Instruction-Following Evaluation of Large Vision-Language Models

TL;DR

The paper shows that LVLMs lose instruction-following ability after visual instruction tuning, and that explicitly including output-format instructions in the fine-tuning data can mitigate this degradation. It introduces synthetic data pipelines (FOVIT/FOIT and related No variants) and a verbalizer-based evaluation to isolate the impact of output-format cues and visual information. Key findings reveal that output-format guidance substantially improves adherence (via $F_1$-based metrics) and that even a small amount of such data can restore instruction-following without harming general visual understanding. The work provides practical guidance for LVLM training, suggesting that formatting directives should be incorporated during visual instruction tuning to preserve alignment with user directives in real-world tasks.

Abstract

Following the initial flourishing of large language models (LLMs), there has been a surge in proposed large vision-language models (LVLMs) that integrate LLMs with vision capabilities. However, it has been observed that LVLMs, after tuning to visual instruction using commonly used training datasets, often fail to exhibit the instruction-following ability that was present in the LLM before integration, leading to results in which they do not follow task instructions as expected. This study quantitatively demonstrates that LVLMs' instruction-following ability declines after fine-tuning and analyzes its underlying causes. In particular, we constructed new training datasets highlighting whether the output format is specified. Then, we investigated how explicitly indicating the output format during fine-tuning affects LVLMs' instruction-following ability. Our quantitative evaluation confirmed that LVLMs' instruction-following ability declines after fine-tuning with commonly used datasets. Furthermore, we found that LVLMs trained with datasets, including instructions on output format, tend to follow instructions more accurately than models that do not. These findings suggest that including samples with instructions on output format during (visual) instruction tuning may help mitigate the decline in instruction-following abilities.

Figures (7)

• Figure 1: LVLMs (left) show lower instruction-following ability than LLMs (right). We examine this gap quantitatively and explore the factors that contribute to reductions in instruction-following ability.
• Figure 2: Example of a (visual) instruction training synthetic dataset consisting of COCO images, with captions generated by GPT-4V, instructions on the output format extracted from the IFEval dataset, and question texts and answers generated by GPT-4 (Section \ref{['subsec:create_instruction_datasets']}).
• Figure 3: Audit examples of synthetic QA annotations. Each panel displays an image with its instruction and the resulting question-answer (QA) pair, evaluated on two criteria: Image Grounding (Consistency) and Instruction Adherence. (a) All success case: the QA about a surfing scene is visually grounded and satisfies the lexical constraint (includes "environment" and "human"). (b) Image-text inconsistency: despite adhering to the lowercase constraint, the question incorrectly presupposes "two horses pulling a traditional carriage," leading to a grounding failure. (c) Instruction violation: the QA about a chef and a wood-fired oven is grounded but the answer ignores the formatting constraint (e.g., required repetitions of "story"), resulting in an adherence failure. Green check marks denote satisfied criteria; red crosses denote failures. In a manual audit of $100$ sampled FOVIT cases annotated by GPT-4 and GPT-4V, more than $80\%$ fell into category (a).
• Figure 4: A method of creating evaluation datasets through verbalizer manipulation using the SST-2 dataset. Three labels, "Natural," "Neutral," and "Unnatural," are defined, based on the consistency between the context and the label's semantic representation of the label.
• Figure 5: $\text{F}_1$ scores for the evaluation dataset are reported for $\text{LVLM}_{\text{FOVIT}}$, $\text{LVLM}_{\text{NoFOVIT}}$, $\text{LLM}_{\text{FOIT}}$, $\text{LLM}_{\text{NoFOIT}}$, and $\text{LVLM}_{\text{LLaVA}}$, which represent the models fine-tuned on the FOVIT, NoFOVIT, FOIT, NoFOIT, and LLaVA-Instruct-150K datasets, respectively. The figure is split into two panels: (a) results when the language backbone is $\text{Llama 2-Chat 7B}$; (b) results when the backbone is $\text{Llama 3.1 8B Instruct}$. Bars report scores for the "Natural," "Neutral," and "Unnatural" subsets (with their sample counts shown in the legend), and "All" refers to the macro average of the $\text{F}_1$ scores across these three conditions.