Research map / Multi-agent systems and games / LLM agents

IR-Agent: Expert-Inspired LLM Agents for Structure Elucidation from Infrared Spectra

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Agent data synthesis and behavior studies

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes IR-Agent, a multi-agent LLM framework (table-interpretation, retriever, and structure-elucidation experts) built on a Transformer IR-to-SMILES translator, for molecular structure elucidation from experimental IR spectra, with chemical information injected via prompts.

Problem

Molecular structure elucidation from infrared spectra is hard and existing ML methods use fixed input formats and cannot flexibly incorporate diverse chemical knowledge or expert analysis workflows.

Approach

Decompose IR interpretation into specialized LLM expert agents: a table-based peak-to-substructure expert, a spectral-retrieval expert providing global context from similar database spectra, and a final expert integrating both over Transformer-generated SMILES candidates; extra chemical info is added by appending sentences to prompts without retraining.

Key results

• Multi-agent IR-Agent with o3-mini reaches Top-1 0.103 and Top-10 0.216 versus 0.098 Top-1 for the standalone Transformer translator (Top-3 candidates).
• Multi-agent consistently beats single-agent across GPT-4o-mini, GPT-4o, and o3-mini backbones.
• Adding atom types as prompt info raises o3-mini-based multi-agent Top-1 from 0.103 to 0.127 and Top-10 to 0.278.

Surprising findings

• The Patch-Based Self-Attention Transformer attains higher Top-1 (0.129) than IR-Agent (0.103), though IR-Agent wins on other Top-K and Tanimoto metrics.
• Llama-3-8B fails to produce valid SMILES even after QLoRA fine-tuning on numeric IR tokens.

Limitations

• Top-1 accuracy is low in absolute terms (about 0.10), and the Transformer translator's candidates bound performance.
• Relies on proprietary OpenAI LLMs; LLM inference is slow and costly for high-throughput use.
• Workflow is deterministic rather than dynamic; the ReAct comparison shows agent-driven tool selection underperforms in this setting.

Experimental setup

Datasets: NIST IR spectra (9,052 experimental)

Benchmarks: NIST IR spectra test split (10%)

Metrics: Top-K exact match accuracy (InChI) Tanimoto similarity (MACCS, RDKit, Morgan)

Baselines: Transformer IR translator IR-Agent (single-agent) Patch-Based Self-Attention Transformer LLaMA-3-8B ReAct agent

Builds on: Transformer ChemCrow

Closest papers by meaning

• Analytical-Chemistry-Informed Transformer for Infrared Spectra Modeling (AAAI 2025 · similarity 0.84)
• Can LLMs Solve Molecule Puzzles? A Multimodal Benchmark for Molecular Structure Elucidation (NeurIPS 2024 · similarity 0.82)
• MOOSE-Chem: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses (ICLR 2025 · similarity 0.82)
• Can Agents Price a Reaction? Evaluating LLMs on Chemical Cost Reasoning (NeurIPS 2026 · similarity 0.81)
• Unraveling Molecular Structure: A Multimodal Spectroscopic Dataset for Chemistry (NeurIPS 2024 · similarity 0.80)
• From What to Why: A Multi-Agent System for Evidence-based Chemical Reaction Condition Reasoning (ICLR 2026 · similarity 0.80)