Blog · arXiv Analysis · Last reviewed July 2, 2026

The Reaction Rule Becomes the Verification Loop

Daniel Armstrong, Maarten Dobbelaere, Valentas Olikauskas, Helena Avila, Octavian Susanu, Jérôme Waser, and Philippe Schwaller's July 2026 arXiv paper turns reaction classification into a useful test case for agentic science: can language models generate symbolic rules that become reliable only after deterministic verification?

For this essay, a reaction-rule receipt is the record that binds a generated chemistry rule to its seed reactions, taxonomy path, SMIRKS pattern, validation examples, false-positive tests, ordering decision, abstention behavior, fallback route, expert checks, and reproducibility settings.

The Claim

The paper, arXiv:2607.01061 [cs.AI; cs.CL], was submitted on July 1, 2026. arXiv lists the title as Agentic generation of verifiable rules for deterministic, self-expanding reaction classification.

The headline claim is that a multi-agent LLM pipeline can classify reactions and generate computable rules across 665,901 US patent reactions, expanding a standard reaction taxonomy from 68 seed classes to 14,073 classes without manual curation.

The useful governance claim is narrower: the LLM is not trusted because it writes plausible chemistry prose. Its outputs become useful when they are converted into symbolic rules, tested against a corpus, ordered to control false positives, and paired with abstention and fallback mechanisms.

The Paper Frame

Computer-assisted synthesis planning needs reaction rules. A target molecule is decomposed into accessible precursors by applying transformations that have names, scopes, constraints, and computable encodings. The paper focuses on SMIRKS rules, which encode graph transformations around a reaction center and local atomic environment.

The problem is long-tailed chemistry. Manual rule libraries can be precise, but they freeze the coverage of known taxonomies and require expert curation. Neural classifiers can be broad, but they usually return probabilities, and a silent misclassification can propagate into route planning, molecular design, or downstream analysis.

The paper's architecture separates two jobs. A deterministic symbolic layer handles known chemistry quickly. A generative LLM layer is invoked when the symbolic layer cannot classify the reaction, proposing taxonomy extensions and new rules that must then face verification.

Multi-Agent Pipeline

The classification pipeline uses specialized agents rather than one monolithic prompt. Reactions are grouped into template-level cohorts so mechanistically equivalent transformations can receive identical labels. A hierarchy agent maps each cohort to a coarse class. A detailed agent refines the assignment within the taxonomy. A verifier agent audits the proposed label.

If the existing taxonomy is insufficient, a generator agent proposes a new taxonomy entry with a code, name, and hierarchical position. An aggregator then integrates proposals and dynamic mappings so later cohorts can reuse newly added entries rather than reinventing them.

That is the right shape for a scientific agent: generation, verification, memory of accepted changes, and conflict handling are separate stages. The paper reports that Gemini 3 Flash handled lower-cost classification stages, while Gemini 3 Pro handled verification, generation, and aggregation, with low-temperature inference and resumable checkpointing.

Rule Generation

For rule generation, the agent writes generalized SMIRKS patterns from class examples. The rule is then tested on worked examples and against other classes. A pattern that matches reactions outside its intended class becomes a false-positive problem, not a successful generation.

The refinement loop asks the LLM to narrow or split patterns while preserving true positives. It includes rollback: if a refinement loses too much recall or fails validation, the older rule can be kept. This matters because a rule can become safer by becoming more specific, but it can also become less useful if it stops covering the class it was meant to describe.

The paper reports 4,964 generalized SMIRKS patterns. Testing 665,901 reactions against those patterns revealed a false-positive graph, and the ordering algorithm eliminated 95.8 percent of false positives; the remaining 4.2 percent came from mutual relationships inside 215 strongly connected components.

Deterministic Classifier

The resulting ReactionClassifier supports two modes. Hybrid strict uses a lightweight MLP gate trained on differential reaction fingerprints, then applies exact templates within the predicted branch. Generalized SMIRKS skips the gate and applies the ordered database of 4,964 generalized patterns globally, returning the first matching class.

On in-distribution held-out USPTO splits, the paper reports strong tier-3 accuracy for Hybrid strict: 97.9 percent overall on the frequency-weighted split and 94.1 percent on the template-balanced split, with covered accuracy of 98.7 percent and 97.5 percent. Generalized SMIRKS is lower overall at tier 3, 85.9 percent and 75.0 percent, but its covered accuracy is closer, 90.1 percent and 86.8 percent.

The ordering step is not cosmetic. Applying the same 4,964-pattern set in arbitrary order lowers tier-3 covered accuracy from 90.1 percent to 70.5 percent on one split and from 86.8 percent to 77.5 percent on the other. The rule database only works because the conflicts between rules are themselves modeled.

Fallback and Growth

The out-of-distribution tests show why a living taxonomy matters. On CRD-2025, a corpus of 9,296 single-reaction-center academic reactions from 2025 and later, Hybrid strict covers 64.0 percent and Generalized SMIRKS covers 68.3 percent, while NameRXN covers 89.3 percent. On the RingBreaker corpus, the paper reports 80.6 percent coverage for Hybrid strict and 77.3 percent for Generalized SMIRKS, exceeding NameRXN's 73.7 percent and Rxn-INSIGHT's 6.9 percent.

When the deterministic layer abstains on CRD-2025, the LLM fallback is invoked. The paper reports final classifications for 2,911 of 2,990 processed single-center abstention reactions. Combined with Hybrid strict coverage, the two-layer architecture classifies 8,861 of 9,296 CRD-2025 reactions, or 95.3 percent.

The fallback adds 1,942 hierarchy entries across single-center and multicenter abstention sets: 135 at L3, 687 at L4, 836 at L5, and 284 at L6/L7. The authors read these additions as underrepresented established methods, not necessarily newly invented chemistry. Examples include alpha-selenenylation of carbonyl compounds, C-H functionalization, and reductive or radical cross-coupling families.

Expert Check

The paper also includes blind expert-chemist validation on 69 reactions from CRD-2025, each graded by three raters. The LLM-derived taxonomy is marked fully Correct in 82.6 percent of cases, compared with 61.8 percent for NameRXN. NameRXN has more Acceptable but imprecise labels, while outright wrong rates are similar: 8.2 percent for the LLM taxonomy and 7.2 percent for NameRXN.

The authors report that the LLM assignment is strictly preferred over NameRXN in 47 of 61 discordant reactions. That supports the paper's precision claim, but it also exposes a limit: fallback labels for previously unclassified reactions are less stable, with 66.7 percent Correct in the new pool versus 90.1 percent in the original pool.

The right reading is not "the LLM replaces chemists." It is that the agent can generate a candidate symbolic infrastructure that becomes inspectable enough for chemists to validate, contest, and improve.

Governance Reading

This paper is a concrete example of agentic science done with a visible trust boundary. The agent does not merely answer a chemistry question. It emits taxonomy entries and executable rules that can reshape a database used by future classifiers.

That is powerful because it converts language-model output into symbolic infrastructure. It is risky for the same reason. A bad rule is not a one-off hallucination; it can become a reusable label, a branch in a hierarchy, a first-match decision, or a false-positive interaction that quietly affects later chemistry.

The governance object is therefore the verification loop. For each generated rule, the record must show the seed corpus, matched examples, missed examples, false positives, ordering dependence, abstentions, conflicts, fallback additions, expert review, software versions, random seed, and deployment boundary.

Reaction-Rule Receipts

A useful reaction-rule receipt should include the source reactions, template cohort, atom mapping, seed taxonomy path, agent outputs, verifier decision, generated hierarchy line, SMIRKS strings, validation fold, held-out fold, true-positive recall, false-positive edges, ordering position, strongly connected component membership, rollback history, and final classifier mode.

For deployment, the receipt should also include abstention conditions. The safest symbolic classifier is not the one that always answers. It is the one that knows when no rule fired, when multiple rules conflict, when the MLP gate and symbolic match disagree, and when a new class is only a proposal.

For scientific governance, the receipt should separate corpus verification from experimental verification. A rule that correctly classifies recorded patent or literature reactions is not proof that a proposed synthesis route is feasible, safe, high yielding, scalable, or acceptable under lab and regulatory constraints.

Limits

The main limitation is domain coverage. The pipeline is tuned to the USPTO corpus, which overrepresents pharmaceutical chemistry and underrepresents organometallic, materials, and process-scale transformations. Extending the method to electrochemistry, flow chemistry, biocatalysis, or other specialized corpora still requires careful curation.

The seed hierarchy matters. The system grows an existing taxonomy; it does not reliably construct one from nothing. It also remains retrospective in this paper: it organizes and generalizes recorded chemistry rather than proving new chemistry in the lab.

The expert-validation sample is useful but modest, and inter-rater agreement is moderate. The system should be read as a promising architecture for self-expanding symbolic classifiers, not as a final guarantee that every generated rule is chemically authoritative.

Source Discipline

This page treats the arXiv abstract, arXiv HTML, and PDF as the source set. The extracted PDF was used for exact methods and result figures where the arXiv HTML view hides some math-rendered values.

The page does not claim independent reproduction, experimental synthesis validation, or review of a released code repository. It reads the reported numbers as author-reported evidence and keeps the trust claim attached to verification artifacts rather than to LLM fluency.

Sources

Return to Blog