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

🕒 Published (v1): 2025-08-22 06:07 UTC · Source: Arxiv · 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 for full molecular structure elucidation from infrared spectra, modeled on expert analytical workflows. It decomposes the task across three specialized agents — table interpretation, spectral retrieval, and integrative structure elucidation — each contributing complementary local or global structural evidence. The framework requires no retraining to incorporate new chemical information; new knowledge is added by appending text to agent prompts.

Problem

Existing ML methods for IR-based structure elucidation (Transformer, RL, CNN) use fixed input formats and require architectural redesign to accommodate new chemical information (atom types, carbon count, scaffold). Single-agent LLMs attempting all sub-tasks simultaneously suffer degraded performance because local absorption features can be overwhelmed by global retrieval context, and handling high-dimensional numerical IR data alongside tables and SMILES candidates within one context window leads to incomplete reasoning.

Method

The pipeline has four stages:

1. IR Spectra Translator: A Transformer model generates an initial candidate pool \(\mathcal{C} = \{s_1, \ldots, s_K\}\) via beam search decoding from the raw spectrum \(X \in \mathbb{R}^{1 \times L}\).

2. Table Interpretation (TI) Expert: An LLM agent uses the IR Peak Table Assigner tool — which compares neighboring wavenumber absorbances to detect peaks, then looks them up in an IR absorption table — to extract local substructure hypotheses. The agent cross-references these with \(\mathcal{C}\) and assigns per-substructure confidence levels: $\(\mathcal{A}_{TI} = \text{TI Expert}(P_{TI}, \text{PeakAssigner}(X, T), \mathcal{C})\)$

3. Retriever (Ret) Expert: An LLM agent uses the IR Spectra Retriever tool, which computes cosine similarity between \(X\) and all database spectra, returns top-\(N\) SMILES with similarity scores, and extracts shared global structural features (weighting by similarity): $\(\mathcal{A}_{Ret} = \text{Ret Expert}(P_{Ret}, \text{Retriever}(X))\)$

4. Structure Elucidation (SE) Expert: Integrates \(\mathcal{A}_{TI}\), \(\mathcal{A}_{Ret}\), and \(\mathcal{C}\) to produce a ranked top-\(K\) SMILES list: $\(\mathcal{A}_{SE} = \text{SE Expert}(P_{SE}, \mathcal{A}_{TI}, \mathcal{A}_{Ret}, \mathcal{C})\)$

Additional chemical information (atom types, carbon count, molecular scaffold) 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 for molecular structure elucidation from IR spectra.
• Three-expert multi-agent decomposition (local table-guided + global retrieval-guided + integrative) that outperforms a single-agent handling all sub-tasks simultaneously.
• Tool-augmented agents: IR Peak Table Assigner for precise numerical peak extraction; IR Spectra Retriever for cosine-similarity-based spectral database search.
• Extensible architecture: new chemical modalities integrated via prompt modification, bypassing architectural redesign or retraining.

Results

Evaluated on 9,052 experimental IR spectra from the NIST database (80/10/10% train/valid/test split); metric is Top-K exact match accuracy (SMILES compared via InChI).

• Transformer baseline (standalone): Top-1 = 0.098, Top-3 = 0.169, Top-5 = 0.176, Top-10 = 0.176
• IR-Agent single-agent (all sub-tasks in one LLM): Top-1 = 0.072, Top-3 = 0.118, Top-5 = 0.133, Top-10 = 0.157
• IR-Agent multi-agent: stated to exceed both baselines on Top-K accuracy (exact numbers for multi-agent rows were truncated in the provided text)
• Multi-agent variants tested with GPT-4o-mini, GPT-4o, and o3-mini backbones; an o3-mini direct-ranking ablation is also evaluated.

Limitations

• Exact multi-agent numerical results are not fully reproduced in the available excerpt, limiting direct comparison.
• Chemical formula is excluded from the base setting; some prior methods (Alberts et al. 2024a; Wu et al. 2025) use ground-truth formulas, making comparisons only partially aligned.
• IR spectra fundamentally lack stereochemical, stoichiometric, and molecular-weight information, capping the ceiling of IR-only elucidation regardless of framework design.
• Retrieval uses raw cosine similarity on absorbance vectors — no learned embedding — which may underperform on noisy or phase-shifted experimental spectra.
• Framework performance depends on quality and coverage of the IR absorption table and retrieval database; poor database coverage degrades the Ret Expert.

Relevance to Harnesses / Meta-Harnesses

IR-Agent is a concrete scientific harness instantiation: it wraps a pre-trained Transformer translator, two domain-specific tools (Peak Table Assigner, Spectra Retriever), and external knowledge bases into a structured multi-agent pipeline with defined inter-agent interfaces. The extensibility mechanism — injecting new knowledge sources into existing agents via prompt appending, without retraining — is a lightweight meta-harness design pattern relevant to anyone building harnesses that must adapt to heterogeneous or evolving input modalities. The paper also provides a direct empirical validation of a core harness design principle: decomposing a complex pipeline into specialized sub-agents outperforms a monolithic single-agent that handles all tasks in one context window, which is a useful data point for harness architecture decisions.