Pith. sign in

REVIEW 2 major objections 5 minor 25 references

ChemWorld: Programmable Chemical Worlds for Controlled and Replayable Agent Experimentation

T0 review · 2 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read ChemWorld makes the chemical world itself a controlled experimental variable: reusable components compile into executable worlds whose hidden laws can be changed one at a time while the agent-facing task stays fixed and every trajectory…

desk verdict Programmable chemical worlds with exact replay and single-law forks, but the fork attribution needs a tighter noise-identity story. read the letter →

arxiv 2608.10792 v1 pith:25X3MFDP submitted 2026-08-11 cs.AI cs.LG

Jiangjie Qiu , Yijun Li , Xiaonan Wang This is my paper · ORCID
classification cs.AIcs.LG
keywords programmablechemicalenvironmentsautonomouschemistryagentsdeterministicreplaytransactionalexecutioncontrolledcounterfactualsworldcompilationagentexperimentationsimulation
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

The paper introduces ChemWorld, a programmable chemical environment that treats the world itself, not just the agent's task, as the experimental variable. Reusable reaction, separation, and observation components are compiled into executable worlds that share one agent-facing public contract (operations, instruments, observations, resources, failure rules), while constitutive laws, material properties, and hidden parameters stay evaluator-owned. Every action runs transactionally, and complete environment–action trajectories replay exactly at zero numerical tolerance. The payoff is controlled counterfactuals: change one private law, hold the public task and action sequence fixed, and attribute the resulting trajectory differences to that intervention. Six parent–child world forks and eight deterministic lifecycles, including a planned rollback with recovery, demonstrate the substrate within the declared component domain.

What carries the argument

The load-bearing objects are the public/private world split $W = (W_{\mathrm{pub}}, \theta)$ and the three-level identity bookkeeping: world-spec ID $\mathrm{id}(W_{\mathrm{pub}}, \theta)$, scenario ID $\mathrm{id}(W_{\mathrm{pub}}, \theta, \zeta_{\mathrm{init}}, \zeta_{\mathrm{dyn}}, \zeta_{\mathrm{obs}})$, and task–world unit. The public contract $T$ is what an agent ever sees; $\theta$ is what an evaluator can change. The transactional per-step pipeline — preflight admission $P$, runtime preconditions, candidate generation $F_\theta(s_t, R_t, a_t, \xi_t)$, post-execution validation $C$, commit or rollback with branch-specific ledger $G_b$ — is what makes records process-complete: failures, resource deltas, and RNG restoration are all recorded, so a replay can resubmit the exact typed-action trace including rejections and rollbacks. Controlled forks pair $W_p = (W_{\mathrm{pub}}, \theta_p)$ and $W_c = (W_{\mathrm{pub}}, \theta_c)$ with identical public contract, action sequence, and bound randomness, so any divergence in state or observation is attributable to the one registered private-law change.

What would settle it

Submit an action that fails post-execution validation so the rollback branch restores the observation-RNG state, replay the identical trace on the same frozen environment, and compare the next observation's noise: any divergence at zero tolerance would falsify the exact-replay and RNG-restoration claims, since the paper's mechanism says the restored state must reproduce identical future randomness. A second test is to re-run any frozen world on a different platform or with rebuilt dependencies and compare trajectories at zero numerical tolerance; any mismatch confirms the replay guarantee is bound to the exact software snapshot.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that chemical worlds can be built from declared components and compiled like software, with the public/private split doing the experimental work. A world is written $W = (W_{\mathrm{pub}}, \theta)$: the public contract $T$ exposes typed actions, instruments, observations, resources, termination and evaluation, while $\theta$ holds the private constitutive and material laws, hidden parameters, and private initialization. A compatibility compiler accepts only declarations that pass dependency, state-ownership, unit, domain, resource, instrument, and lifecycle checks, so invalid worlds fail before construction. The transactional runtime admits each typed action through preflight and precondition gates, generates a candidate transition, validates it, and either commits or rolls back, recording branch-specific attempt costs and restoring the observation-RNG state on failure. Replay reconstructs the bound world with its seeds ($\zeta_{\mathrm{init}}, \zeta_{\mathrm{dyn}}, \zeta_{\mathrm{obs}}$) and intervention record and resubmits the full trace, including rejections and rollbacks, comparing observations, rewards, events, and state integrity at zero tolerance. The qualification census (64 reference units, 1,786 recipes, 52 generated compositions, 192 negative probes, 8 lifecycles, 6 fork pairs) reports complete passage, and an independent agent closed a 15-action lifecycle in a non-reference world through the same public interface. The paper explicitly limits these guarantees to the declared component and model domain: replay is version-bound, not cross-platform, and the synthetic instruments are not calibrated digital twins.

Load-bearing premise

Exact replay and single-law attribution assume the runtime is fully deterministic within the bound software environment and that the recorded seeds and typed-action trace capture every source of randomness; the paper itself states the guarantee does not extend to cross-platform or cross-version replay.

Editorial extensions

If this is right

  • Researchers can run strictly matched counterfactuals — same public task, instruments, actions, and noise seeds — where only one hidden law differs, and attribute trajectory differences to that single change.
  • Every agent lifecycle, including failures, rollbacks, and resource debits, produces an evaluator-complete record that can be replayed exactly at zero numerical tolerance, making the experimental process itself auditable rather than just its outcome.
  • New chemical worlds enter the same transaction and replay machinery without redesigning the agent-facing contract, since world authoring is separated from process-model implementation and agent integration.
  • An independent agent can operate a non-reference world through the same public interface while the evaluator retains the full hidden-state record, supporting evaluation without exposing the mechanism being evaluated.
  • Within the declared domain, failure becomes a recorded experimental event: a planned invalid action rolls back to committed state, and recovery continues from that state within the same experimental unit.

Reading between the lines

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

  • If the substrate holds, agent evaluation could shift from terminal outcomes toward process records — for example using the paper's 19 process coordinates to compare when agents gather evidence, how they respond to rollbacks, and whether they detect a changed hidden law across forks; the paper defines these coordinates but does not use them to rank agents.
  • The same public/private split suggests a testable generalization: calibrated kinetic or instrument models entering through the declared interfaces could let software worlds serve as cheap pre-screens for physical self-driving laboratories, with replayable conditions then transferred as hypotheses; the paper names this workflow as a natural direction but does not demonstrate transfer.
  • A boundary condition of the zero-tolerance replay claim is that the guarantee inherits every numerical detail of the bound software stack, so the practical unit of reproducibility is the frozen release (v0.1.0) rather than the world description alone; porting a world across platforms or versions would likely break exactness.
  • Single-law attribution is only as strong as the completeness of the recorded randomness: any nondeterminism not captured by the three seeds and the typed-action trace would silently flow into both parent and child runs, so the fork design implicitly argues that all nondeterminism in the system is seed-bound.
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

2 major / 5 minor

Summary. The paper introduces ChemWorld, a programmable chemical-world simulator in which reusable process and observation components are compiled into executable worlds with a public agent-facing contract and evaluator-owned private chemical/material laws. The core methodological claims are: (1) worlds can be composed and varied while holding the public task and interaction conditions fixed; (2) transactional execution records all operations, failures, resource changes, and state transitions, enabling exact replay; and (3) single-private-law forks allow controlled counterfactual attribution of trajectory differences. The manuscript reports a protocol-frozen qualification census (64 reference task–world units, 1,786 recipes, 52 generated compositions, 192 negative probes, 32 module probes, 7 interface paths, 8 deterministic use cases, 6 fork pairs), plus one independent-agent lifecycle on a non-reference world. The paper explicitly limits its claims to the declared component/model domain and states that its module fixtures and directionality oracles are internal consistency checks rather than independent physical or cross-implementation validation.

Significance. If the claims hold, ChemWorld would be a valuable substrate for controlled, repeatable experiments over chemical worlds and agent behavior, complementing physical laboratories and fixed-task benchmark suites. The paper's strengths include a frozen qualification protocol, a full census with explicit denominators, a versioned open-source release, an offline manifest verifier, and unusually candid statements of scope and limitation (§4.2, §7.3). The separation of world qualification from agent execution is also methodologically sound. However, the central controlled-fork attribution claim is currently under-specified: the paper fixes seeds but does not establish that the realized observation-noise stream is identical across parent and child forks. Because the observation-RNG state evolves with the committed trajectory, fixing the seed alone does not fix the noise realization after a law change. This is a load-bearing gap for Contribution 3, and it must be addressed before the attribution claim can be accepted as stated.

major comments (2)
  1. [§5.2 and §3.3] The conclusion in §5.2 that trajectory differences are 'attributable to the registered private-law change under fixed actions and noise identity' is not supported by the protocol as written. Section 3.3 states that committed state binds the observation-RNG state ρ_t and that candidate transitions consume recorded random variates ξ_t. Because ρ_t evolves with the committed trajectory, a child run with a changed private law will generally consume different random draws than the parent after the point of divergence, even when the scenario seeds ζ_obs are identical. The manuscript does not specify a common-random-numbers mechanism, such as a dedicated RNG stream advanced on a state-independent schedule or injection of the parent's recorded ξ_t into the child run. Without such a mechanism, the child's observation noise is not established to be the same realization as the parent's, so the observed differences are not strictly attributable to the law change alone. This is load-bearing for the paper's central controlled-counterfactual claim and needs to be either implemented and documented or explicitly flagged as a residual confound.
  2. [Appendix A.4 and Table 3] The controlled-fork acceptance criteria in Appendix A.4 contain no gate that compares the realized observation-noise stream between parent and child. The listed gates—lineage, exactly-one-private-target, public-contract invariance, same-sequence executability, expected-state/observation divergence, and exact-replay—are all consistent with a child whose noise draws differ from the parent's. Passing these gates therefore does not rule out noise-realignment effects. The paper should either add an explicit noise-identity check to the fork protocol (for example, verifying that the aligned observation-RNG states are identical or that the child consumed exactly the parent's recorded variates), or it should soften the attribution statement in §5.2 to say that differences are attributable to the law change plus any residual noise realignment.
minor comments (5)
  1. [Figure 1C] The figure states 'Differences attributable only to private-law change' without the caveat 'within the declared model domain and with a verified common-noise mechanism'; this wording should be qualified to match the scope stated in §7.3.
  2. [Abstract and §7.3] The abstract says trajectories 'can be replayed exactly,' which could be read as a stronger guarantee than the version-bound replay defined in §7.3. Consider adding 'within the same bound software version' or an equivalent qualifier to the abstract.
  3. [§3.1] The symbol T is used both for the public task contract tuple and for 'task ID' in the identity-level definitions; this overloading is confusing and should be disambiguated.
  4. [Table 3] The row label 'partition-base response K1.00 → K1.75' appears to be a typo for 'phase-partition response' or 'partition response'; please correct it.
  5. [§4.3] The sentence stating that solver-diagnostic and candidate-observation fault paths are implemented but were not assigned separate qualification denominators is easy to miss; it would be helpful to state explicitly in the main text that these paths are not covered by the 192-probe census.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's replay and fork claims are internal determinism/consistency checks that the paper explicitly scopes as such, and no fitted parameter is renamed as a prediction.

full rationale

ChemWorld is an engineering/systems paper, not a derivation of scientific predictions from first principles, so the circularity patterns sought here largely do not apply. The exact-replay claim (Section 3.3, Appendix A.5) is a self-consistency determinism check: the runtime reconstructs the bound world from recorded identities and seeds, resubmits the full typed-action trace, and compares against the recorded trajectory at zero tolerance. This is an empirical verification of implementation determinism, not a prediction that reduces by construction to its inputs. The controlled-fork attribution claim (Section 5.2, Table 3, A.4) uses protocol-frozen direction and magnitude oracles; the thresholds were fixed before execution and are not fitted to the observed fork outcomes. The paper repeatedly and explicitly discloses that its qualification oracles are internal: 'Their numerical fixtures and directional oracles are internal qualification checks, not independent reference implementations or external physical validation' (Section 4.2) and 'Module fixtures and directionality oracles primarily test consistency with the authored models rather than an independent reference implementation or empirical ground truth' (Section 7.3). It also scopes replay: 'Exact replay is likewise restricted to the bound software identities, world identity and environment–action trace; it does not imply policy re-execution, cross-platform numerical identity or cross-version archival replay' (Section 7.3). The agent experiment is deliberately separated from world qualification ('World qualification and agent execution were kept as independent experimental units so that agent success could not serve as evidence for world correctness', Section 6.2), avoiding the self-validation circularity that would otherwise be plausible. The skeptic's concern about controlled forks—that fixing seeds ζ_init, ζ_dyn, ζ_obs may not fix the realized observation-noise stream because the observation-RNG state ρ_t evolves with the committed trajectory—is a real soundness gap in the attribution claim, but it is an under-specification of the fork protocol, not a logical reduction of the conclusion to its premises. No self-citation is load-bearing, no uniqueness theorem is imported from the authors' prior work, and no known result is renamed.

Assumptions & free parameters 3 free parameters · 4 assumptions · 2 invented entities

The platform's claims rest on four domain assumptions about determinism, model coherence, sampling representativeness, and information-boundary enforcement. Two author-invented abstractions (identity scheme and process coordinates) organize the framework. No numerical parameters are fitted to external data; the listed free parameters are hand-set design choices for time limits, fork thresholds, and coverage bounds.

free parameters (3)
  • Process-time limits t_max per pattern = e.g., phase-observation 0 s, reaction-thermal 3,600 s, reaction-distillation 10,440 s, continuous flow 7,200 s
    Hand-set in Table 4 from t_required + t_reserve + t_repeat to gate resource feasibility; they define which action sequences are admissible but are design choices, not fitted to data.
  • Controlled-fork acceptance thresholds = partition: delta >= 1e-4 mol abs and rel >= 0.05; electrochemical: delta >= 1e-6 mol abs and rel >= 0.01; plus…
    Protocol-frozen divergence oracle criteria in Table 3 and A.4; chosen by the authors to define passing forks, not estimated from data.
  • Continuous coverage bounds = e.g., heat 350-390 K, duration 600-1800 s, reflux 1.0-3.0, potential 0.5-1.8 V, etc.
    Authored sampling domains in Appendix C.2 that define the generated worlds; arbitrary design choices that shape the qualification sample.
assumptions (4)
  • domain assumption Deterministic execution given bound identities and recorded seeds
    Exact replay and fork attribution require that resubmitting the same typed actions under the same world/scenario identities and seeds reproduces the same trajectory; the paper tests this but does not prove it for all future platforms (Section 3.3, 7.3).
  • domain assumption Authored component models are coherent within their declared domains
    The chemical and material laws (e.g., Arrhenius kinetics, van't Hoff solubility, Fenske distillation) are taken as given; physical validity is declared out of scope, but the usefulness of the substrate for agent studies depends on internal coherence (Table 8, Section 7.3).
  • domain assumption Coverage sampling represents the declared construction domain
    The qualification uses pairwise rows and Latin hypercube samples over authored bounds; the paper explicitly limits conclusions to this finite sample (Section 4.1, C.2).
  • domain assumption Public/private information boundary is enforceable at runtime
    The agent-facing packets are checked for undeclared private fields, but the guarantee rests on the runtime correctly separating public and evaluator projections (Section 3.3, A.5).
invented entities (2)
  • Three-level identity scheme (world-spec ID, scenario ID, task-world unit)
    purpose: Separates world identity, randomness identity, and task identity to define overlap and to bind replay and fork comparisons.
    Author-defined abstraction with no external existence; its value is internal to the framework and is not independently verified.
  • 19 process coordinates
    purpose: Provide a non-composite record of agent experimental strategy (e.g., closed lifecycle fraction, post-measure continuation prevalence, maximum incumbent drawdown).
    Author-defined ledger entries with fixed formulas and null rules (Appendix D.1); not calibrated against external benchmarks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ChemWorld: Programmable Chemical Worlds for Controlled and Replayable Agent Experimentation." pith.science (2026). https://pith.science/paper/25X3MFDP

@misc{pith2026260810792,
  author       = {Pith},
  title        = {Pith review of: ChemWorld: Programmable Chemical Worlds for Controlled and Replayable Agent Experimentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/25X3MFDP}},
  note         = {Machine review of arXiv:2608.10792}
}
read the original abstract

Autonomous chemistry increasingly depends on environments in which agents can repeatedly act, observe, and adapt.Physical laboratories provide essential real-material evidence but are costly to repeat and difficult to use for tightly matched interventions, whereas most digital environments keep the underlying experimental world largely fixed. We introduce ChemWorld, a programmable chemical environment in which reusable process and observation components are compiled into executable worlds. ChemWorld separates the public experimental contract available to an agent from evaluator-owned chemical and material laws. Researchers can therefore vary world composition and operating conditions, or change a single hidden law while holding the public task and interaction conditions fixed. Transactional execution records operations, failures, resource changes, and state transitions, allowing complete environment-action trajectories to be replayed exactly and audited. Full-census qualification covered the reference registry, 52 generated compositions, and module, interface, compilation, and invalid-action tests. Eight deterministic experimental cases demonstrated shared lifecycle semantics, failure recovery, and exact replay, while six parent-child world-fork pairs isolated the effects of single private-law interventions under matched public conditions. An independent agent also completed a full lifecycle in a non-reference world through the same public interface. Within the declared component and model domain, ChemWorld provides a controlled and replayable substrate for studying experimentation across systematically varied chemical worlds, complementary to physical-laboratory evidence and calibration.

Figures

Figures reproduced from arXiv: 2608.10792 by the authors.

Figure 1
Figure 1. ChemWorld composes declared modules into executable worlds with transactional semantics, deter [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Construction coverage and qualification design. A, [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Execution, intervention and agent access. A, [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 14 canonical work pages

  1. [1]

    Boiko, Robert MacKnight, Ben Kline, and Gabe Gomes

    Daniil A. Boiko, Robert MacKnight, Ben Kline, and Gabe Gomes. Autonomous chemical research with large lan- guage models.Nature, 624:570–578, 2023. doi: 10.1038/ s41586-023-06792-0

    work page 2023

  2. [2]

    Bran, Sam Cox, Oliver Schilter, Carlo Bal- dassari, Andrew D

    Andres M. Bran, Sam Cox, Oliver Schilter, Carlo Bal- dassari, Andrew D. White, and Philippe Schwaller. Aug- menting large language models with chemistry tools.Na- ture Machine Intelligence, 6:525–535, 2024. doi: 10.1038/ s42256-024-00832-8

    work page 2024

  3. [3]

    Szymanski et al

    Nathan J. Szymanski et al. An autonomous laboratory for the accelerated synthesis of inorganic materials.Nature, 624:86–91, 2023. doi: 10.1038/s41586-023-06734-w

    doi:10.1038/s41586-023-06734-w 2023

  4. [4]

    Szczyp- iński, et al

    Tianyi Dai, Sriram Vijayakrishnan, Filip T. Szczyp- iński, et al. Autonomous mobile robots for exploratory synthetic chemistry.Nature, 635:890–897, 2024. doi: 10.1038/s41586-024-08173-7

    doi:10.1038/s41586-024-08173-7 2024

  5. [5]

    OR- GANA: A robotic assistant for automated chemistry ex- perimentation and characterization.Matter, 8(2), 2025

    Kourosh Darvish, Marta Skreta, Yuchi Zhao, et al. OR- GANA: A robotic assistant for automated chemistry ex- perimentation and characterization.Matter, 8(2), 2025. doi: 10.1016/j.matt.2024.10.015

    doi:10.1016/j.matt.2024.10.015 2025

  6. [6]

    A multiagent- driven robotic AI chemist enabling autonomous chemical research on demand.Journal of the American Chemical Society, 147(15):12534–12545, 2025

    Tao Song, Man Luo, Xiaolong Zhang, et al. A multiagent- driven robotic AI chemist enabling autonomous chemical research on demand.Journal of the American Chemical Society, 147(15):12534–12545, 2025. doi: 10.1021/jacs. 4c17738

    doi:10.1021/jacs 2025

  7. [7]

    Au- toLabs: cognitive multi-agent systems with self-correction for autonomous chemical experimentation.Scientific Re- ports, 16:19554, 2026

    Gihan Panapitiya, Emily Saldanha, Heather Job, et al. Au- toLabs: cognitive multi-agent systems with self-correction for autonomous chemical experimentation.Scientific Re- ports, 16:19554, 2026. doi: 10.1038/s41598-026-45593-z

    doi:10.1038/s41598-026-45593-z 2026

  8. [8]

    Bayley, et al

    Simone Pilon, Elia Savino, Oliver M. Bayley, et al. A flexible and affordable self-driving laboratory for auto- mated reaction optimization.Nature Synthesis, 2026. doi: 10.1038/s44160-026-01053-0

    work page doi:10.1038/s44160-026-01053-0 2026

Show all 25 references
  1. [9]

    Prince, Tao Zhou, Henry Chan, and Mathew J

    Aikaterini Vriza, Michael H. Prince, Tao Zhou, Henry Chan, and Mathew J. Cherukara. Operating advanced scientific instruments with AI agents that learn on the job.npj Computational Materials, 12:160, 2026. doi: 10.1038/s41524-026-02005-0

    2026 doi

  2. [10]

    Felton, Jan G

    Kobi C. Felton, Jan G. Rittig, and Alexei A. Lapkin. Sum- mit: Benchmarking machine learning methods for reaction optimisation.Chemistry–Methods, 1:116–122, 2021. doi: 10.1002/cmtd.202000051

    2021 doi

  3. [11]

    Hickman, Loïc M

    Florian Häse, Matteo Aldeghi, Riley J. Hickman, Loïc M. Roch, Melodie Christensen, Elena Liles, Jason E. Hein, and Alán Aspuru-Guzik. Olympus: a benchmarking framework for noisy optimization and experiment planning.Machine Learning: Science and Technology, 2:035021, 2021. doi:...

    2021 doi

  4. [12]

    PC-Gym: Benchmark environments for process control problems, 2024

    Maximilian Bloor, José Torraca, Ilya Orson Sandoval, et al. PC-Gym: Benchmark environments for process control problems, 2024. Preprint, arXiv:2410.22093

    2024 arXiv

  5. [13]

    ChemGymRL: A customizable inter- active framework for reinforcement learning for digital chemistry.Digital Discovery, 3:742–758, 2024

    Chris Beeler, Sriram Ganapathi Subramanian, Kyle Sprague, et al. ChemGymRL: A customizable inter- active framework for reinforcement learning for digital chemistry.Digital Discovery, 3:742–758, 2024. doi: 10.1039/D3DD00183K

    2024 doi

  6. [14]

    DiscoveryWorld: A virtual environment for developing and evaluating auto- mated scientific discovery agents

    Peter Jansen, Marc-Alexandre Côté, Tushar Khot, Erin Bransom, Bhavana Dalvi Mishra, Bodhisattwa Prasad Ma- jumder, Oyvind Tafjord, and Peter Clark. DiscoveryWorld: A virtual environment for developing and evaluating auto- mated scientific discovery agents. InAdvances in Neural...

    2024 doi

  7. [15]

    Li, Lyle Goodyear, Agam Bhatia, Louise Li, Aditi Bhaskar, Mohammed Zaman, and Noah D

    Kanishk Gandhi, Michael Y. Li, Lyle Goodyear, Agam Bhatia, Louise Li, Aditi Bhaskar, Mohammed Zaman, and Noah D. Goodman. BoxingGym: Benchmarking progress in automated experimental design and model discovery,

  8. [16]

    Maddison

    Haonan Duan, Stephen Zhewen Lu, Caitlin Fiona Harri- gan, Nishkrit Desai, Jiarui Lu, Michał Koziarski, Leonardo Cotta, and Chris J. Maddison. Measuring scientific capa- bilities of language models with a systems biology dry lab,

  9. [17]

    Agentic explo- ration of physics models.Physical Review X, 16:031002,

    Maximilian Nägele and Florian Marquardt. Agentic explo- ration of physics models.Physical Review X, 16:031002,

  10. [18]

    Preprint, arXiv:2507.02083

    arXiv

  11. [19]

    Malik, Tiarnan Doherty, Panagiotis Tigas, Muhammed Razzak, Stephen J

    Shreshth A. Malik, Tiarnan Doherty, Panagiotis Tigas, Muhammed Razzak, Stephen J. Roberts, Aron Walsh, and Yarin Gal. MADE: Benchmark environments for closed- loop materials discovery. InInternational Conference on Machine Learning, 2026

    2026

  12. [20]

    LabUtopia: High- fidelity simulation and hierarchical benchmark for scientific embodied agents

    Rui Li, Zixuan Hu, Wenxi Qu, et al. LabUtopia: High- fidelity simulation and hierarchical benchmark for scientific embodied agents. InAdvances in Neural Information Processing Systems, Datasets and Benchmarks Track, 2025

    2025

  13. [21]

    Nguyen, et al

    Tianshi Zheng, Kelvin Kiu-Wai Tam, Newt Hue-Nam K. Nguyen, et al. NewtonBench: Benchmarking generalizable scientific law discovery in llm agents. InInternational Conference on Learning Representations, 2026

    2026

  14. [22]

    Scaling scientific discovery environments for turn-level agentic RL, 2026

    Yucheng Xu, Keyi Zhang, Yuyang Yu, Min Zhang, Shiyuan Meng, Pei Chu, and Zhongying Tu. Scaling scientific discovery environments for turn-level agentic RL, 2026. Preprint, arXiv:2607.28990. 11 Programmable Chemical Worlds arXiv preprint Appendix A. Qualification and Experiment...

    2026 arXiv

  15. [24]

    Labimus: A simulation and benchmark for humanoid dexterous manipulation in chemical laboratory, 2026

    Yuhan Wu, Zhao Jin, Tao Li, et al. Labimus: A simulation and benchmark for humanoid dexterous manipulation in chemical laboratory, 2026. Preprint

    2026

  16. [2025]

    Preprint, arXiv:2501.01540

  17. [2026]

    doi: 10.1103/xnqc-q6nt

    doi

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.