文章背景与核心概要

在大语言模型（LLM）的实际应用中，统一分配推理预算往往会带来巨大的计算开销，甚至导致“过度思考（over-thinking）”带来的性能惩罚，这在依赖复杂视觉布局驱动难度的文档中心化任务中尤为明显。为了解决这一挑战，本文作者推出了 BudgetDoc，这是首个旨在为三种不同文档任务中的模型-预算-性能权衡提供显式监督的多模态基准。

借助 BudgetDoc，作者训练了 DRB（Document-Reasoning Balancer）——一个参数量约为 10 亿的预检（pre-flight）估计器，它结合了 SigLIP-2 与 Qwen3-0.6B。DRB 能够准确预测不同预算级别下模型的序数性能，加权 F1 分数达到了 0.753。在跨 5 个前沿模型和 3 个数据集动态分配推理预算时，DRB 在 15 种配置中的 9 种达到了甚至超越了全最大预算基线的 F1 分数，同时大幅削减了计算成本。此外，初步评估还凸显了 DRB 在泛化至跨模型选择方面的巨大潜力。

Can a Lightweight Multimodal Model Estimate LLM Reasoning Performance? A Study for Compute-Optimal Document Inference

Authors: Zishan Ahmad, Vishal Vaddina
Submitted: August 19, 2026
Primary Subject: Artificial Intelligence (cs.AI)
arXiv ID: arXiv:2608.18591 [cs.AI]

Summary

Uniformly allocating reasoning budgets to Large Language Models (LLMs) is often computationally expensive and can lead to over-thinking penalties, particularly in document-centric tasks where complex visual layouts drive the difficulty.

To overcome this challenge, the authors introduce BudgetDoc, the first multimodal benchmark designed to provide explicit supervision for model-budget-performance trade-offs across three distinct document tasks. Using BudgetDoc, they train DRB (Document-Reasoning Balancer)—an approximately 1-billion-parameter pre-flight estimator combining SigLIP-2 and Qwen3-0.6B. DRB accurately predicts ordinal model performance across different budget levels, achieving a 0.753 weighted F1 score.

When dynamically allocating reasoning budgets across five frontier models and three datasets, DRB matched or improved F1 scores compared to always-maximum-budget baselines in 9 out of 15 configurations while drastically reducing computational costs. Furthermore, preliminary evaluations highlight DRB's promising potential to generalize toward cross-model selection.

Metadata

Links & Full-Text Access

• PDF: View PDF
• HTML: Experimental HTML Version
• Source: TeX Source
• External Citations: Google Scholar | Semantic Scholar | NASA ADS