chemical validity checking

This work addresses the lack of systematic evaluation for symbolic, verifiable reasoning over molecular graph structures in current chemical large language models. Existing benchmarks often suffer from label bias or information leakage, hindering precise diagnosis of model shortcomings. To bridge this gap, we propose MolecularIQ—the first evaluation framework specifically designed for symbolic reasoning on molecular graphs. By integrating molecular graph representations, symbolic logic verification, and carefully structured reasoning tasks, MolecularIQ establishes a fine-grained benchmark that effectively uncovers systematic failure modes of contemporary models across specific molecular structures and reasoning challenges. This framework provides interpretable diagnostic insights and actionable directions for developing chemical large language models with faithful structural understanding capabilities.

Existing benchmarks for chemical reasoning evaluation focus solely on final answers, making it difficult to detect logical errors in intermediate reasoning steps. To address this limitation, this work proposes ChemCoTBench-V2, a novel benchmark that employs expert-designed structured templates to guide models in generating verifiable intermediate reasoning states. By integrating deterministic chemical rules, reference trajectory alignment, and oracle-verifiable state constraints, the framework enables low-cost, auditable process-level evaluation without requiring human or LLM-based adjudication. This approach is the first to support state-constraint verification and precise error localization in open-ended tasks, revealing a significant discrepancy between answer correctness and reasoning consistency across mainstream large language models. It further facilitates fine-grained model comparison and identification of the first erroneous step in reasoning trajectories.

Large language models (LLMs) frequently generate chemically inaccurate descriptions in molecular science and lack capabilities for error detection and interpretable correction, undermining their scientific reliability. Method: We introduce MolErr2Fix, the first fine-grained benchmark for diagnosing and correcting chemical reasoning errors, comprising 1,193 human-annotated instances covering structural and semantic errors. It features a novel quadruple annotation schema—error type, location, explanation, and correction—integrating structured error taxonomy with domain-specific chemical knowledge validation. The benchmark includes modular task designs and an open-source evaluation API. Contribution/Results: Comprehensive evaluation of mainstream LLMs reveals significant deficiencies in error localization and correction. MolErr2Fix effectively exposes model weaknesses, serving as a critical evaluation tool and data foundation for advancing trustworthy chemical reasoning in LLMs.

This work aims to achieve computable chemical synthesis—i.e., precise, code-driven control of reaction pathways on general-purpose reconfigurable hardware to automate the synthesis of any stable, isolable molecule while satisfying mass conservation, finite reaction time, and analytical detectability constraints. Method: We introduce the “chemputation” paradigm, modeling synthesis as graph transformations over the space Reagents × Process × Catalyst. We formally define and prove the Universal Chemical Synthesis Theorem, introduce the notion of “analytically reachable quantity” for molecules, and establish dynamic error correction as essential. Our end-to-end implementation integrates the Chemputer hardware platform, the chempiler compiler, assembly-theory–driven reachability analysis, and a real-time sensing feedback framework. Results: Experimental validation demonstrates that chemical reactions are intrinsically programmable, observable, and correctable graph operations. We identify reactor count and sensor bandwidth as critical scalability bottlenecks for chemputation.

Current large language models (LLMs) exhibit significant limitations in molecular structure reasoning, particularly in leveraging critical structural features—such as functional groups—to predict molecular properties. Method: We propose Molecular Structure Reasoning (MSR), the first framework to explicitly incorporate molecular structural sketches into LLM-based reasoning. MSR establishes a dual-path paradigm for reasoning over both known and unknown molecules, integrating SMILES and graph-based structural encodings, structure-aware prompt engineering, and a multi-stage reasoning chain to achieve interpretable, structure-to-language mapping. Contribution/Results: Evaluated across multiple molecular property prediction and functional group identification tasks, MSR consistently achieves substantial accuracy improvements over baseline LLMs. These results empirically validate that explicit structural modeling is both effective and essential for enhancing LLMs’ chemical understanding—bridging a key gap between symbolic chemical knowledge and neural language reasoning.

This work addresses the challenge that SMILES strings generated by large language models are often invalid due to violations of syntactic or chemical rules, and existing repair methods struggle to simultaneously preserve chemical validity and structural semantic fidelity. To overcome the limitations of conventional post-processing or single-point correction strategies, the authors propose a molecule-identity-preserving recovery paradigm that introduces a trajectory-level, multi-candidate exploration mechanism. This approach integrates RDKit-executable edits, molecular-aware similarity assessment, and proxy-guided multi-trajectory search to efficiently restore validity while maintaining the intended molecular semantics. Experiments on the ChEBI-20 invalid drafts demonstrate that the proposed method significantly outperforms current baselines across structural, exact-match, and string-level metrics, achieving state-of-the-art recovery performance.

This work addresses the inconsistency in predictions and explanations arising from multiple graph representations of the same molecule in molecular graph machine learning, which violates chemical identity. To resolve this, the authors propose InChIfied Invariants—strictly invariant features constructed at node, edge, and graph levels based on the International Chemical Identifier (InChI)—that inherently guarantee identical representations, predictions, and attributions for chemically equivalent graphs. Evaluation on the large-scale PubChem Substances dataset demonstrates that the method achieves consistent representations for 99.62% of chemically equivalent graph pairs, a dramatic improvement over the 0.35% consistency achieved by conventional Daylight invariants. Furthermore, it maintains competitive predictive performance on MoleculeNet benchmark tasks while significantly enhancing model interpretability and chemical plausibility.

This work addresses the pervasive hallucination problem in large chemical reasoning models, where structural descriptions in chain-of-thought (CoT) reasoning often diverge from actual molecular structures and are decoupled from answer correctness. Through systematic analysis of chemical CoT mechanisms, the study reveals for the first time that CoT exhibits both hallucinatory and functional characteristics, challenging the conventional view that treats CoT as reliable evidence of valid reasoning. Employing attribution analysis, SMILES-based draft perturbations, and cross-model comparisons across four model families and twelve chemical tasks, the authors demonstrate that correct answers frequently co-occur with inaccurate structural descriptions, while perturbing the molecular drafts significantly degrades output quality. These findings indicate that CoT plays a causal role in generation, underscoring the critical importance of process supervision in chemical reasoning.

Molecular property prediction is often compromised by label noise arising from experimental errors, database inconsistencies, or weak annotations, leading to biased model learning. To address this challenge, this work proposes MOLAR, a novel framework that explicitly disentangles latent clean properties from observed noisy labels. MOLAR leverages graph neural networks and text encoders to extract multimodal residual evidence and incorporates a class-aware noise channel to jointly infer label reliability and modality-specific confidence. Evaluated on both real-world noisy datasets and benchmarks with synthetically flipped labels, MOLAR consistently outperforms existing baselines. Visualization analyses further demonstrate its capability to effectively diagnose label quality and modality contributions, thereby enhancing model robustness and interpretability.

This work addresses the challenge of subtle, long-tail errors—such as incorrect molecular formulas, space groups, or formation energies—frequently generated by large language models in chemical and materials reasoning. The authors propose a hierarchical verification-and-correction framework that first extracts verifiable claims from model outputs, then efficiently validates them against authoritative databases and physical constraints. A gated correction loop, guided by condition-anchored prompting, enables precise error rectification. Evaluated across 528 conditional prompts, the method reduces molecular formula error rates from 22% to 4%, achieves correction success rates of 80–97%, and incurs only 31% of the retrieval cost of exhaustive search. Notably, it significantly enhances reasoning reliability on long-tail tasks, including isotope half-life prediction.