Recently, the visual programming framework (VisProg) has emerged as a significant framework for executing compositional visual tasks due to its interpretability and flexibility. However, the performance of VisProg on specific Visual Reasoning (VR) tasks is markedly inferior compared to well-trained task-specific models since its employed visual sub-modules have limited generalization capabilities. Due to the non-differentiability of VisProg, it is quite challenging to improve these visual sub-modules within VisProg for the specific VR task while maintaining their generalizability on the un-seen tasks. Attempt to overcome these difficulties, we propose CLVP, a Continuous Learning paradigm for VisProg across various visual reasoning tasks. Specifically, our CLVP distills the capabilities of well-trained task-specific models into the visual sub-modules in a stepwise and anti-forgetting manner. This can continually improve the performance of VisProg on multiple visual tasks while preserving the flexibility of VisProg. Extensive and comprehensive experimental results demonstrate that our CLVP obtains significant performance gains on specific VR benchmarks, i.e., GQA (+1.4%) and NLVRv2 (+5.6%), compared to the VisProg baseline, and also maintains a promising generalizability for VR on un-seen and previous learned tasks.

Today, large language models (LLMs) are taught to use new tools by providing a few demonstrations of the tool's usage. Unfortunately, demonstrations are hard to acquire, and can result in undesirable biased usage if the wrong demonstration is chosen. Even in the rare scenario that demonstrations are readily available, there is no principled selection protocol to determine how many and which ones to provide. As tasks grow more complex, the selection search grows combinatorially and invariably becomes intractable. Our work provides an alternative to demonstrations: tool documentation. We advocate the use of tool documentation, descriptions for the individual tool usage, over demonstrations. We substantiate our claim through three main empirical findings on 6 tasks across both vision and language modalities. First, on existing benchmarks, zero-shot prompts with only tool documentation are sufficient for eliciting proper tool usage, achieving performance on par with few-shot prompts. Second, on a newly collected realistic tool-use dataset with hundreds of available tool APIs, we show that tool documentation is significantly more valuable than demonstrations, with zero-shot documentation significantly outperforming few-shot without documentation. Third, we highlight the benefits of tool documentations by tackling image generation and video tracking using just-released unseen state-of-the-art models as tools. Finally, we highlight the possibility of using tool documentation to automatically enable new applications: by using nothing more than the documentation of GroundingDino, Stable Diffusion, XMem, and SAM, LLMs can re-invent the functionalities of the just-released Grounded-SAM and Track Anything models.