PaperlayerView exact revision on arXiv

Source-linked AI summary

Molecular LLM Agents: From Architectural Design to Scientific Autonomy

Jiatong Li, Wengyu Zhang, Weida Wang, Yuxuan Ren, Wei Liu, Chenyang Mao, Yuqiang Li, Yatao Bian, Changmeng Zheng, Xiaoyong Wei, Qing Li

arXiv:2608.23104v1cs.CLcs.AI

TL;DR

Molecular LLM agents need a framework for coordinating chemically grounded perception, tools, feedback, and decision-making across discovery workflows. This survey develops architectural and scientific-autonomy taxonomies, compares existing systems and evaluations, and concludes that trustworthy autonomy requires more than planning or tool use alone.

  • Problem

    Molecular agents lack a shared design framework, autonomy roadmap, and governance boundary despite differing representations, tools, feedback, optimization, and human oversight.

  • Method

    The survey organizes agents architecturally and classifies scientific autonomy by the outermost feedback loop reliably closed without mandatory human intervention.

  • Results

    The analysis concludes that scientific autonomy requires chemically faithful perception, grounded and verifiable tool use, feedback-aware decisions, provenance, optimization, and explicit safety domains.

  • Takeaways & Limitations

    The framework connects molecular-agent system design with demonstrated decision authority across computational and physical molecular workflows.

  • Takeaways & Limitations

    Adversarial inputs and jailbreaks can subvert safety alignment, creating dual-use risks for hazardous molecular design.

Abstract

from arXiv · show

Molecular science represents an important frontier for LLM-based agents. Unlike general agents that mainly operate over natural language, code, or web environments, molecular LLM agents must perceive, reason about, and act upon chemical objects across symbolic strings, molecular graphs, 3D conformations, spectra, simulations, and wet-lab measurements. Their capabilities depend on chemically faithful molecular perception, an LLM-centered agent framework, domain-specific tool grounding, and computational or experimental feedback, in addition to planning and tool use. This work develops a conceptual framework for molecular LLM agents from two complementary perspectives. First, we introduce an architectural view of molecular-agent design, covering molecular representation and perception, the agent framework, domain-specific toolboxes, and learning and optimization. Second, we propose a scientific autonomy ladder inspired by staged autonomy in engineering systems, categorizing agents into four levels: L1 assistive or fixed workflows, L2 adaptive computational agents, L3 feedback-aware physical experiment agents, and L4 scientific-agenda agents. Together, these two perspectives establish a comprehensive framework for comparing existing molecular LLM agents, identifying missing capabilities and deployment risks, and guiding the design, evaluation, and deployment of future agents in molecular discovery workflows.

1 Introduction

Molecular LLM agents address the orchestration bottleneck in molecular discovery by connecting molecular representations, reasoning, tools, and computational or experimental feedback. The paper proposes complementary architectural and autonomy frameworks to compare systems, expose capability gaps, and assess deployment risks.

  • Molecular discovery must translate desired functions into chemically valid, synthesizable, safe molecules supported by computational or experimental evidence.
  • Conventional molecular AI often solves isolated prediction, generation, or ranking tasks while users coordinate evidence, tools, failure repair, and subsequent decisions.
  • Molecular LLM agents use an LLM-centered controller to interpret goals, decompose actions, invoke chemistry tools, observe feedback, and revise plans.
  • Chemical agents require faithful handling of identity, connectivity, stereochemistry, geometry, units, and experimental conditions across representations and tools.Their observations span heuristic scores, simulations, spectra, assays, and hardware logs with different uncertainties and costs.
  • The field lacks a shared design framework, autonomy roadmap, and governance boundary, making systems difficult to compare and deployment risks difficult to assess.
  • The proposed framework combines four architectural components with an L1–L4 autonomy ladder spanning assistive workflows, adaptive computation, physical feedback, and scientific agendas.

2 Architecture Design of Molecular Agents

The architectural view organizes molecular LLM agents as interacting components that connect molecular state to decisions, scientific tools, observations, and optimization. A complementary autonomy taxonomy classifies systems by the feedback loop they reliably close without mandatory human intervention.

  • Architectural taxonomy: The architecture comprises molecular perception, an LLM-centered agent framework, domain-specific toolboxes, and learning and optimization.These components form a closed chain from molecular state to decision, action, observation, and optimization.
  • Architectural taxonomy: The perception layer routes molecular objects among strings, graphs, geometries, images, and structured records.
  • Architectural taxonomy: The agent framework plans actions, maintains memory, reflects on feedback, and may coordinate specialized agents.
  • Architectural taxonomy: The toolbox grounds decisions in databases, cheminformatics, simulation, synthesis, and experimental interfaces.
  • Scientific-autonomy taxonomy: Scientific autonomy is classified by the outermost feedback loop an agent can reliably close without mandatory human intervention, from L1 through L4.The rubric concerns decision responsibility and feedback source rather than tool count, reasoning steps, model components, or repeated trials.

2.1 Molecular Representation and Perception

Molecular agents require representations that preserve chemically relevant information while supporting LLM control and tool exchange. The section surveys strings, graphs, 3D structures, images, experimental records, tokenization strategies, and routing across representations.

  • Perception converts chemical objects into states that an LLM controller can read, edit, verify, and pass to tools.Representation choice determines which chemical facts are explicit and which actions are easy or brittle.
  • Representation substrates: SMILES and related strings provide compact, parser-friendly interfaces, but implicit graph structure makes topology editing and long-range reasoning fragile.
  • Representation substrates: SELFIES guarantees that every string maps to a chemically valid molecule, reducing unparsable outputs during repeated sampling, mutation, or optimization.The guarantee does not ensure stability, synthesizability, or task quality.
  • Representation substrates: Two-dimensional graphs expose atoms, bonds, and stereochemical features directly, while structured languages such as MolJSON and MoleCode make relational structure visible to LLMs.
  • Representation substrates: Three-dimensional representations encode spatial arrangement for geometry-dependent properties, whereas images support structure recognition from diagrams and multimodal records.Experimental records additionally preserve conditions, units, identifiers, uncertainty, and provenance as state-bearing observations.
  • Tokenization: Tokenization defines the molecular action granularity and influences error propagation, optimization dynamics, and the agent’s ability to reflect on edits.Fragment-level tokens can make chemically coherent edits more efficient, while validity constraints leave stability, synthesis feasibility, and task quality unresolved.
  • Representation routing: Molecular agents commonly route among representations by generating strings, parsing graphs, converting to 3D, querying databases, and rendering images for review.Memory and tool-use policies can include prior tool outputs, evidence, and task history in the perceived state.

2.2 Agent Framework

The agent framework is an LLM-centered controller that converts molecular goals into executable, feedback-responsive scientific trajectories. Its faculties include reasoning, planning, memory, reflection and self-correction, and multi-agent collaboration, but autonomy depends on whether evidence changes consequential decisions.

  • The agent framework selects tools or subagents, interprets outputs, and decides whether a trajectory should continue, revise, stop, or request approval.
  • Molecular controllers implement an observe, plan, act, and revise loop through reasoning, planning, memory, reflection and self-correction, and multi-agent collaboration.
  • Planning: Planning extends molecular agents from one-shot tool use to executable in-silico campaigns involving representations, tools, constraints, and feedback signals.
  • Planning: Interleaved reason-act planning alternates local reasoning, tool invocation, observation, and revision, while hierarchical designs separate higher-level planning from lower-level execution.
  • Memory: Memory supports longer workflows by preserving intermediate state and action history, but stored context alone does not raise autonomy unless it changes a later scientific decision.
  • Autonomy evidence: The survey records 96 faculty-level assignments: 39 at L1, 48 at L2, and 9 at L3; no surveyed paper meets the L4 criterion.

2.3 Domain-Specific Molecular Toolboxes

Domain-specific molecular toolboxes ground language-level plans in chemically executable operations over structures, databases, simulations, spectra, and laboratory protocols. Their value depends on reliable state passing, uncertainty handling, recovery, cost control, and calibration, not tool access alone.

  • Molecular toolboxes translate language-level plans into executable operations over structures, databases, simulations, spectra, and laboratory protocols.
  • Molecular perception: Entity resolution maps names, identifiers, targets, and assay descriptions into canonical molecular and biological representations for downstream tool use.
  • Molecular perception: Cheminformatics toolkits parse and validate molecular formats, preserve or check stereochemistry, compute descriptors, generate fingerprints, and convert file formats.
  • Scientific grounding: Databases and literature tools provide identity, structural, bioactivity, assay, reaction, availability, and experimental evidence with identifiers, units, curation history, and provenance.
  • Computational evaluation: Simulation and screening tools provide feedback ranging from low-cost descriptors and predictors to docking, quantum chemistry, molecular dynamics, and free-energy workflows.
  • Orchestration and safety: Robust orchestration must detect invalid inputs, lost stereochemistry, convergence failures, unit errors, database mismatches, and unsafe protocols, then retry, switch tools, or request human intervention.

2.4 Learning and Optimization

Learning and optimization concern what molecular agents update from experience and what candidates, actions, workflows, or experiments they improve. Evidence for autonomy requires feedback to change a later consequential scientific decision, with higher levels requiring physical feedback or cross-campaign agenda revision.

  • Learning updates policies, memories, models, acquisition strategies, tool-use behaviors, and workflows, whereas optimization improves decision variables against objectives.
  • Agency evidence requires observations, evaluations, or accumulated experience to change a later consequential scientific decision.
  • Optimization targets: Optimization targets have broadened from molecular candidates to candidate selection, tool invocation, evidence trust, resource allocation, failure repair, and stopping decisions.
  • Optimization targets and search spaces: Representation determines which molecular features can be changed, but agentic behavior depends on whether feedback alters later proposals, evaluations, tool choices, or selection decisions.
  • Multi-objective decision making: Multi-objective optimization balances conflicting properties through composite objectives, constraints, threshold filters, or Pareto selection, while safety and executability may be hard constraints.
  • Autonomy evidence: Computational optimizers support L2 only when evaluations affect later decisions, and experimental objectives support L3 only with physical execution and incorporated experimental feedback.
  • Autonomy evidence: Current evidence spans candidate and workflow optimization at L1 and L2 and bounded physical optimization and learning at L3, but not cross-campaign agenda revision at L4.

3 Evaluation and Benchmarking

Benchmark scores answer different scientific questions and are not directly comparable across panels. Evaluation should therefore pair local metrics with execution quality, feedback efficiency, coverage, and autonomy-relevant safeguards.

  • Benchmark interpretation: Figure 5 uses source-faithful, benchmark-specific result slices rather than a cross-benchmark aggregate.Molecular-dynamics or chemistry subsets are used when available; otherwise panels retain explicitly marked overall results and source metrics.
  • Benchmark interpretation: Benchmarks measure distinct targets, including procedural reasoning, executable tool sequences, research artifacts, and feedback-guided search.Similar percentages may represent answer accuracy, executable completion, reproduction success, or search efficiency.
  • Configuration sensitivity: Configuration-sensitive outcomes show that adding tools is not uniformly better across reported settings.The reported MolBench slice gives MolClaw-CC 81.1% accuracy versus 51.4% for vanilla agents and 45.1% for a standalone-LLM mean; a matched ChemCost comparison reports CTA@25 changing from 2.5–4.1% to 16.3% for GPT-4o-mini.
  • Evaluation design: Evaluation should combine partial credit with end-to-end success and stratification by horizon or difficulty because errors can accumulate across specification, execution, diagnosis, and repair.The displayed discussion frames partial credit as informative but insufficient on its own.
  • Evaluation design: Metrics should expose efficiency and coverage alongside success, including oracle-budget efficiency, expert-knowledge conditions, and unsupported agent–task pairs.MADE measures acceleration over a random baseline under a fixed oracle budget, while SciAgentArena’s heatmap includes strong, failed, and incompatible combinations.
  • Autonomy boundaries: Autonomy claims require evidence beyond headline scores, including loop adaptation, recovery, reproducibility, intermediate validity, operational scope, and mandatory human gates.The stated requirements differ across L1, L2, and L3 autonomy levels.

4 Evidence for Scientific Autonomy Levels

The framework defines scientific autonomy by the outermost feedback loop an agent reliably closes without mandatory human intervention, distinguishing computational adaptation, physical experimentation, and scientific-agenda control. Four levels and complementary reliability and human-involvement measures clarify boundaries between automation, competence, and decision authority.

  • Autonomy framework: Scientific autonomy is an operational rubric based on the outermost feedback loop reliably closed without mandatory human intervention, not on tool count or reasoning complexity.The rubric is presented as a survey framework rather than a universal field consensus.
  • Four autonomy levels: Computational autonomy requires tool or environment evidence to revise a consequential state or decision without an intermediate human decision.Fixed pipelines and syntax repair do not qualify, whereas evidence may change a candidate, hypothesis, plan, tool choice, or stopping rule.
  • Four autonomy levels: Physical autonomy additionally requires designing or adapting and executing an experiment while incorporating execution or measurement feedback into recovery, interpretation, state, or completion.A robotic platform replaying a fixed protocol is automation, whereas a single physical episode can qualify when observations change the agent’s actions or conclusions.
  • Four autonomy levels: Scientific-agenda autonomy additionally requires using accumulated evidence to formulate, prioritize, and pursue new questions, including evidence-conditioned transitions across task families.Generating candidate hypotheses alone is insufficient; the agent must select questions, allocate actions, and revise research direction from evidence.
  • Four autonomy levels: L1 systems remain assistive or fixed, L2 adapt computational scientific states from digital evidence, L3 complete feedback-aware physical workflows, and L4 control scientific-agenda evolution.The levels are hierarchical, with physical autonomy presupposing adaptive computational planning, analysis, or control, and scientific-agenda autonomy presupposing experimental and computational process management.
  • Evaluation and boundaries: Autonomy must be reported relative to a domain and complemented by safety, loop-reliability, and mandatory-human-intervention measures.These measures distinguish systems at the same nominal level that differ in reliability or dependence on human intervention, while autonomy and competence remain distinct.

5 Safety and Challenges

Molecular-agent safety depends on coupled architecture-wide failure modes, from corrupted molecular identity and unreliable tools to optimization that amplifies weaknesses. Risk increases with operational authority and physical integration, requiring governed execution, representation fidelity, verifiable reasoning, adaptive orchestration, and recovery.

  • System-wide safety: Risk is determined by operational authority, action reversibility, affected substances and equipment, and the consequences of erroneous decisions.Safety therefore concerns interactions among software, hardware, materials, people, and operating procedures.
  • System-wide safety: Molecular-agent failures can propagate across perception, controller design, tool interaction, and optimization, producing increasingly consequential behavior as action authority increases.The lifecycle view treats these components as coupled rather than isolated sources of failure.
  • Safety and governance: Adversarial inputs and jailbreaks can redirect molecular capabilities toward hazardous compounds, controlled substances, or chemical weapons.The same capabilities supporting beneficial drug discovery can be repurposed for dual-use objectives.
  • Safety and governance: Automated laboratory integration makes digital decision errors physically consequential, including through reagent sequencing, temperature control, and hazardous-material handling.Reported consequences include spills, fires, explosions, or personal injury, motivating rigorous controls and sustained human oversight.
  • Safety and governance: A governance contract should record verified molecular identity, permissions, approval gates, abort conditions, and auditable traces while evaluating unsafe episodes, escalations, recovery, costs, and human interventions.These are proposed reporting requirements rather than a claim that one control stack suffices for every laboratory.
  • Molecular perception: Representation conversions can silently alter stereochemistry, protonation, tautomeric state, atom mapping, conformers, or measurement context, requiring round-trip validation and persistent identity tracking.Strings, graphs, 3D structures, images, and spectra expose different molecular properties.
  • Agent framework: Current agents lack machine-checked chemical proofs and formal certificates, while text and SMILES remain interpretable but lossy compared with harder-to-audit graph, 3D, and latent states.Formal verification remains open for molecule edits, synthesis constraints, safety rules, and protocol preconditions.
  • Orchestration and evaluation: Tool use can help some chemistry tasks while hurting others, so controllers must adaptively choose among direct answers, retrieval, tools, specialists, and human approval.Evaluation should assess controller behavior and safety controls rather than relying solely on final task scores or refusal rates.

6 Future Directions

Future directions frame autonomy as a sequence of testable milestones, each producing artifacts and measurements that justify greater authority. The progression moves from verifiable molecular-state contracts and reproducible governed computational loops toward bounded physical episodes, iterative campaigns, and evidence-based scientific autonomy.

  • Milestone framework: Future milestones should produce artifacts and measurements that justify each expansion of agent authority rather than merely enumerate missing capabilities.The proposed progression treats autonomy as an evidence-building process.
  • Milestone 1: A verifiable molecular-state contract should specify typed identity, structure, conditions, units, uncertainty, provenance, and permitted edits, with round-trip tests across representations.The milestone is reached when declared conversions preserve identity and required metadata on a public test suite.
  • Milestone 2: A credible L2 loop should be replayable under matched evidence and resource budgets while recording tools, states, uncertainty, stopping decisions, escalations, recovery, cost, reproducibility, and human gates.Evaluation should compare against fixed-workflow and no-memory baselines.
  • Milestone 3: L3 progression requires a declared experimental domain with validated protocols, instrument limits, permissions, abort conditions, and measurement-quality checks.Physical execution should be evaluated as a whole software–hardware–material system.
  • Milestone 4: Reliable iterative L3 campaigns require measurements to select or change subsequent experiments and should be compared with fixed designs and established acquisition policies under matched budgets.Replicated measurements, calibration, sample efficiency, stopping behavior, and recovery are proposed evaluation targets.
  • Milestone 5: L4 should remain an evidence standard requiring cross-campaign evidence preservation, new-question formulation, resource allocation, task-family selection, and agenda revision when assumptions are contradicted.Evaluation should use multiple task families and prospective rather than retrospective evidence.

7 Conclusion

The survey presents molecular LLM agents through an architectural framework and a scientific-autonomy ladder, linking system design to demonstrated decision authority. Its analysis concludes that autonomy requires chemically faithful perception, verifiable grounding, feedback-aware decisions, provenance, realistic optimization, and explicit safety—not planning or tool use alone.

  • Framework: The framework combines molecular representation and perception, an LLM-centered agent framework, domain-specific toolboxes, and learning and optimization with four autonomy levels from L1 to L4.The autonomy view classifies agents by the outermost feedback loop they can reliably close without mandatory human intervention.
  • Conclusion: Planning, tool use, stronger language models, or multi-agent coordination alone do not establish scientific autonomy.Progress also requires chemically faithful perception, grounded and verifiable tools, feedback-aware decisions, persistent provenance, realistic optimization, and safety within an explicit operational domain.
Loading 2608.23104v1…