Research map / Large language models / Agents, code and math
MolecularIQ: Characterizing Chemical Reasoning Capabilities Through Symbolic Verification on Molecular Graphs
Christoph Bartmann, Johannes Schimunek, Mykyta Ielanskyi, Philipp Seidl, Günter Klambauer, Sohvi Luukkonen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use a higher reasoning budget (High vs Med) (ablation) | GPT-OSS 20B · MolecularIQ overall | accuracy (%): 27.2 -> 35.7 | Improved |
| use a higher reasoning budget (High vs Med) (ablation) | GPT-OSS 120B · MolecularIQ overall | accuracy (%): 26.5 -> 47.5 | Improved |
| chemistry fine-tuning of general LLMs (ablation) | chemistry-specialized LLMs · MolecularIQ overall | accuracy | Hurt |
| chemistry fine-tuning (ChemDFM-R) (ablation) | ChemDFM-R · MolecularIQ overall | accuracy | No clear effect |
| chemistry tuning effect on output format (ablation) | chemistry-tuned LLMs · MolecularIQ | type-validity rate: 18 percentage points lower on average | Hurt |
| randomize, kekulize or relabel rings in SMILES (ablation) | top-10 LLMs · MolecularIQ counting and indexing | accuracy | Hurt |
| increase multitask load (ablation) | top LLMs · MolecularIQ counting | accuracy | Hurt |
| increase multitask load vs independence baseline (ablation) | top LLMs · MolecularIQ | success rate: observed n-task success exceeds p_single^n | Improved |
| increase molecular Bertz complexity (ablation) | top LLMs · MolecularIQ counting | accuracy | Hurt |
| switch from counting to indexing for same property (ablation) | top LLMs · MolecularIQ | accuracy: decrease of ~5-30% depending on model | Hurt |
| lower prevalence of constraint combination or 3+ constraints (ablation) | top LLMs · MolecularIQ generation | accuracy | Hurt |
| use 8 rollouts instead of 3 (ablation) | subset of models · MolecularIQ | accuracy | No clear effect |
Closest papers by meaning
- Beyond Chemical QA: Evaluating LLM's Chemical Reasoning with Modular Chemical Operations (NeurIPS 2025 · similarity 0.93)
- Can LLMs Solve Molecule Puzzles? A Multimodal Benchmark for Molecular Structure Elucidation (NeurIPS 2024 · similarity 0.91)
- FGBench: A Dataset and Benchmark for Molecular Property Reasoning at Functional Group-Level in Large Language Models (NeurIPS 2025 · similarity 0.90)
- Mol-LLaMA: Towards General Understanding of Molecules in Large Molecular Language Model (NeurIPS 2025 · similarity 0.89)
- Training a Scientific Reasoning Model for Chemistry (NeurIPS 2025 · similarity 0.89)
- Structured Chemistry Reasoning with Large Language Models (ICML 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
