arXiv News

First-hand information for everyone

Site

AboutPrivacy PolicyTerms of Use

© 2026 arXiv News

A single‑task trained LLM discovers universal machine‑learning configurations that outperform experts on four vision, audio and text‑to‑image tasks | arXiv News

arXiv News

EnglishJapanese

Switch language

EnglishJapanese
Loading account…

First-hand information for everyone

Back to front page
Natural Language ProcessingEnglishPublished

A single‑task trained LLM discovers universal machine‑learning configurations that outperform experts on four vision, audio and text‑to‑image tasks

November 11, 2026arXiv: 2611.12791v1

Researchers have introduced a new encoding scheme, NCode, that turns mixed design choices – model structures, training parameters and data‑mix ratios – into a single sequence of numbers. An LLM‑based searcher learns from optimization paths collected on a single vision benchmark and then ranks candidate NCodes, avoiding the need to generate code directly.

Figure showing how the NCode scheme encodes diverse AutoML design choices into a single numerical sequence, supporting the claim that NCode unifies heterogeneous components.
Fig. 1. Overview of the proposed LLM‑based open‑world AutoML framework and the NCode representation. (a) Conventional AutoML methods adapt models separately for each task, whereas our approach uses a universal encoding to jointly optimize architectures, training hyper‑parameters, and data ratios across tasks, domains, and modalities. (b) NCode maps heterogeneous design choices – including architectures, operations, hyper‑parameters, and data sampling weights – into unified discrete numerical sequences. (c) Illustrative encoding of models with varying topologies, operations, training hyper‑parameters, and data ratios.Hu, Yuxuan, Liu, Jihao, Shi, Weikang, Huang, Xijie, Wei, Linda, Zhan, Mingjie et al. · ·
arXiv:2611.12791v1 · Fig. 1:
CC BY 4.0

Because the LLM treats every configuration as a discrete code, it can jointly search architectures, hyper‑parameters and data ratios in a zero‑shot setting. The method was trained only on NAS‑Bench‑301 trajectories and has demonstrated state‑of‑the‑art performance on image segmentation, as reported in Table I.

Reinforcement‑learning based Chain‑of‑Thought prompting helped explain the LLM’s shift from early‑stage exploration to late‑stage exploitation.

However, the initial training data came from a single domain, which may limit the breadth of prior knowledge available for unrelated tasks. Further work is needed to confirm generality beyond the reported modalities.

The figure and table attached visually illustrate how NCode maps heterogeneous design choices into a unified sequence and show the joint‑optimization advantage over baseline methods.

Source paper

Large Language Models as a Universal Searcher for Open-World Automated Machine Learning

Natural Language Processing
View on arXiv