NatureLM: Unified Scientific Language Model

• NatureLM is a unified sequence-based foundation model that tokenizes diverse scientific entities—such as molecules, proteins, DNA/RNA, materials, and text—for cross-domain discovery.
• It employs modality delimiters and a two-stage training process on 143B tokens, integrating specialized scientific vocabularies and leveraging pretrained LLM backbones.
• The model expands into bioacoustics via NatureLM-audio, achieving state-of-the-art zero-shot classification and detection for species identification and ecological tasks.

NatureLM, or Nature LLM, denotes a line of foundation models built on the premise that many scientific and ecological entities can be represented as sequences constituting a shared “language of nature.” In its sequence-based formulation, NatureLM is a science foundation model spanning text, small molecules, proteins, DNA, RNA, materials, and even cells, with the explicit goal of enabling scientific discovery rather than only scientific question answering (Xia et al., 11 Feb 2025). In a closely related multimodal formulation, NatureLM-audio extends the same general program to bioacoustics by pairing audio with language for species identification, detection, captioning, and other biodiversity tasks (Robinson et al., 2024).

1. Conceptual basis and scope

NatureLM was proposed against a background in which scientific foundation models were typically trained in isolation within a single domain: protein-only models for proteins, chemistry-only models for molecules, DNA/RNA models for genomics, materials models for crystals, and text-only science LLMs for literature. The stated limitation of that regime is that such systems do not naturally connect across domains. NatureLM addresses this by treating heterogeneous scientific objects as sequences that can be tokenized and modeled jointly, thereby enabling cross-domain scientific understanding, text-guided generation and optimization, and discovery-oriented tasks (Xia et al., 11 Feb 2025).

The core representational move is to cast scientific entities into a unified autoregressive interface. Small molecules are represented as SMILES, proteins as amino-acid FASTA sequences, DNA and RNA as nucleotide sequences, materials as flattened sequences including composition, space group, and, for 3D prediction, coordinates, while text remains ordinary natural language. The model uses modality delimiters such as mol ... /mol, protein ... /protein, dna ... /dna, rna ... /rna, and material ... /material (Xia et al., 11 Feb 2025). This design is intended to let one model learn relations such as text-to-molecule, protein-to-molecule, DNA-to-RNA, and composition-to-crystal structure.

A second usage of the name appears in bioacoustics. NatureLM-audio is described as the first audio-language foundation model specifically designed for bioacoustics and as the first concrete instantiation of a broader NatureLM vision that could extend beyond audio to modalities such as motion and images (Robinson et al., 2024). Accordingly, “NatureLM” is best understood not as a single frozen artifact but as a modeling program: one that treats diverse natural-science signals as parts of a common representational language.

2. Architecture, tokenization, and training of the sequence model

The sequence-based NatureLM family comprises three reported versions: NatureLM-1B, NatureLM-8B, and NatureLM-8x7B, the last being a Mixture-of-Experts model with total 46.7B parameters. The 8B model is initialized from Llama 3 8B, the 8x7B model from Mixtral 8x7B, and the family is trained by continuing pre-training from strong existing LLM backbones while adding scientific vocabularies (Xia et al., 11 Feb 2025).

Pre-training uses 143B tokens spanning text, small molecules, proteins, DNA, RNA, materials, and cross-domain pairs. The reported token distribution is text 14.4B, small molecule 4.2B, protein 65.2B, DNA 19.8B, RNA 27.5B, material 0.02B, and cross-domain 12.7B. The model is therefore heavily weighted toward scientific sequences, with only about 10% pure text (Xia et al., 11 Feb 2025). The pre-training procedure is explicitly two-stage. In Stage 1, only the newly added scientific tokens are trained while original model parameters are frozen. In Stage 2, the whole model is trained jointly. This staging is described as a way to avoid destabilizing the pretrained LLM.

Post-training uses about 5.1 million instruction-response pairs across more than 60 task categories, including molecule optimization, protein design, guide RNA design, material generation, and general text instructions. At inference time, the implementation uses vLLM with PagedAttention and Selective Batching; reported throughput for the 8x7B model is about 525 tokens/sec on 2 NVIDIA A100s with BF16 (Xia et al., 11 Feb 2025).

Scaling is a central empirical theme. Across 22 evaluated task categories, 18 improve with larger model size, validation loss decreases consistently from 1B to 8B to 8x7B, and the largest model is usually best overall (Xia et al., 11 Feb 2025). Within the paper’s framing, this is evidence that joint scientific-sequence pre-training and increased capacity improve validity, task adherence, cross-domain generalization, and generation quality.

3. Generative and discovery-oriented capabilities

NatureLM’s principal claim is that a unified sequence model can support text-instructed generation, optimization, and translation across multiple scientific domains. In small-molecule generation, NatureLM-8x7B reports 98.8% validity and 98.8% uniqueness for unconditional SMILES generation, compared with 77.9%/35.1% for Llama 3 (8B), 72.6%/35.1% for Mixtral (8x7B), and 99.6%/54.6% for GPT-4 (Xia et al., 11 Feb 2025). For property-to-molecule generation, the paper evaluates QED, HBA, HBD, FSP3, RotBonds, and TPSA using Spearman correlation and a correctness metric, stating that for some properties, such as TPSA, the Spearman correlation exceeds 0.8 and that the model can handle multiple properties jointly.

Drug-discovery tasks form a major application cluster. For target-aware hit compound generation, NatureLM reports stronger docking and QED than Pocket2Mol, TargetDiff, and TamGen; the 8x7B model reaches Vina $-6.95$, QED 0.62, SAS 0.82, Diversity 0.75, LogP$\in[0,5]$ 0.84, and Ro5 0.99 (Xia et al., 11 Feb 2025). In binding-affinity optimization on 12 held-out targets, more than 90% of generated molecules are novel with respect to ChEMBL, and one highlighted case improves affinity from 410 nM to 53 nM. For ADMET and CYP optimization, the 8B model reports BBBP 0.549 and CYP Avg 0.865. In retrosynthesis on USPTO-50K, NatureLM-8x7B reaches 71.9% Top-1 and 87.4% Top-3, surpassing LocalRetro, R-SMILES, and EditRetro on the reported Top-1 metric.

Protein design is treated as a parallel generative regime rather than a separate specialist problem. In unconditional protein generation, NatureLM-8x7B reports average length 318, diversity 0.989, and average pLDDT 75.9, whereas GPT-4 reports 46, 0.816, and 65.1, and Mixtral 8x7B reports 53, 0.906, and 69.9 (Xia et al., 11 Feb 2025). For text-guided protein generation, the model is evaluated on stability and solubility; the 8x7B version reaches average prediction 0.655 and data ratio 0.812 for stability score $> 0.5$. On antigen-binding CDR-H3 design evaluated on RAbD, NatureLM-8x7B reaches AAR 0.376, competitive with specialized antibody-design models though not always best. For protein-to-text description generation, NatureLM-8x7B reports Rouge-L 0.585, above GPT-4o at 0.091 and fine-tuned Llama 3 (8B) at 0.324.

The same sequence interface extends to RNA and materials. For unconditional RNA generation, NatureLM-8x7B reports MFE $-177.1$ kcal/mol and retrieves 165 Rfam families; for guide RNA design it reports validity 0.957 and Top-1 accuracy 0.699 (Xia et al., 11 Feb 2025). In protein-to-RNA design, success rates rise with scale, from 40.9% for 1B to 44.8% for 8x7B. In materials, unconditional generation is evaluated with SMACT validity and stability via energy above hull, while composition-conditioned and property-conditioned generation are evaluated with validity, stability, precision, and novelty. The 8x7B model reports 94.75% SMACT and 53.60% stability in property-conditioned generation, and the paper highlights two discovered materials with DFT bulk moduli of 390 GPa and 394 GPa, close to the target 400 GPa. A companion fine-tuned model, NatureLM-Mat3D (1B), is reported to be especially strong on MPTS-52 for crystal-structure prediction, with MR/RMSE 30.20/0.0837 (Xia et al., 11 Feb 2025).

These results are used to support a broader claim: NatureLM is not merely multi-task but multi-domain and cross-domain. The paper explicitly lists cross-domain generation regimes such as protein-to-molecule, protein-to-RNA, DNA-to-RNA, and text-plus-materials-property generation (Xia et al., 11 Feb 2025).

4. NatureLM-audio and bioacoustics

NatureLM-audio adapts the NatureLM program to ecological acoustics. Its architecture consists of a BEATs audio encoder, a window-level Q-Former following a SALMONN-style design, and a Llama 3.1-8B language backbone adapted with LoRA. The audio encoder and Q-Former are trainable, while the LLM backbone is frozen except for the LoRA adapter layers (Robinson et al., 2024). Training is curriculum-like. A first stage performs perception pretraining on focal species classification, and a second stage performs generalization fine-tuning with additional bioacoustic tasks including detection, captioning, lifestage prediction, and call-type prediction, while also adding speech and music data.

The training corpus is curated across bioacoustics, general audio, music, and speech. Bioacoustic sources include Xeno-canto, iNaturalist, Animal Sound Archive, and Watkins for species-level tasks, along with soundscape data from Sapsucker Woods, Sierra Nevada, and the University of Hawai‘i at Hilo (Robinson et al., 2024). Captioning uses AnimalSpeak and reprocessed Xeno-canto captions generated with Gemini-1.0-pro. Auxiliary music tasks are derived from NSynth, including pitch prediction in Hz, instrument classification, velocity, timbre and quality description, and mixtures of 1–3 instruments. Speech contributes a synthetic speaker-diarization-style counting task derived from LibriSpeech.

Evaluation centers on BEANS-Zero, which extends BEANS with zero-shot and unseen-species testing. Standard tasks include esc50, watkins, cbi, humbugdb, dcase, enabirds, hiceas, rfcx, and gibbons; new tasks include unseen-cmn, unseen-sci, lifestage, call-type, captioning, and zf-indv (Robinson et al., 2024). The reported metrics are accuracy for classification, macro-averaged F1 for detection, and SPIDEr for captioning.

NatureLM-audio is reported to set state of the art on 6 of 9 standard zero-shot classification and detection datasets. The table values include esc50 0.635, watkins 0.646, cbi 0.755, humbugdb 0.073, dcase 0.052, enabirds 0.279, hiceas 0.390, rfcx 0.039, and gibbons 0.003 (Robinson et al., 2024). On held-out species, NatureLM-audio reaches 0.116 on unseen-cmn and 0.196 on unseen-sci, above a CLAP baseline at 0.034 and 0.004. On novel tasks it reports lifestage 0.763, call-type 0.810, captioning 0.494 SPIDEr, and zf-indv 0.383, all stated as new state of the art. A targeted ablation on zebra-finch counting reports 0.377 accuracy with speech in stage-2 training and 0.243 without speech, supporting the paper’s transfer claim from speech speaker counting to vocalizing-individual counting.

The paper also records limitations and risks: bird bias in the training distribution, misuse risk for endangered-species monitoring, and the possibility of behavioral or ecological impacts from large-scale passive monitoring (Robinson et al., 2024). These caveats are integral to the model’s scientific positioning.

5. Derived systems and post-training adaptations

NatureLM has also been reused as a backbone inside larger multimodal pipelines. In MolChord, a reproduction of NatureLM-1B serves as the autoregressive sequence generator for protein-guided drug design, unifying text, protein FASTA, and molecular SMILES while a diffusion-based structure encoder based on FlexRibbon supplies contextual 3D features through a lightweight gated MLP adapter (Zhang et al., 31 Oct 2025). MolChord trains in three stages: Stage A aligns the structure encoder and NatureLM using 676K protein examples, 316K small-molecule examples, and 94K protein–ligand complexes; Stage B performs supervised fine-tuning on CrossDocked2020 with pocket structural features injected into NatureLM and a VAE noise mechanism for diversity; Stage C performs DPO refinement on a preference dataset constructed from generated candidates. On CrossDocked2020, MolChord reports Vina Dock $-7.62$, High Affinity 55.1%, QED 0.56, SA 0.77, Diversity 0.76, and Success Rate 33.2%, while MolChord-RL reports $-8.59$, 74.6%, 0.56, 0.78, 0.71, and 53.4%, respectively.

A separate line of work studies the effects of fine-tuning on NatureLM-audio and introduces model merging as a corrective mechanism. NatureLM-audio is described there as a LoRA fine-tuning of Llama-3.1-8B-Instruct on about 2M audio-text pairs, mostly bioacoustics but also music and human sounds (Marincione et al., 7 Nov 2025). The central observation is that domain-specific fine-tuning yields strong performance on bioacoustic benchmarks but reduces instruction-following flexibility. The reported merge is linear weight interpolation,

$$\mathbf{\Theta}^{(\text{merge})} = (1 - \alpha)\mathbf{\Theta}^{(1)} + \alpha \mathbf{\Theta}^{(2)},$$

with the endpoints taken as the base Llama-3.1-8B-Instruct and NatureLM-audio (Marincione et al., 7 Nov 2025). For a combined prompt requesting both scientific and common species names, intermediate interpolation substantially improves behavior: on Watkins, accuracy rises from 6% to 45%, and on CBI, from 12% to 63%, when moving from $\alpha = 1$ toward approximately $\alpha \approx 0.7$. On the unseen-family-cmn split of BEANS-Zero, the merged model reports F1 $= 0.28$, compared with F1 $\in[0,5]$0 for NatureLM-audio, described as over a 200% relative improvement and a new state of the art for that closed-set zero-shot task.

Together, these downstream studies show two distinct properties of the NatureLM paradigm. First, the sequence interface is sufficiently generic to be embedded into structure-aware systems such as MolChord. Second, post-training specialization can create measurable trade-offs in prompt adherence and zero-shot behavior, making lightweight post hoc interventions such as model merging scientifically consequential.

6. Evaluation, trade-offs, and open questions

A broad external assessment appears in BioMol-LLM-Bench, a cross-scale benchmark with 26 downstream tasks organized into four difficulty levels: L0 general bio-molecular knowledge understanding, L1 small-molecule tasks, L2 protein-level tasks, and L3 multiple-molecules interaction tasks (Xu et al., 3 Apr 2026). The benchmark evaluates 13 representative models, including NatureLM-8x7B, and incorporates both direct prompting and tool-augmented workflows through named APIs such as ADMETAI_predict_solubility_lipophilicity_hydration, ADMETAI_predict_BBB_penetrance, and ADMETAI_predict_stress_response.

The reported findings are cautious. Chain-of-thought data provides limited benefit and may even reduce performance on biological tasks. Hybrid mamba-attention architectures are more effective for long bio-molecular sequences. Supervised fine-tuning improves specialization at the cost of generalization. Classification tasks are substantially easier than regression tasks, for which none of the evaluated models achieves meaningful performance on the hardest cases (Xu et al., 3 Apr 2026). The paper also states that NatureLM-8x7B performs poorly in many settings and that deep fine-tuning can produce zeros on some tasks when the fine-tuning does not transfer.

Tool use partly changes that picture. In the benchmark’s tool-augmented evaluation, tools improve performance especially on regression tasks, reduce extreme errors, and consistently outperform direct prompting (Xu et al., 3 Apr 2026). This suggests, though does not prove, that the most robust realization of the NatureLM agenda may be hybrid: a language-centered generalist model for representation and orchestration, coupled to external scientific tools for numerically exact computation.

The resulting interpretation is neither that NatureLM collapses into a standard domain-specific generator nor that it already furnishes mechanistically faithful scientific reasoning. Rather, the literature portrays NatureLM as a unification strategy with substantial empirical reach—across molecules, proteins, RNA, materials, and bioacoustics—but also with clearly documented tensions between specialization and generalization, prompt following and domain adaptation, and fluent generation and quantitative scientific accuracy (Xia et al., 11 Feb 2025).