A Study of Adaptive Modeling Towards Robust Generalization

Zihao Jing, Qiuhao Zeng, Ruiyi Fang, Yan Yi Li, Yan Sun, Boyu Wang, Pingzhao HuarXiv 2026 · DOI · arXiv

TL;DR

Cuttlefish is a new multimodal framework that allows Large Language Models (LLMs) to reason across diverse biomolecular structures—including proteins, DNA, RNA, and small molecules—using a unified "all-atom" approach. By implementing Scaling-Aware Patching, the model adaptively allocates more tokens to complex structures while using a Geometry Grounding Adapter to inject explicit 3D spatial cues, significantly reducing structural hallucinations compared to fixed-length or sequence-only models.

What Problem Does This Paper Address?

Current AI models for biology often suffer from two major bottlenecks:

  1. Modality Silos: Most models are specialized for just one type of molecule (e.g., only proteins or only small molecules).
  2. Rigid Tokenization: Existing "structure-aware" LLMs typically use sequence-style encodings (like SMILES strings) or a fixed number of "connector tokens" (like a Q-Former).
  3. Sequence-only inputs lack 3D geometric context, leading the model to "hallucinate" chemical properties that aren't physically possible.
  4. Fixed-length connectors create a "compression bottleneck." If a model uses 64 tokens for every input, it over-allocates resources to tiny molecules (diluting attention) and over-compresses massive proteins (losing critical detail).

The authors aim to create a unified, geometry-grounded interface that scales its computational budget based on the actual complexity of the biological entity.

Methods

The authors introduce Cuttlefish, which uses a three-part architecture to bridge microscopic geometry with linguistic reasoning:

  • SE(3)-Equivariant Encoder: An Equivariant Graph Neural Network (EGNN) processes all-atom inputs (coordinates and features) to ensure the model's understanding remains stable regardless of how a molecule is rotated or translated in 3D space.
  • Scaling-Aware Patching: This is the core innovation. Instead of a fixed token count, Cuttlefish uses an instruction-conditioned gating policy. It identifies "anchor atoms" based on the user's prompt and grows variable-sized "soft patches" around them. This allows the token budget to grow naturally with the size and complexity of the structure.
  • Geometry Grounding Adapter: This module refines the patch tokens through cross-attention with the full atom-level embeddings. It then injects these "geometry-informed" tokens directly into the LLM’s embedding space, ensuring the language output is physically grounded in the 3D structure.

The training involves a two-stage process: first, aligning the structural tokens with the text (modality alignment), and then fine-tuning the LLM backbone to accommodate the new structural data.

Datasets

The researchers evaluated Cuttlefish on several established benchmarks:

  • Mol-Instructions: A large-scale biomolecular instruction dataset.
  • DNA-Chat & RNA-QA: Specialized datasets for nucleic acid reasoning.

Crucially, they introduced a new dataset called GEO-AT. This is the first unified all-atom instruction dataset that includes 3D coordinates and language annotations for molecules, proteins, and nucleic acids. It serves as a "community resource" to catalyze the development of geometry-grounded LLMs.

Key Insights and Ideas

  • Instruction-Conditioned Selection: Unlike previous models that compress a molecule the same way every time, Cuttlefish changes which parts of the structure it "focuses" on based on the question asked.
  • The "Aha!" of Mixed-Modality Training: The authors found that training on molecules, proteins, and DNA/RNA simultaneously actually improved performance. The model benefits from "cross-modality diversity," learning general chemical relations that transfer across different types of biological entities.
  • Ending the Compression Bottleneck: By allowing the number of tokens to scale with the number of atoms (as seen in Fig. 6), Cuttlefish maintains stable performance on large proteins where fixed-capacity models (like Mol-Llama) typically see a sharp drop in accuracy.

Key Results

  • Reduced Hallucinations: In a functional group hallucination test, Cuttlefish achieved the lowest hallucination rate (0.07 for molecules, 0.04 for proteins) compared to strong baselines like Mol-Llama and ProtChatGPT.
  • Superior Reasoning: Across 17 tasks in Mol-Instructions and 20 tasks in DNA-Chat, Cuttlefish consistently outperformed modality-specific models. For example, it achieved a METEOR score of 0.715 on molecular captioning, surpassing the previous state-of-the-art.
  • Zero-Shot Entity Recognition: Even when given only structural coordinates without any text hints, Cuttlefish could accurately identify whether it was looking at a small molecule, a protein, or a nucleic acid.

Limitations and Future Work

  • Extremely Large Entities: While Cuttlefish scales better than previous models, it still struggles with "extremely large" entities (e.g., >50,000 nodes). The authors suggest a "Retrieval-Augmented Hierarchical Reasoning" approach for future work to handle entire cellular complexes.
  • Upstream Noise: The model's performance is highly sensitive to the quality of the input 3D coordinates. If the provided structure is poorly resolved, the model's reasoning suffers.
  • Coarse Supervision: Currently, the model relies on the general reasoning capabilities of the underlying LLM (like Llama-3). The authors hope to build more specialized "chain-of-thought" datasets for all-atom reasoning to improve deep scientific deliberation.

Summary updated

Abstract

Large language models (LLMs) increasingly support reasoning over biomolecular structures, but most existing approaches remain modality-specific and rely on either sequence-style encodings or fixed-length connector tokens for structural inputs. These designs can under-expose explicit geometric cues and impose rigid fusion bottlenecks, leading to over-compression and poor token allocation as structural complexity grows. We present a unified all-atom framework that grounds language reasoning in geometric information while adaptively scaling structural tokens. The method first constructs variable-size structural patches on molecular graphs using an instruction-conditioned gating policy, enabling complexity-aware allocation of query tokens. It then refines the resulting patch tokens via cross-attention with modality embeddings and injects geometry-informed tokens into the language model to improve structure grounding and reduce structural hallucinations. Across diverse all-atom benchmarks, the proposed approach yields consistent gains in heterogeneous structure-grounded reasoning. An anonymized implementation is provided in the supplementary material.
BibTeX
@misc{jing2026study,
  author = {Zihao Jing and Qiuhao Zeng and Ruiyi Fang and Yan Yi Li and Yan Sun and Boyu Wang and Pingzhao Hu},
  title = {{A Study of Adaptive Modeling Towards Robust Generalization}},
  year = {2026},
  eprint = {2602.02780},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.02780},
}
Figure