Muru Zhang

Ofir Press, Muru Zhang, Sewon Min et al. · allen-ai, uw

We investigate the ability of language models to perform compositional reasoning tasks where the overall solution depends on correctly composing the answers to sub-problems. We measure how often models can correctly answer all sub-problems but not generate the overall solution, a ratio we call the compositionality gap. We evaluate this ratio by asking multi-hop questions with answers that require composing multiple facts unlikely to have been observed together during pretraining. In the GPT-3 family of models, as model size increases we show that the single-hop question answering performance improves faster than the multi-hop performance does, therefore the compositionality gap does not decrease. This surprising result suggests that while more powerful models memorize and recall more factual knowledge, they show no corresponding improvement in their ability to perform this kind of compositional reasoning. We then demonstrate how elicitive prompting (such as chain of thought) narrows the compositionality gap by reasoning explicitly. We present a new method, self-ask, that further improves on chain of thought. In our method, the model explicitly asks itself (and answers) follow-up questions before answering the initial question. We finally show that self-ask's structured prompting lets us easily plug in a search engine to answer the follow-up questions, which additionally improves accuracy.

Aaron Walsman, Muru Zhang, Klemen Kotar et al.

Visual understanding of geometric structures with complex spatial relationships is a fundamental component of human intelligence. As children, we learn how to reason about structure not only from observation, but also by interacting with the world around us -- by taking things apart and putting them back together again. The ability to reason about structure and compositionality allows us to not only build things, but also understand and reverse-engineer complex systems. In order to advance research in interactive reasoning for part-based geometric understanding, we propose a challenging new assembly problem using LEGO bricks that we call Break and Make. In this problem an agent is given a LEGO model and attempts to understand its structure by interactively inspecting and disassembling it. After this inspection period, the agent must then prove its understanding by rebuilding the model from scratch using low-level action primitives. In order to facilitate research on this problem we have built LTRON, a fully interactive 3D simulator that allows learning agents to assemble, disassemble and manipulate LEGO models. We pair this simulator with a new dataset of fan-made LEGO creations that have been uploaded to the internet in order to provide complex scenes containing over a thousand unique brick shapes. We take a first step towards solving this problem using sequence-to-sequence models that provide guidance for how to make progress on this challenging problem. Our simulator and data are available at github.com/aaronwalsman/ltron. Additional training code and PyTorch examples are available at github.com/aaronwalsman/ltron-torch-eccv22.

Yi Li, Muru Zhang, Markus Grotz et al.

Segmentation and tracking of unseen object instances in discrete frames pose a significant challenge in dynamic industrial robotic contexts, such as distribution warehouses. Here, robots must handle object rearrangement, including shifting, removal, and partial occlusion by new items, and track these items after substantial temporal gaps. The task is further complicated when robots encounter objects not learned in their training sets, which requires the ability to segment and track previously unseen items. Considering that continuous observation is often inaccessible in such settings, our task involves working with a discrete set of frames separated by indefinite periods during which substantial changes to the scene may occur. This task also translates to domestic robotic applications, such as rearrangement of objects on a table. To address these demanding challenges, we introduce new synthetic and real-world datasets that replicate these industrial and household scenarios. We also propose a novel paradigm for joint segmentation and tracking in discrete frames along with a transformer module that facilitates efficient inter-frame communication. The experiments we conduct show that our approach significantly outperforms recent methods. For additional results and videos, please visit \href{https://sites.google.com/view/stow-corl23}{website}. Code and dataset will be released.

Guang Yang, Muru Zhang, Lin Qiu et al.

Optical music recognition (OMR) aims to convert music notation into digital formats. One approach to tackle OMR is through a multi-stage pipeline, where the system first detects visual music notation elements in the image (object detection) and then assembles them into a music notation (notation assembly). Most previous work on notation assembly unrealistically assumes perfect object detection. In this study, we focus on the MUSCIMA++ v2.0 dataset, which represents musical notation as a graph with pairwise relationships among detected music objects, and we consider both stages together. First, we introduce a music object detector based on YOLOv8, which improves detection performance. Second, we introduce a supervised training pipeline that completes the notation assembly stage based on detection output. We find that this model is able to outperform existing models trained on perfect detection output, showing the benefit of considering the detection and assembly stages in a more holistic way. These findings, together with our novel evaluation metric, are important steps toward a more complete OMR solution.

Muru Zhang, Ofir Press, William Merrill et al.

A major risk of using language models in practical applications is their tendency to hallucinate incorrect statements. Hallucinations are often attributed to knowledge gaps in LMs, but we hypothesize that in some cases, when justifying previously generated hallucinations, LMs output false claims that they can separately recognize as incorrect. We construct three question-answering datasets where ChatGPT and GPT-4 often state an incorrect answer and offer an explanation with at least one incorrect claim. Crucially, we find that ChatGPT and GPT-4 can identify 67% and 87% of their own mistakes, respectively. We refer to this phenomenon as hallucination snowballing: an LM over-commits to early mistakes, leading to more mistakes that it otherwise would not make.