Paper Shows Small Open Model Outperforming Gemini 3.1 Pro

The paper, titled “DualView: Leveraging Broad and Output-Specific Evidence for LLM-based Multi-Output Text Regression,” addresses a problem that has gotten less attention than single-number prediction: scoring a piece of text on more than one axis at once, such as rating an essay for grammar, argument strength and clarity in a single pass. The authors present DualView, which pulls two kinds of representation out of a frozen large language model, a broad reading of the whole text and an output-specific reading tuned to each score the system needs to produce, then combines them through a dual-branch architecture to generate the full vector of scores together.

The authors tested DualView across five heterogeneous regression tasks and eight large language models, including one with 122 billion parameters, and compared it against seven supervised regression baselines. They report that DualView beat all seven while fitting far fewer parameters than the three strongest of those baselines. The gains held up when the researchers looked at individual output dimensions separately rather than only at aggregate scores, and the method kept its advantage even when training data was scarce.

The detail most relevant to this site is the comparison with Gemini. The paper states that DualView, when built on a smaller open-weight language model, outperformed unsupervised prediction carried out with Gemini 3.1 Pro. That is a narrow claim: it covers unsupervised use of Gemini specifically, not every way Gemini could be deployed for the task, and the authors frame it as evidence that a lightweight, tuned approach on a smaller model can beat a larger general-purpose model that has had no task-specific adaptation applied to it.

The numbers

The paper’s headline architecture detail is its scale: the largest language model tested in the comparison carried 122 billion parameters, the size the authors print to establish how far up the model-size ladder their method was checked. Ablations in the paper further show that the broad and output-specific branches contribute separately rather than overlapping, and that lightweight shared predictors can reach strong results with comparatively few trainable parameters, reinforcing the efficiency claim at the center of the work.

For anyone currently routing text-scoring workloads through the Gemini API by default, the paper is a reminder that an unsupervised general-purpose call is not automatically the strongest or cheapest option once a dedicated, smaller method is fitted to the specific scoring task.