Software Engineering and Methodology, 2025

    2025

  19. [19]

    Llm agents for interactive workflow provenance: Reference architecture and evaluation methodology

    Renan Souza, Timothy Poteet, Brian Etz, Daniel Rosendo, Amal Gueroudji, Woong Shin, Prasanna Balaprakash, and Rafael Ferreira da Silva. Llm agents for interactive workflow provenance: Reference architecture and evaluation methodology. InProceedings of the SC’25 Workshops of the International Conference for High Performance Computing, Networking, Storage a...

    2025

  20. [20]

    R-lam: Reproducibility-constrained large action models for scientific workflow automation.arXiv preprint arXiv:2601.09749, 2026

    Suriya Sureshkumar. R-lam: Reproducibility-constrained large action models for scientific workflow automation.arXiv preprint arXiv:2601.09749, 2026

    arXiv 2026

  21. [21]

    Audit trails for accountability in large language models.arXiv preprint arXiv:2601.20727, 2026

    Victor Ojewale, Harini Suresh, and Suresh Venkatasubramanian. Audit trails for accountability in large language models.arXiv preprint arXiv:2601.20727, 2026

    arXiv 2026

  22. [22]

    Pydantic v2: Data validation using Python type hints.https://docs

    Samuel Colvin et al. Pydantic v2: Data validation using Python type hints.https://docs. pydantic.dev, 2024. Accessed March 2026

    2024

  23. [23]

    Universalcontrolprotocolforscientificinstrumentation using extended Model Context Protocol

    RobertodosReisandVinayakP.Dravid. Universalcontrolprotocolforscientificinstrumentation using extended Model Context Protocol. Northwestern University Invention Disclosure Disc-ID- 25-05-22-002 (Technology ID 2025-136), accepted 3 June 2025, 2025. Assignee: Northwestern University

    2025

  24. [24]

    Model context protocol (mcp) tool descriptions are smelly! towards improving ai agent efficiency with augmented mcp tool descriptions.arXiv preprint arXiv:2602.14878, 2026

    Mohammed Mehedi Hasan, Hao Li, Gopi Krishnan Rajbahadur, Bram Adams, and Ahmed E Hassan. Model context protocol (mcp) tool descriptions are smelly! towards improving ai agent efficiency with augmented mcp tool descriptions.arXiv preprint arXiv:2602.14878, 2026

    Pith/arXiv arXiv 2026

  25. [25]

    The FAIR guiding principles for scientific data management and stewardship

    Mark D Wilkinson, Michel Dumontier, IJsbrand Jan Aalbersberg, Gabrielle Appleton, Myles Axton, Arie Baak, Niklas Blomberg, Jan-Willem Boiten, Luiz Bonino da Silva Santos, Philip E Bourne, et al. The FAIR guiding principles for scientific data management and stewardship. Scientific data, 3(1):1–9, 2016

    2016

  26. [26]

    Nodeology: creating graph-based agentic workflows for ai-assisted electron microscopy.Mi- croscopy and Microanalysis, 31(Supplement_1):ozaf048–1108, 2025

    Xiangyu Yin, Chuqiao Shi, Benjamin Fein-Ashley, Yu-Tsun Shao, Yimo Han, and Yi Jiang. Nodeology: creating graph-based agentic workflows for ai-assisted electron microscopy.Mi- croscopy and Microanalysis, 31(Supplement_1):ozaf048–1108, 2025

    2025

  27. [27]

    Four-dimensional scanning transmission electron microscopy (4d-stem): From scanning nanodiffraction to ptychography and beyond.Microscopy and Microanalysis, 25(3): 563–582, 2019

    Colin Ophus. Four-dimensional scanning transmission electron microscopy (4d-stem): From scanning nanodiffraction to ptychography and beyond.Microscopy and Microanalysis, 25(3): 563–582, 2019

    2019

  28. [28]

    py4dstem: A software package for four-dimensional scanning transmission electron microscopy data analysis.Microscopy and Microanalysis, 27(4):712–743, 2021

    Benjamin H Savitzky, Steven E Zeltmann, Lauren A Hughes, Hamish G Brown, Shiteng Zhao, Philipp M Pelz, Thomas C Pekin, Edward S Barnard, Jennifer Donohue, Luis Rangel DaCosta, et al. py4dstem: A software package for four-dimensional scanning transmission electron microscopy data analysis.Microscopy and Microanalysis, 27(4):712–743, 2021

    2021

  29. [29]

    Mea- surement of atomic electric fields and charge densities from average momentum transfers using scanning transmission electron microscopy.Ultramicroscopy, 178:62–80, 2017

    Knut Müller-Caspary, Florian F Krause, Tim Grieb, Stefan Löffler, Marco Schowalter, Armand Béché, Vincent Galioit, Dennis Marquardt, Josef Zweck, Peter Schattschneider, et al. Mea- surement of atomic electric fields and charge densities from average momentum transfers using scanning transmission electron microscopy.Ultramicroscopy, 178:62–80, 2017

    2017

  30. [30]

    Phase contrast stem for thin samples: Integrated differential phase contrast.Ultramicroscopy, 160:265–280, 2016

    Ivan Lazić, Eric GT Bosch, and Sorin Lazar. Phase contrast stem for thin samples: Integrated differential phase contrast.Ultramicroscopy, 160:265–280, 2016

    2016

  31. [31]

    Electron microscopy (big and small) data analysis with the open source software package hyperspy.Microscopy and Microanalysis, 23(S1):214–215, 2017

    Francisco de la Peña, Tomas Ostasevicius, Vidar Tonaas Fauske, Pierre Burdet, Petras Jokubauskas, Magnus Nord, Mike Sarahan, Eric Prestat, Duncan N Johnstone, Joshua Taillon, et al. Electron microscopy (big and small) data analysis with the open source software package hyperspy.Microscopy and Microanalysis, 23(S1):214–215, 2017

    2017

  32. [32]

    Springer Science & Business Media, 2011

    Ray F Egerton.Electron energy-loss spectroscopy in the electron microscope. Springer Science & Business Media, 2011. 20

    2011

  33. [33]

    PyZMQ: Python bindings for ZeroMQ.https://pyzmq.readthedocs

    The ZeroMQ Authors. PyZMQ: Python bindings for ZeroMQ.https://pyzmq.readthedocs. io, 2024. Accessed March 2026

    2024

  34. [34]

    Ollama: Run large language models locally.https://ollama.com, 2024

    Ollama. Ollama: Run large language models locally.https://ollama.com, 2024. Accessed March 2026

    2024

  35. [35]

    React: Synergizing reasoning and acting in language models.arXiv preprint arXiv:2210.03629, 2022

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models.arXiv preprint arXiv:2210.03629, 2022

    Pith/arXiv arXiv 2022

  36. [36]

    LangChain and LangGraph: Python libraries for building LLM-powered applications.https://python.langchain.com, 2024

    LangChain AI. LangChain and LangGraph: Python libraries for building LLM-powered applications.https://python.langchain.com, 2024. Accessed March 2026

    2024

  37. [37]

    Empowering electron ptychography with generative artificial intelligence and agentic workflows

    Yi Jiang, Xiangyu Yin, Chuqiao Shi, Benjamin Fein-Ashley, Yu-Tsun Shao, and Yimo Han. Empowering electron ptychography with generative artificial intelligence and agentic workflows. Microscopy and Microanalysis, 31(Supplement_1):ozaf048–1061, 2025

    2025

  38. [38]

    Automated electron microscope tomography using robust prediction of specimen movements.Journal of structural biology, 152(1):36–51, 2005

    David N Mastronarde. Automated electron microscope tomography using robust prediction of specimen movements.Journal of structural biology, 152(1):36–51, 2005

    2005

  39. [39]

    Kalinin, and Gerd Duscher

    Utkarsh Pratiush, Sergei V. Kalinin, and Gerd Duscher. stemOrchestrator: A software framework for integrating detector and camera APIs for automated STEM workflows.engrXiv preprint, 2025. URLhttps://engrxiv.org/preprint/view/4645

    2025

  40. [40]

    Experiences with Model Context Protocol servers for science and high performance computing.arXiv preprint arXiv:2508.18489, 2025

    Haochen Pan, Ryan Chard, Reid Mello, Christopher Grams, et al. Experiences with Model Context Protocol servers for science and high performance computing.arXiv preprint arXiv:2508.18489, 2025

    Pith/arXiv arXiv 2025

  41. [41]

    MCP-Bench: Benchmarking tool-using LLM agents with complex real-world tasks via Model Context Protocol servers

    Zhenting Wang, Qi Chang, Hemani Patel, Shashank Biju, Cheng-En Wu, Quan Liu, Aolin Ding, Alireza Rezazadeh, Ankit Shah, Yujia Bao, and Eugene Siow. MCP-Bench: Benchmarking tool-using LLM agents with complex real-world tasks via Model Context Protocol servers. arXiv preprint arXiv:2508.20453, 2025

    Pith/arXiv arXiv 2025

  42. [42]

    nuance-mcp-bridge/1.0

    Jingkang Liang, Niklas Groll, and Gürkan Sin. Large language model agent for user-friendly chemical process simulations, 2026. URLhttps://arxiv.org/abs/2601.11650. 21 Appendix A Bridge protocolnuance-mcp-bridge/1.0 This appendix specifies the versioned JSON contract that host-process-bound adapter implementa- tions must satisfy. The contract is vendor-neu...

    Pith/arXiv arXiv 2026