Scaling-Aware Adapter for Structure-Grounded LLM Reasoning

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

Read Scaling-Aware Adapter for Structure-Grounded LLM Reasoning by Zihao Jing, Qiuhao Zeng, Ruiyi Fang, Yan Yi Li, Yan Sun, Boyu Wang, Pingzhao Hu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-02780.txt.

Description

Zihao Jing and 6 co-authors ground a language model's reasoning in the geometry of molecular and atomic structures, in a 2026 paper. Models reasoning over two- and 3D structures hallucinate when links compress geometry into fixed token budgets. Existing methods are also specific to one modality, blocking unified all-atom reasoning.

Their system, Cuttlefish, first applies scaling-aware patching. An instruction-conditioned gate makes variable-size patches over structural graphs, scaling the token budget with structural complexity. A geometry grounding adapter then refines those tokens through cross-attention to modality vectors and feeds them into the language model. Explicit geometric cues reach the reasoning process head-on. Across interdisciplinary all-atom benchmarks, Cuttlefish beat rivals at heterogeneous structure-grounded reasoning. Token budgets that breathe with the input, rather than fixed funnels, are the central design lesson.

Catalog record

CREATORZihao Jing, Qiuhao Zeng, Ruiyi Fang, Yan Yi Li, Yan Sun, Boyu Wang, Pingzhao Hu
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE144 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-02780

Browse the full library · About the Tower of Babel