Rethinking Interpretability in the Era of Large Language Models

Interpretable machine learning has exploded as an area of interest over the last decade, sparked by the rise of increasingly large datasets and deep neural networks. Simultaneously, large language models (LLMs) have demonstrated remarkable capabilities across a wide array of tasks, offering a chance to rethink opportunities in interpretable machine learning. Notably, the capability to explain in natural language allows LLMs to expand the scale and complexity of patterns that can be given to a human. However, these new capabilities raise new challenges, such as hallucinated explanations and immense computational costs. In this position paper, we start by reviewing existing methods to evaluate the emerging field of LLM interpretation (both interpreting LLMs and using LLMs for explanation). We contend that, despite their limitations, LLMs hold the opportunity to redefine interpretability with a more ambitious scope across many applications, including in auditing LLMs themselves. We highlight two emerging research priorities for LLM interpretation: using LLMs to directly analyze new datasets and to generate interactive explanations.

Teaching Language Models to Self-Improve through Interactive Demonstrations

The self-improving ability of large language models (LLMs), enabled by prompting them to analyze and revise their own outputs, has garnered significant interest in recent research. However, this ability has been shown to be absent and difficult to learn for smaller models, thus widening the performance gap between state-of-the-art LLMs and more cost-effective and faster ones. To reduce this gap, we introduce TriPosT, a training algorithm that endows smaller models with such self-improvement ability, and show that our approach can improve a LLaMA-7b's performance on math and reasoning tasks by up to 7.13%. In contrast to prior work, we achieve this by using the smaller model to interact with LLMs to collect feedback and improvements on its own generations. We then replay this experience to train the small model. Our experiments on four math and reasoning datasets show that the interactive experience of learning from and correcting its own mistakes is crucial for small models to improve their performance.

MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts

Although Large Language Models (LLMs) and Large Multimodal Models (LMMs) exhibit impressive skills in various domains, their ability for mathematical reasoning within visual contexts has not been formally examined. Equipping LLMs and LMMs with this capability is vital for general-purpose AI assistants and showcases promising potential in education, data analysis, and scientific discovery. To bridge this gap, we present MathVista, a benchmark designed to amalgamate challenges from diverse mathematical and visual tasks. We first taxonomize the key task types, reasoning skills, and visual contexts from the literature to guide our selection from 28 existing math-focused and visual question answering datasets. Then, we construct three new datasets, IQTest, FunctionQA, and PaperQA, to accommodate for missing types of visual contexts. The problems featured often require deep visual understanding beyond OCR or image captioning, and compositional reasoning with rich domain-specific tools, thus posing a notable challenge to existing models. We conduct a comprehensive evaluation of 11 prominent open-source and proprietary foundation models (LLMs, LLMs augmented with tools, and LMMs), and early experiments with GPT-4V. The best-performing model, Multimodal Bard, achieves only 58% of human performance (34.8% vs 60.3%), indicating ample room for further improvement. Given this significant gap, MathVista fuels future research in the development of general-purpose AI agents capable of tackling mathematically intensive and visually rich real-world tasks. Preliminary tests show that MathVista also presents challenges to GPT-4V, underscoring the benchmark's importance. The project is available at https://mathvista.github.io/.

Self-Verification Improves Few-Shot Clinical Information Extraction

Extracting patient information from unstructured text is a critical task in health decision-support and clinical research. Large language models (LLMs) have shown the potential to accelerate clinical curation via few-shot in-context learning, in contrast to supervised learning which requires much more costly human annotations. However, despite drastic advances in modern LLMs such as GPT-4, they still struggle with issues regarding accuracy and interpretability, especially in mission-critical domains such as health. Here, we explore a general mitigation framework using self-verification, which leverages the LLM to provide provenance for its own extraction and check its own outputs. This is made possible by the asymmetry between verification and generation, where the latter is often much easier than the former. Experimental results show that our method consistently improves accuracy for various LLMs in standard clinical information extraction tasks. Additionally, self-verification yields interpretations in the form of a short text span corresponding to each output, which makes it very efficient for human experts to audit the results, paving the way towards trustworthy extraction of clinical information in resource-constrained scenarios. To facilitate future research in this direction, we release our code and prompts.

Chameleon: Plug-and-Play Compositional Reasoning with Large Language Models

Large language models (LLMs) have achieved remarkable progress in various natural language processing tasks with emergent abilities. However, they face inherent limitations, such as an inability to access up-to-date information, utilize external tools, or perform precise mathematical reasoning. In this paper, we introduce Chameleon, a plug-and-play compositional reasoning framework that augments LLMs to help address these challenges. Chameleon synthesizes programs to compose various tools, including LLM models, off-the-shelf vision models, web search engines, Python functions, and rule-based modules tailored to user interests. Built on top of an LLM as a natural language planner, Chameleon infers the appropriate sequence of tools to compose and execute in order to generate a final response. We showcase the adaptability and effectiveness of Chameleon on two tasks: ScienceQA and TabMWP. Notably, Chameleon with GPT-4 achieves an 86.54% accuracy on ScienceQA, significantly improving upon the best published few-shot model by 11.37%; using GPT-4 as the underlying LLM, Chameleon achieves a 17.8% increase over the state-of-the-art model, leading to a 98.78% overall accuracy on TabMWP. Further studies suggest that using GPT-4 as a planner exhibits more consistent and rational tool selection and is able to infer potential constraints given the instructions, compared to other LLMs like ChatGPT.

Instruction Tuning with GPT-4

Prior work has shown that finetuning large language models (LLMs) using machine-generated instruction-following data enables such models to achieve remarkable zero-shot capabilities on new tasks, and no human-written instructions are needed. In this paper, we present the first attempt to use GPT-4 to generate instruction-following data for LLM finetuning. Our early experiments on instruction-tuned LLaMA models show that the 52K English and Chinese instruction-following data generated by GPT-4 leads to superior zero-shot performance on new tasks to the instruction-following data generated by previous state-of-the-art models. We also collect feedback and comparison data from GPT-4 to enable a comprehensive evaluation and reward model training. We make our data generated using GPT-4 as well as our codebase publicly available.

Interactive Text Generation

Users interact with text, image, code, or other editors on a daily basis. However, machine learning models are rarely trained in the settings that reflect the interactivity between users and their editor. This is understandable as training AI models with real users is not only slow and costly, but what these models learn may be specific to user interface design choices. Unfortunately, this means most of the research on text, code, and image generation has focused on non-interactive settings, whereby the model is expected to get everything right without accounting for any input from a user who may be willing to help. We introduce a new Interactive Text Generation task that allows training generation models interactively without the costs of involving real users, by using user simulators that provide edits that guide the model towards a given target text. We train our interactive models using Imitation Learning, and our experiments against competitive non-interactive generation models show that models trained interactively are superior to their non-interactive counterparts, even when all models are given the same budget of user inputs or edits.

Check Your Facts and Try Again: Improving Large Language Models with External Knowledge and Automated Feedback

Large language models (LLMs), such as ChatGPT, are able to generate human-like, fluent responses for many downstream tasks, e.g., task-oriented dialog and question answering. However, applying LLMs to real-world, mission-critical applications remains challenging mainly due to their tendency to generate hallucinations and their inability to use external knowledge. This paper proposes a LLM-Augmenter system, which augments a black-box LLM with a set of plug-and-play modules. Our system makes the LLM generate responses grounded in external knowledge, e.g., stored in task-specific databases. It also iteratively revises LLM prompts to improve model responses using feedback generated by utility functions, e.g., the factuality score of a LLM-generated response. The effectiveness of LLM-Augmenter is empirically validated on two types of scenarios, task-oriented dialog and open-domain question answering. LLM-Augmenter significantly reduces ChatGPT's hallucinations without sacrificing the fluency and informativeness of its responses. We make the source code and models publicly available.

Guiding Large Language Models via Directional Stimulus Prompting

We introduce a new framework, Directional Stimulus Prompting, that uses a tuneable language model (LM) to provide guidance for the black-box frozen large language model (LLM) on downstream tasks. Unlike prior work that manually or automatically finds the optimal prompt for each task, we train a policy LM to generate discrete tokens as ``directional stimulus'' of each input, which is a hint/cue such as keywords of an article for summarization. The directional stimulus is then combined with the original input and fed into the LLM to guide its generation toward the desired target. The policy LM can be trained through 1) supervised learning from annotated data and 2) reinforcement learning from offline and online rewards to explore directional stimulus that better aligns LLMs with human preferences. This framework is flexibly applicable to various LMs and tasks. To verify its effectiveness, we apply our framework to summarization and dialogue response generation tasks. Experimental results demonstrate that it can significantly improve LLMs' performance with a small collection of training data: a T5 (780M) trained with 2,000 samples from the CNN/Daily Mail dataset improves Codex (175B)'s performance by 7.2% in ROUGE-Avg scores; 500 dialogues boost the combined score by 52.5%, achieving comparable or even better performance than fully trained models on the MultiWOZ dataset.

Enhancing Task Bot Engagement with Synthesized Open-Domain Dialog

Many efforts have been made to construct dialog systems for different types of conversations, such as task-oriented dialog (TOD) and open-domain dialog (ODD). To better mimic human-level conversations that usually fuse various dialog modes, it is essential to build a system that can effectively handle both TOD and ODD and access different knowledge sources. To address the lack of available data for the fused task, we propose a framework for automatically generating dialogues that combine knowledge-grounded ODDs and TODs in various settings. Additionally, we introduce a unified model PivotBot that is capable of appropriately adopting TOD and ODD modes and accessing different knowledge sources in order to effectively tackle the fused task. Evaluation results demonstrate the superior ability of the proposed model to switch seamlessly between TOD and ODD tasks.