Improving Automatic VQA Evaluation Using Large Language Models

It is crucial to choose an appropriate LLM as LAVE’s performance directly hinges on its capabilities. We pose VQA evaluation as a close-ended answer-verification task and adapt a frozen LLM through in-context learning. Hence, we opt for an instruction-tuned LLM, which has demonstrated superior performance in transferring to new tasks with limited demonstrations (Wei et al. 2022a). Instruction-tuned LLMs are also more robust to prompt selection, and they can match the few-shot performance of much larger LLMs pretrained with self-supervised objectives. Considering all these factors, we first select the Flan-T5 (Chung et al. 2022) model family for our metric. In addition, Flan-T5 is finetuned on chain-of-thought (CoT) data, enabling it to provide reasoning for its answers.

To demonstrate LAVE’s robustness across different LLMs, we also consider Vicuna-v1.3 (Chiang et al. 2023) and GPT-3.5-Turbo (aka ChatGPT (OpenAI 2022)). We optimize our prompt for Flan-T5 (Sec. 5.4) and subsequently use the same prompt with the other LLMs. This opens the possibility of enhancing our metric without extra effort as stronger LLMs become available in the future.

We frame VQA evaluation as an answer-rating task amenable to in-context learning with LLMs. We adopt a rating scale ranging from 1 to 3 (as opposed to a binary rating) to account for ambiguous questions or incomplete answers. Our prompt (Fig. 3) comprises the typical components: task description, a few demonstrations of input/output, and the input for a test example. We draw inspiration from SNI (Wang et al. 2022) to structure our task description, as it is one of the main sources of training data for instruction-tuned LLMs. We also append “Give the rationale before rating.” to elicit a justification for the assigned rating, which improves explainability. Each demonstration consists of a question $q$, a set of reference answers $R$, the candidate answer $c$, the answer rating $r$, and an explanation $e$ for the rating. We observed binary questions are particularly challenging to evaluate (App. A.3), so we manually curate two sets of 8 demonstrations, one for binary questions and the other for general questions. We ensure demonstrations are diverse and cover various question types, numbers of reference answers, levels of agreement, candidate answer precision and verbosity, ratings, etc (refer to App. A.2 for the complete list). While these sets of demonstrations are designed to be comprehensive, users of our metric could also provide their own demonstrations to cover different cases. Additionally, to account for noise in the annotations, we filter out outlier reference answers that have a frequency lower than 25% of the maximum answer frequency. Finally, the test example only includes $q$, $R$ and $c$, and the LLM is expected to generate an explanation $e$ followed by a rating $r$. We found incorporating visual context from the image $i$ into the prompt does not provide significant benefits (see Sec. 5.4 and App. A.5 for more details).

Given the LLM’s generated text for the test example, we extract the rating $r$ from the last character (either 1, 2 or 3) and linearly map it to a score $s$ in the range $[0,1]$: $s=(r-1)/2$. Inspired by Liu et al. (2023), we also explored the possibility of using the probabilities of output tokens to normalize the ratings and take their weighted sum as the final rating, but we did not observe any improvements in our task.

We collected human judgments about answer quality using Amazon Mechanical Turk (MTurk) with the same web interface as Agrawal et al. (2023). Specifically, we collected judgments for answers generated by BLIP-2 on VQAv2, VG-QA and OK-VQA, PromptCap on VQAv2 and OK-VQA, BLIP_VQA on VG-QA and OK-VQA, and BLIP_VG on VQAv2 and OK-VQA ($2450$ questions each²²2We initially collected human judgments on $2500$ questions per model-dataset pair, but had to remove some to control data quality.). In total, our test set contains $22.1$k questions. We additionally collected validation/development sets of human judgments for answers generated by BLIP-2 on VQAv2 and VG-QA, and BLIP on VQAv2 and VG-QA ($1000$ questions each). In total, our validation set contains $4$k questions, which serve to guide our design choices. We emphasize that PromptCap and OK-VQA are completely unseen during metric development to show LAVE’s generality. Following (Agrawal et al. 2023), each answer was assessed by 5 annotators who were asked to provide a binary rating (correct/incorrect) based on the corresponding image and question. An important difference between the task posed to turkers and to the LLM is that the LLM is provided with a list of reference answers annotated by humans, while turkers are not. After filtering out low-quality annotations, we obtain an inter-annotator agreement measured by Krippendorff’s $\alpha$ of $62.0$. Upon manual inspection, we observed that model responses (and even questions) are often ambiguous in nature, which would explain the relatively low inter-annotator agreement (see App. A.1 for more details). We derive a single “quality” score from the $5$ binary ratings per answer as follows: $1.0$ if at least $4$ annotators rate the answer as correct, $0.5$ if only $2$ or $3$ did so, and $0.0$ otherwise. Considering partial scores is crucial to acknowledge the ambiguity inherent in certain questions, particularly when dealing with generative models which can produce technically accurate answers that may not align with the intended meaning of the question.

To evaluate LAVE, we measure its correlation with human judgment using two widely accepted rank correlation coefficients: Spearman’s $\rho$ and Kendall’s $\tau$ (in the appendix). These provide a robust measure of the association between metrics, without assuming a linear relationship or a specific distribution of the data. Both coefficients range from $-1$ to $1$, with $-1$/$+1$ meaning perfect inverse/direct correlation.

Spearman correlations between VQA metrics (ours and baselines) and human judgment on every evaluated pair of VQA model and dataset are shown in Tab. 2 (see App. A.3 for additional results and observations). We verify statistical significance by bootstrapping with 5000 resamples and running a t-test with a significance level of 5% between pairs of correlations. For each setting, we bold the best results which are significantly better than the second-best result. The main observations are summarized as follows:

Overall, LAVE is significantly more aligned with human judgment than all the considered baselines, independently of the underlying LLM. Among the considered LLMs, we observe GPT-3.5 provides the highest overall correlation with human judgment. This is in line with recent LLM leaderboards (Zheng et al. 2023) which, at the time of writing, place GPT-3.5 (along GPT-4 (OpenAI 2023) and Claude (Anthropic 2023)) one step above any other LLM across a diverse set of benchmarks. However, we are excited to report that, on the VQA evaluation task, open-source LLMs such as Flan-T5 or Vicuna also outperform all baselines on average. Interestingly, despite being trained on user-shared conversations with ChatGPT, Vicuna falls behind Flan-T5 in our task.

LAVE generalizes to new VQA models and benchmarks. We did not use human judgments of answers produced by PromptCap (across all datasets) or to OK-VQA questions (from all models) to guide our prompt design (see Sec. 5.4). Still, our metric correlates better with human judgment than all baselines in these hold-out settings. This indicates our design choices are not overfitted to the particular settings used during metric development, and that LAVE appears promising for evaluating answers generated by various models and across different datasets.

Questions from VQAv2 and OK-VQA answered by BLIP follow a different trend. In these settings, VQA Accuracy’s correlation with human judgment is considerably higher than for zero-shot VQA models, whereas LAVE has only a slightly higher correlation. We observe human score is much higher for BLIP-2 and PromptCap answers ($0.7552$ on average) compared to BLIP answers ($0.5293$ on average). Therefore, BLIP answers are more frequently incorrect or incomplete, which is expected as open-ended generative models are known to perform better on OOD data. The higher correlation for VQA Accuracy in these settings can be attributed to its efficacy in identifying incorrect candidate answers, while LLMs might label some as correct. For instance, GPT-3.5 labels “sink“ as a correct answer to the question “What is this kind of sink called?”, or “refrigerator” as a correct answer to “What does this device generally do?”. Thus, the different trend in correlation with human judgment can be explained by a higher frequency of incorrect answers. This trend does not hold for VG-QA because LAVE outperforms the baselines when there is a single reference answer (see App. A.3).

We compute correlation between LAVE _FT5 and human judgment when ablating for different prompt design choices. The best overall configuration is used to compute correlation on the test sets. Tab. 3 summarizes our ablation results.

Aside from having better overall correlation with human judgment, we would like to know how LAVE behaves in the failure modes of VQA Accuracy highlighted in Sec. 3. As a reminder, these are all cases where human annotators collectively labeled the candidate answer as correct (score of $1.0$), while VQA Accuracy was below $0.5$ (either $0.3$ or $0.0$). Therefore, we would expect our metric to give these candidate answers a score of $1.0$ (excluding incorrect cases – $8.25\%$)³³3Note that, in this setting, it is not possible to compute correlation with human judgment since it is constant ($1.0$).. Out of the $22.1$k questions from our test sets, this is the case for $3601$ questions ($16.33\%$). For completeness, we found the reverse scenario, collective human score of $0.0$ and VQA Accuracy above $0.5$, occurs in 379 questions ($1.72\%$); these are cases where the new human annotators disagree with the original annotations of the VQA datasets, indicating some noise in our collected human judgments.

Fig. 4 shows the average score of VQA Accuracy, Soft VQA Accuracy, LAVE _FT5 and LAVE _GPT-3.5 on the 400 manually-labeled examples analyzed in Sec. 3. We observe that LAVE is significantly more aligned with human judgment than both VQA Accuracy and Soft VQA Accurarcy, especially when candidate answers are more verbose or they are a synonym of the reference answers. As previously mentioned, broad questions or which have multiple correct answers may be overly subjective, so it is harder for an LLM to determine whether the candidate answer is correct. It is interesting to see, however, that in these cases Flan-T5 generally performs better than GPT-3.5. In summary, this indicates that LAVE is able to recover a considerable fraction of correct candidate answers wrongly labeled as incorrect by VQA Accuracy.

Tab. 4 contains a few selected examples where LAVE _GPT-3.5 fixes failures of VQA Accuracy. For instance, example (a) shows our metric is able to identify that the candidate answer is a synonym of several references, even though the form is different. Example (b) demonstrates our metric is robust to answers of diverse verbosity. In example (c), our metric is capable of identifying the candidate answer is equivalent to the references, even though they belong to different lexical categories.