Vision-language models (VLMs) achieve impressive zero-shot performance on multimodal reasoning tasks. Typically, best reported performance is achieved with a zero- or a few-shot prompt. We observe that asking the model to take other routes of solving the same task, such as through code generation, hurts performance. Furthermore, training sets are typically no longer useful for improving model performance through few-shot learning, due to their use in training. Indeed, we observe that auto-prompting techniques such as DSPy (CITATION), when applied on training sets, do not produce few-shot examples that further improve validation performance. Further, when used in conjunction with program-of-thought, performance becomes even worse.Our work overcomes these limitations by introducing a novel self-play programming interface which leverages the ability of VLMs to first generate code to decompose a complex visual reasoning task in sub-tasks, then use itself, or other models, as a tool to solve decomposed tasks. Our approach enables DSPy to not suffer from performance drops, when applied iteratively on training sets. Furthermore, it outperforms zero-shot baselines on difficult chart reasoning benchmarks. We report the performance of our approach on ChartQA, PlotQA and ChartFC. This enables large models, such as Gemini or GPT to autonomously learn how to use themselves as tools and iteratively improve without the need for additional data.

We introduce ScreenQA, a novel benchmarking dataset designed to advance screen content understanding through question answering. The existing screen datasets are focused either on low-level structural and component understanding, or on a much higher-level composite task such as navigation and task completion for autonomous agents. ScreenQA attempts to bridge this gap. By annotating 86k question-answer pairs over the RICO dataset, we aim to benchmark the screen reading comprehension capacity, thereby laying the foundation for vision-based automation over screenshots. Our annotations encompass full answers, short answer phrases, and corresponding UI contents with bounding boxes, enabling four subtasks to address various application scenarios. We evaluate the dataset’s efficacy using both open-weight and proprietary models in zero-shot, fine-tuned, and transfer learning settings. We further demonstrate positive transfer to web applications, highlighting its potential beyond mobile applications.

Vision-language models (VLMs) are achieving increasingly strong performance on multimodal tasks. However, reasoning capabilities remain limited particularly for smaller VLMs, while those of large-language models (LLMs) have seen numerous improvements. We pro-pose a technique to transfer capabilities from LLMs to VLMs. On the recently introduced ChartQA, our method obtains state-of-the-artperformance when applied on the PaLI3-5B VLM by Chen et al. (2023c), while also enabling much better performance on PlotQA and FigureQA.We first improve the chart representation by continuing the pre-training stage using an improved version of the chart-to-table translation task by Liu et al. (2023a). We then propose constructing a 20x larger dataset than the original training set. To improve general reasoning capabilities and improve numerical operations, we synthesize reasoning traces using the table representation of charts. Lastly, our model is fine-tuned using the multitask loss introduced by Hsieh et al. (2023).Our variant ChartPaLI-5B outperforms even 10x larger models such as PaLIX-55B without using an upstream OCR system, while keeping inference time constant compared to the PaLI3-5B baseline. When rationales are further refined with a simple program-of-thought prompt (Chen et al., 2023a), our model outperforms the recently introduced Gemini Ultra and GPT-4V.

While self-correction has shown promise in improving LLM outputs in terms of style and quality (e.g. Chen et al., 2023b; Madaan et al.,2023), recent attempts to self-correct logical or reasoning errors often cause correct answers to become incorrect, resulting in worse performances overall (Huang et al., 2023). In this paper, we show that poor self-correction performance stems from LLMs’ inability tofind logical mistakes, rather than their ability to correct a known mistake. Firstly, we benchmark several state-of-the-art LLMs ontheir mistake-finding ability and demonstrate that they generally struggle with the task, even in highly objective, unambiguous cases. Secondly, we test the correction abilities of LLMs – separately from mistake finding – using a backtracking setup that feeds ground truth mistake location information to the model. We show that this boosts downstream task performance across our 5 reasoning tasks, indicating that LLMs’ correction abilities are robust. Finally, we show that it is possible to obtain mistake location information without ground truth labels or in-domain training data. We train a small classifier with out-of-domain data, which exhibits stronger mistake-finding performance than prompting a large model. We release our dataset of LLM-generated logical mistakes, BIG-Bench Mistake, to enable further research into locating LLM reasoning mistakes.

Cascades are a common type of machine learning systems in which a large, remote model can be queried if a local model is not able to accurately label a user’s data by itself. Serving stacks for large language models (LLMs) increasingly use cascades due to their ability to preserve task performance while dramatically reducing inference costs. However, applying cascade systems in situations where the local model has access to sensitive data constitutes a significant privacy risk for users since such data could be forwarded to the remote model. In this work, we show the feasibility of applying cascade systems in such setups by equipping the local model with privacy-preserving techniques that reduce the risk of leaking private information when querying the remote model. To quantify information leakage in such setups, we introduce two privacy measures. We then propose a system that leverages the recently introduced social learning paradigm in which LLMs collaboratively learn from each other by exchanging natural language. Using this paradigm, we demonstrate on several datasets that our methods minimize the privacy loss while at the same time improving task performance compared to a non-cascade baseline.

Despite recent advances, evaluating how well large language models (LLMs) follow user instructions remains an open problem. While evaluation methods of language models have seen a rise in prompt-based approaches, limited work on the correctness of these methods has been conducted. In this work, we perform a meta-evaluation of a variety of metrics to quantify how accurately they measure the instruction-following abilities of LLMs. Our investigation is performed on grounded query-based summarization by collecting a new short-form, real-world dataset riSum, containing 300 document-instruction pairs with 3 answers each. All 900 answers are rated by 3 human annotators. Using riSum, we analyze the agreement between evaluation methods and human judgment. Finally, we propose new LLM-based reference-free evaluation methods that improve upon established baselines and perform on par with costly reference-based metrics that require high-quality summaries.

The task of summarisation is notoriously difficult to evaluate, with agreement even between expert raters unlikely to be perfect. One technique for summary evaluation relies on collecting comparison data by presenting annotators with generated summaries and tasking them with selecting the best one. This paradigm is currently being exploited in reinforcement learning using human feedback, whereby a reward function is trained using pairwise choice data. Comparisons are an easier way to elicit human feedback for summarisation, however, such decisions can be bottle necked by the usability of the annotator interface. In this paper, we present the results of a pilot study exploring how the user interface impacts annotator agreement when judging summary quality.