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

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

IR-Agent is a multi-agent LLM framework that elucidates full molecular structures (SMILES) from infrared spectra by decomposing the expert workflow into three specialized agents: one that interprets absorption tables for local substructure cues, one that retrieves globally similar spectra from a database, and one that integrates both analyses into a final ranked prediction. It improves over a standalone Transformer baseline on 9,052 experimental NIST spectra without requiring chemical formula input or model retraining to incorporate new knowledge types.

Problem

Existing ML methods for IR-spectrum-to-SMILES elucidation assume fixed input formats (often requiring ground-truth chemical formulas), cannot flexibly incorporate diverse chemical side-information (atom types, carbon counts, molecular scaffolds), and do not emulate the two-stage expert process of absorption-table analysis plus spectral-database consultation. Extending these models to new input types requires redesigning and retraining the entire architecture.

Method

The framework has two layers: a pretrained IR Spectra Translator (Transformer with beam search) that generates \(K\) initial SMILES candidates \(\mathcal{C} = \{s_1, \ldots, s_K\}\) from spectrum \(X \in \mathbb{R}^{1 \times L}\), seeding downstream agents with plausible structures.

Three specialized LLM agents then operate on \(\mathcal{C}\):

1. Table Interpretation (TI) Expert — calls the IR Peak Table Assigner tool, which detects peaks by comparing neighboring wavenumber absorbances and maps peak ranges to functional groups via an IR absorption table. The agent cross-validates table output against \(\mathcal{C}\), producing confidence-annotated local substructure hypotheses: $\(\mathcal{A}_\text{TI} = \text{TI Expert}(P_\text{TI}, \text{IR Peak Table Assigner}(X, T), \mathcal{C})\)$

2. Retriever (Ret) Expert — calls the IR Spectra Retriever, which computes cosine similarity between \(X\) and all database spectra, returning top-\(N\) matches with their SMILES. The agent extracts shared global structural features, weighting by similarity: $\(\mathcal{A}_\text{Ret} = \text{Ret Expert}(P_\text{Ret}, \text{IR Spectra Retriever}(X))\)$

3. Structure Elucidation (SE) Expert — synthesizes \(\mathcal{A}_\text{TI}\), \(\mathcal{A}_\text{Ret}\), and \(\mathcal{C}\) into a final ranked top-\(K\) SMILES list: $\(\mathcal{A}_\text{SE} = \text{SE Expert}(P_\text{SE}, \mathcal{A}_\text{TI}, \mathcal{A}_\text{Ret}, \mathcal{C})\)$

Additional chemical information (e.g., scaffold, atom types) is incorporated by appending a concise sentence to each agent's existing prompt—no new agents or retraining required.

Key Contributions

• First LLM-agent framework applied to molecular structure elucidation from IR spectra.
• Multi-agent decomposition into TI/Ret/SE experts with complementary local and global reasoning roles.
• Two task-specific tools: IR Peak Table Assigner (deterministic peak-to-substructure lookup) and IR Spectra Retriever (cosine-similarity spectral database search).
• Prompt-only extensibility: new chemical knowledge types added by appending to existing prompts, requiring no architectural change or retraining.
• Operates without ground-truth chemical formula, unlike prior Transformer-based approaches.

Results

Evaluated on 9,052 experimental NIST IR spectra (80/10/10% train/valid/test split), Top-\(K\) exact-match accuracy (SMILES converted to InChI for comparison), averaged over 3 runs:

• Standalone Transformer (IR Spectra Translator only): Top-1 = 0.098, Top-3 = 0.169, Top-5 = 0.176, Top-10 = 0.176
• Single-agent IR-Agent (GPT-4o-mini): Top-1 = 0.072, Top-3 = 0.118, Top-5 = 0.133, Top-10 = 0.157 — underperforms the Transformer in isolation
• The paper asserts the full multi-agent IR-Agent exceeds the Transformer baseline across Top-\(K\) metrics (complete multi-agent table truncated in provided text)
• Framework tested with GPT-4o-mini, GPT-4o, and o3-mini backbones; o3-mini also evaluated in a direct-generation (no agent) setting

Limitations

• Quantitative multi-agent results table was not fully provided; single-agent ablation actually underperforms the Transformer at Top-1, suggesting agent orchestration is critical and fragile without all three experts.
• Cosine similarity retrieval is spectrally shallow—it does not account for structurally diverse molecules with similar spectra or structurally similar molecules with different spectral profiles.
• IR spectra fundamentally lack stereochemical, stoichiometric, and molecular-weight information, capping theoretical ceiling; the framework inherits this physical constraint.
• Quality of all downstream reasoning is bounded by the Transformer translator's initial candidate pool (fixed \(K\) via beam search).
• Evaluation on experimental spectra only from NIST; generalization to other instrument types or sample preparations is not demonstrated.
• Prompt-based integration of chemical information is not rigorously ablated for cases with conflicting or noisy side-information.

Relevance to Harnesses / Meta-Harnesses

IR-Agent instantiates a domain-specific agentic harness with a canonical three-stage topology: a neural pre-processor seeds candidates → parallel specialist agents apply heterogeneous analysis tools → an integrator agent synthesizes outputs into a final answer. This pipeline-with-integration pattern is directly analogous to meta-harness designs that coordinate specialist sub-agents via a synthesis layer. The extensibility mechanism—injecting new knowledge sources through prompt appending rather than architectural redesign—demonstrates a lightweight "plug-in" pattern relevant to harness maintainability. The explicit tool abstraction (IR Peak Table Assigner, IR Spectra Retriever) reflects the broader principle of giving agents deterministic, domain-specific callable tools rather than relying on LLM hallucination for structured lookup, a design discipline transferable to any harness handling structured external knowledge.