cs.CLAug 24, 2026
SaveMolecular LLM Agents: From Architectural Design to Scientific Autonomy
Organizations: The Hong Kong Polytechnic University Hong Kong, China · Shanghai AI Lab Shanghai, China · National University of Singapore Singapore · Shanghai Jiao Tong University Shanghai, China
Abstract
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.
Figures & tables
Figure 1. End-to-end workflow of a molecular LLM agent. Starting from a scientific goal and task-specific constraints, the agent gathers multi-modal molecular evidence, plans and executes chemistry-tool operations, selects candidates, and verifies them through simulation or wet-lab experiments. Verification results are then fed back to refine the molecular design and the agent’s subsequent actions. Five numbered stages connect a scientific goal, multi-modal data collection, agent planning and tool use, computational or wet-lab verification, and feedback-driven optimization.
Figure 2. Two complementary taxonomies for molecular LLM agents. The architectural taxonomy (left) organizes agent systems around molecular representation and perception, an LLM-centered agent framework, domain-specific molecular toolboxes, and learning and optimization, together with multi-agent collaboration and cross-cutting deployment requirements. The scientific-autonomy taxonomy (right) classifies systems by the outermost feedback loop they can reliably close without mandatory human intervention, ranging from L1 assistive or fixed workflows to L4 scientific-agenda agents. The left side shows four architectural components, with agent collaboration and safety surrounding a molecular LLM agent. The right side shows four increasing autonomy levels from assistive or fixed workflows to scientific-agenda discovery.
| Substrate | Typical form | What it makes explicit | Agent-level use and main risk |
|---|---|---|---|
| Line notation | SMILES, InChI | Compact identity and parser-compatible connectivity | Fast generation, search, and tool calls; topology and stereochemistry must be recovered from a brittle sequence. |
| Robust or fragment string | SELFIES, Group SELFIES, SAFE, molecule-native fragments | Validity constraints or chemically meaningful blocks | Useful for repeated editing and optimization; validity does not ensure synthesizability or property quality. |
| Graph or structured text | Molecular graph, MolJSON, MoleCode | Atoms, bonds, local topology, explicit node and edge identities | Better for exact edits and graph reasoning; requires graph encoders or verbose structured contexts. |
| 3D geometry | Conformers, coordinates, pockets, docking poses | Distances, angles, chirality, spatial contacts, physical state | Necessary for binding, quantum, and simulation tasks; expensive and conformer-sensitive. |
| Image | Skeletal diagram or scanned chemical figure | Human-readable structural drawing and visual stereochemistry cues | Connects agents to papers and notebooks; must be recognized into graph or string form before most tools can act. |
| Experimental record | Spectra, chromatograms, assays, time series, protocol events, robot logs | Measured response, conditions, units, acquisition context, and execution state | Grounds interpretation and laboratory control; sample mismatch, lost metadata, or overconfident peak assignment can corrupt later decisions. |
Table 1. Common molecular and experimental perceptual substrates for LLM-based molecular agents.
Figure 3. Two views of the 58 surveyed papers by highest demonstrated autonomy level. The faculty heatmap counts 96 paper–faculty assignments, while the yearly stacked bars count each paper once; hatching marks the L4 evidence gap. Left: a faculty-by-autonomy heatmap for reasoning, planning, memory, reflection and self-correction, and multi-agent collaboration. Right: yearly stacked bars in which each paper is counted once at its highest demonstrated autonomy level. The L4 column is hatched because no surveyed paper qualifies.
Figure 4. Hierarchical learning and optimization in molecular LLM agents. Optimization may target candidates, actions, workflows, or experiments, while learning updates policies, memories, models, strategies, and reusable workflows from experience. These mechanisms can operate at multiple autonomy levels: from fixed computational procedures (L1), through feedback-driven computational decisions (L2) and physical experimentation (L3), to cross-campaign learning that revises higher-level scientific objectives or agendas (L4). Learning and optimization operate across multiple levels of molecular-agent autonomy, from fixed computational procedures to feedback-driven computational decisions, physical experimentation, and cross-campaign learning.
| Benchmark | Scientific scope | Agent interaction or loop | Primary evaluation signal |
|---|---|---|---|
| MolViBench ( Li et al., 2026e ) | 358 molecular tasks, 12 workflows, and five difficulty levels | Generates executable programs for multi-step molecular workflows | Program correctness, execution success, and degradation with workflow complexity |
| MolBench ( Zhang et al., 2026d ) | Molecular screening, optimization, and end-to-end discovery challenges spanning 8–50+ tool calls | Connects filtering, affinity estimation, molecular editing, and workflow execution | Subtask quality and end-to-end challenge completion |
| ChemCost ( Wu et al., 2026 ) | 1,427 reactions, 2,261 chemicals, and 230,775 price quotes | Reasons over reaction components and noisy procurement information | Cost-error tolerance, robustness to noise, and component-level attribution |
| MDGym ( Kumar et al., 2026 ) | 169 molecular simulations and 303 tasks across two MD engines | Configures, runs, diagnoses, and repairs simulation workflows | Executable task success by difficulty, engine, and failure type |
| ChemReason-Bench ( Zhang et al., 2026c ) | 7,306 tasks instantiated from 500 organic reactions in six formats | Produces or validates ordered, condition-aware, schema-constrained procedure steps | Ordering, constraint validation, entity-role grounding, and parseable completion |
| Corral ( Alampara et al., 2025 ) | Four chemistry and materials environments: MD, ML, catalysis, and spectroscopy | Compares tool-calling and ReAct-style agents in executable expert-designed tasks | Task success, tool-use failures, and sensitivity to task–tool alignment |
Table 2. Landscape of representative molecular and scientific-agent benchmarks. The table emphasizes evaluated interaction and evidence rather than ranking benchmark-specific scores, which are not directly comparable.
Figure 5. Compact reported result slices for all 12 benchmarks in Table 2 : (a) MolViBench IR ( Li et al., 2026e ) ; (b) MolBench ( Zhang et al., 2026d ) ; (c) ChemCost ( Wu et al., 2026 ) ; (d) MDGym ( Kumar et al., 2026 ) ; (e) ChemReason-Bench ( Zhang et al., 2026c ) ; (f) MADE ( Malik et al., 2026 ) ; (g) Corral MD ( Alampara et al., 2025 ) ; (h) ScienceAgentBench ( Chen et al., 2025 ) ; (i) SciAgentBench chemistry ( Shen et al., 2026 ) ; (j) SciCode ( Tian et al., 2024 ) ; (k) CORE-Bench ( Siegel et al., 2024 ) ; and (l) SciAgentArena molecule optimization ( Liu et al., 2026a ) . Asterisks mark overall results used where no numeric molecular subgroup is published. In (l), darker cells denote higher task scores on the source’s 0–1 scale and hatching denotes incompatible agent–task combinations. Axes retain source metrics and are comparable only within panels. Twelve compact small-multiple panels provide one reported result slice for every molecular or scientific-agent benchmark in Table 2. The final panel is a ten-task molecule-optimization heatmap with hatched incompatible agent–task combinations. Three panels are explicitly marked as overall results because their sources do not publish numeric molecular subgroups.
| Level | Closed loop | Operational criterion |
|---|---|---|
| L1 | None | The agent retrieves information, invokes tools, provides recommendations, or executes a fixed workflow, but evidence does not autonomously revise a consequential scientific decision. |
| L2 | Adaptive computational | Given a computational objective, the agent uses digital evidence to revise a candidate, hypothesis, plan, tool choice, or stopping decision within a digital or simulated environment. |
| L3 | Physical workflow | Given a high-level experimental objective, the agent designs and executes a physical experiment and incorporates execution or measurement feedback without a mandatory human gate during ordinary operation. |
| L4 | Scientific agenda | The agent formulates and prioritizes new scientific questions, selects suitable computational and experimental task families, and updates its research agenda from accumulated evidence. |
Table 3. Levels of scientific autonomy for molecular LLM agents.
| System | Scientific setting | Strongest demonstrated feedback or action | Assigned level |
|---|---|---|---|
| ChemAgent ( Tang et al., 2025 ) | Chemical reasoning | Updates task memory to improve bounded reasoning, without closing an external scientific feedback loop | L1: assistive memory |
| CACTUS ( McNaughton et al., 2024 ) | Chemistry tool use | Executes a predefined tool-mediated problem-solving workflow | L1: fixed workflow |
| Chemist-X ( Chen et al., 2023 ) | Reaction conditions | Retrieves evidence and recommends conditions; downstream experimental revision remains external | L1: recommendation |
| ChemCrow ( M. Bran et al., 2024 ) | General chemistry | Uses computational chemistry tools and their outputs to revise a tool-mediated solution trajectory | L2: computational loop |
| ChatDrug ( Liu et al., 2024a ) | Molecular editing | Uses retrieval and domain feedback to revise subsequent molecular edits | L2: feedback-driven edit |
| ChemReasoner ( Sprueill et al., 2024 ) | Catalyst and molecular search | Uses quantum-chemical rewards to redirect search over candidate reasoning paths | L2: simulation feedback |
Table 4. Representative molecular-agent systems classified by the strongest feedback loop demonstrated in their reported evaluation. The examples are illustrative rather than exhaustive; level is assigned from evidence-conditioned behavior, not from the number of tools or agents.
Figure 6. Challenges and bottlenecks in the lifecycle of a molecular agent. The diagram maps vulnerabilities across perception, controller design, tool and environment interaction, and optimization; arrows indicate how an upstream error can propagate into more consequential system-level behavior as action authority increases. A four-layer molecular-agent lifecycle showing risks in molecular perception, agent control, tool and environment interaction, and optimization. Arrows indicate that failures can propagate across layers and grow more consequential as the system gains action authority.
More in this topic
LLM Agents
Large language model agents use language models to plan, use tools, and act across multi-step tasks and environments.
See all 705 papers