Yanpeng Zhao

Yanpeng Zhao, Wentao Ding, Hongtao Li et al.

A recent trend in vision-language models (VLMs) has been to enhance their spatial cognition for embodied domains. Despite progress, existing evaluations have been limited both in paradigm and in coverage, hindering rapid, iterative model development. To address these limitations, we propose ESPIRE, a diagnostic benchmark for embodied spatial reasoning. ESPIRE offers a simulated world that physically grounds VLMs and evaluates them on spatial-reasoning-centric robotic tasks, thus narrowing the gap between evaluation and real-world deployment. To adapt VLMs to robotic tasks, we decompose each task into localization and execution, and frame both as generative problems, in stark contrast to predominant discriminative evaluations (e.g., via visual-question answering) that rely on distractors and discard execution. This decomposition further enables a fine-grained analysis beyond passive spatial reasoning toward reasoning to act. We systematically design ESPIRE both at the instruction level and at the environment level, ensuring broad coverage of spatial reasoning scenarios. We use ESPIRE to diagnose a range of frontier VLMs and provide in-depth analysis of their spatial reasoning behaviors.

Yanpeng Zhao, Yiwei Hao, Siyu Gao et al.

Learning object-centric representations from unsupervised videos is challenging. Unlike most previous approaches that focus on decomposing 2D images, we present a 3D generative model named DynaVol-S for dynamic scenes that enables object-centric learning within a differentiable volume rendering framework. The key idea is to perform object-centric voxelization to capture the 3D nature of the scene, which infers per-object occupancy probabilities at individual spatial locations. These voxel features evolve through a canonical-space deformation function and are optimized in an inverse rendering pipeline with a compositional NeRF. Additionally, our approach integrates 2D semantic features to create 3D semantic grids, representing the scene through multiple disentangled voxel grids. DynaVol-S significantly outperforms existing models in both novel view synthesis and unsupervised decomposition tasks for dynamic scenes. By jointly considering geometric structures and semantic features, it effectively addresses challenging real-world scenarios involving complex object interactions. Furthermore, once trained, the explicitly meaningful voxel features enable additional capabilities that 2D scene decomposition methods cannot achieve, such as novel scene generation through editing geometric shapes or manipulating the motion trajectories of objects.

Yanpeng Zhao, Siyu Gao, Yunbo Wang et al.

Unsupervised learning of object-centric representations in dynamic visual scenes is challenging. Unlike most previous approaches that learn to decompose 2D images, we present DynaVol, a 3D scene generative model that unifies geometric structures and object-centric learning in a differentiable volume rendering framework. The key idea is to perform object-centric voxelization to capture the 3D nature of the scene, which infers the probability distribution over objects at individual spatial locations. These voxel features evolve over time through a canonical-space deformation function, forming the basis for global representation learning via slot attention. The voxel features and global features are complementary and are both leveraged by a compositional NeRF decoder for volume rendering. DynaVol remarkably outperforms existing approaches for unsupervised dynamic scene decomposition. Once trained, the explicitly meaningful voxel features enable additional capabilities that 2D scene decomposition methods cannot achieve: it is possible to freely edit the geometric shapes or manipulate the motion trajectories of the objects.

Zhimu Zhou, Yanpeng Zhao, Qiuyu Liao et al.

Visual planning represents a crucial facet of human intelligence, especially in tasks that require complex spatial reasoning and navigation. Yet, in machine learning, this inherently visual problem is often tackled through a verbal-centric lens. While recent research demonstrates the promise of fully visual approaches, they suffer from significant computational inefficiency due to the step-by-step planning-by-generation paradigm. In this work, we present EAR, an editing-as-reasoning paradigm that reformulates visual planning as a single-step image transformation. To isolate intrinsic reasoning from visual recognition, we employ abstract puzzles as probing tasks and introduce AMAZE, a procedurally generated dataset that features the classical Maze and Queen problems, covering distinct, complementary forms of visual planning. The abstract nature of AMAZE also facilitates automatic evaluation of autoregressive and diffusion-based models in terms of both pixel-wise fidelity and logical validity. We assess leading proprietary and open-source editing models. The results show that they all struggle in the zero-shot setting, finetuning on basic scales enables remarkable generalization to larger in-domain scales and out-of-domain scales and geometries. However, our best model that runs on high-end hardware fails to match the zero-shot efficiency of human solvers, highlighting a persistent gap in neural visual reasoning.

Yanpeng Zhao, Ivan Titov

There has been a significant surge of interest in visually grounded grammar induction in recent times. While a variety of models have been developed for the task and have demonstrated impressive performance, they have not been evaluated on text domains that are different from the training domain, so it is unclear if the improvements brought by visual groundings are transferable. Our study aims to fill this gap and assess the degree of transferability. We start by extending VC-PCFG (short for Visually-grounded Compound PCFG~\citep{zhao-titov-2020-visually}) in such a way that it can transfer across text domains. We consider a zero-shot transfer learning setting where a model is trained on the source domain and is directly applied to target domains, without any further training. Our experimental results suggest that: the benefits from using visual groundings transfer to text in a domain similar to the training domain but fail to transfer to remote domains. Further, we conduct data and result analysis; we find that the lexicon overlap between the source domain and the target domain is the most important factor in the transferability of VC-PCFG.

Zhengpeng Shi, Hengli Li, Yanpeng Zhao et al.

AI models capable of comprehending humor hold real-world promise -- for example, enhancing engagement in human-machine interactions. To gauge and diagnose the capacity of multimodal large language models (MLLMs) for humor understanding, we introduce v-HUB, a novel visual-centric video humor understanding benchmark. v-HUB comprises a curated collection of minimally verbal short videos, sourced from classic silent films and online resources, and reflecting real-world scenarios where humor can be appreciated purely through visual cues. Each video clip is paired with rich annotations, including captions, descriptions, and explanations, supporting evaluation tasks like caption matching and humor explanation. To broaden its applicability, we further construct an open-ended video QA task, making it readily integrable into existing video understanding benchmarks. We evaluate a diverse set of MLLMs, from specialized Video-LLMs to versatile OmniLLMs that can process audio, covering both open-source and proprietary domains. The experimental results expose the difficulties MLLMs face in comprehending humor from visual cues alone. For example, all models exhibit a marked performance drop on caption matching when moving from text-based to video-based evaluation (without audio). Our findings also demonstrate that incorporating audio helps with video humor understanding, highlighting the informativeness of sound and the promise of integrating richer modalities for complex video understanding tasks.

Tong Wu, Yanpeng Zhao, Zilong Zheng

Recently, many methods have been developed to extend the context length of pre-trained large language models (LLMs), but they often require fine-tuning at the target length ($\gg4K$) and struggle to effectively utilize information from the middle part of the context. To address these issues, we propose $\textbf{C}$ontinuity-$\textbf{R}$elativity ind$\textbf{E}$xing with g$\textbf{A}$ussian $\textbf{M}$iddle ($\texttt{CREAM}$), which interpolates positional encodings by manipulating position indices. Apart from being simple, $\texttt{CREAM}$ is training-efficient: it only requires fine-tuning at the pre-trained context window (e.g., Llama 2-4K) and can extend LLMs to a much longer target context length (e.g., 256K). To ensure that the model focuses more on the information in the middle, we introduce a truncated Gaussian to encourage sampling from the middle part of the context during fine-tuning, thus alleviating the "Lost-in-the-Middle" problem faced by long-context LLMs. Experimental results show that $\texttt{CREAM}$ successfully extends LLMs to the target length for both Base and Chat versions of $\texttt{Llama2-7B}$ with "Never Miss A Beat". Our code is publicly available at https://github.com/bigai-nlco/cream.

Songlin Yang, Yanpeng Zhao, Kewei Tu

Probabilistic context-free grammars (PCFGs) with neural parameterization have been shown to be effective in unsupervised phrase-structure grammar induction. However, due to the cubic computational complexity of PCFG representation and parsing, previous approaches cannot scale up to a relatively large number of (nonterminal and preterminal) symbols. In this work, we present a new parameterization form of PCFGs based on tensor decomposition, which has at most quadratic computational complexity in the symbol number and therefore allows us to use a much larger number of symbols. We further use neural parameterization for the new form to improve unsupervised parsing performance. We evaluate our model across ten languages and empirically demonstrate the effectiveness of using more symbols. Our code: https://github.com/sustcsonglin/TN-PCFG

Yanpeng Zhao, Liwen Zhang, Kewei Tu

We introduce Latent Vector Grammars (LVeGs), a new framework that extends latent variable grammars such that each nonterminal symbol is associated with a continuous vector space representing the set of (infinitely many) subtypes of the nonterminal. We show that previous models such as latent variable grammars and compositional vector grammars can be interpreted as special cases of LVeGs. We then present Gaussian Mixture LVeGs (GM-LVeGs), a new special case of LVeGs that uses Gaussian mixtures to formulate the weights of production rules over subtypes of nonterminals. A major advantage of using Gaussian mixtures is that the partition function and the expectations of subtype rules can be computed using an extension of the inside-outside algorithm, which enables efficient inference and learning. We apply GM-LVeGs to part-of-speech tagging and constituency parsing and show that GM-LVeGs can achieve competitive accuracies. Our code is available at https://github.com/zhaoyanpeng/lveg.

Chen Zhu, Yanpeng Zhao, Shuaiyi Huang et al.

Visual attention, which assigns weights to image regions according to their relevance to a question, is considered as an indispensable part by most Visual Question Answering models. Although the questions may involve complex relations among multiple regions, few attention models can effectively encode such cross-region relations. In this paper, we demonstrate the importance of encoding such relations by showing the limited effective receptive field of ResNet on two datasets, and propose to model the visual attention as a multivariate distribution over a grid-structured Conditional Random Field on image regions. We demonstrate how to convert the iterative inference algorithms, Mean Field and Loopy Belief Propagation, as recurrent layers of an end-to-end neural network. We empirically evaluated our model on 3 datasets, in which it surpasses the best baseline model of the newly released CLEVR dataset by 9.5%, and the best published model on the VQA dataset by 1.25%. Source code is available at https: //github.com/zhuchen03/vqa-sva.

Yapeng Mi, Hengli Li, Yanpeng Zhao et al.

Reasoning-augmented machine learning systems have shown improved performance in various domains, including image generation. However, existing reasoning-based methods for image generation either restrict reasoning to a single modality (image or text) or rely on high-quality reasoning data for fine-tuning. To tackle these limitations, we propose MILR, a test-time method that jointly reasons over image and text in a unified latent vector space. Reasoning in MILR is performed by searching through vector representations of discrete image and text tokens. Practically, this is implemented via the policy gradient method, guided by an image quality critic. We instantiate MILR within the unified multimodal understanding and generation (MUG) framework that natively supports language reasoning before image synthesis and thus facilitates cross-modal reasoning. The intermediate model outputs, which are to be optimized, serve as the unified latent space, enabling MILR to operate entirely at test time. We evaluate MILR on GenEval, T2I-CompBench, and WISE, achieving state-of-the-art results on all benchmarks. Notably, on knowledge-intensive WISE, MILR attains an overall score of 0.63, improving over the baseline by 80%. Our further analysis indicates that joint reasoning in the unified latent space is the key to its strong performance. Moreover, our qualitative studies reveal MILR's non-trivial ability in temporal and cultural reasoning, highlighting the efficacy of our reasoning method.

Huimin Wu, Xiaojian Ma, Haozhe Zhao et al.

Text-guided image editing involves modifying a source image based on a language instruction and, typically, requires changes to only small local regions. However, existing approaches generate the entire target image rather than selectively regenerate only the intended editing areas. This results in (1) unnecessary computational costs and (2) a bias toward reconstructing non-editing regions, which compromises the quality of the intended edits. To resolve these limitations, we propose to formulate image editing as Next Editing-token Prediction (NEP) based on autoregressive image generation, where only regions that need to be edited are regenerated, thus avoiding unintended modification to the non-editing areas. To enable any-region editing, we propose to pre-train an any-order autoregressive text-to-image (T2I) model. Once trained, it is capable of zero-shot image editing and can be easily adapted to NEP for image editing, which achieves a new state-of-the-art on widely used image editing benchmarks. Moreover, our model naturally supports test-time scaling (TTS) through iteratively refining its generation in a zero-shot manner. The project page is: https://nep-bigai.github.io/

Yanpeng Zhao, Shanyan Guan, Yunbo Wang et al.

We introduce NeoWorld, a deep learning framework for generating interactive 3D virtual worlds from a single input image. Inspired by the on-demand worldbuilding concept in the science fiction novel Simulacron-3 (1964), our system constructs expansive environments where only the regions actively explored by the user are rendered with high visual realism through object-centric 3D representations. Unlike previous approaches that rely on global world generation or 2D hallucination, NeoWorld models key foreground objects in full 3D, while synthesizing backgrounds and non-interacted regions in 2D to ensure efficiency. This hybrid scene structure, implemented with cutting-edge representation learning and object-to-3D techniques, enables flexible viewpoint manipulation and physically plausible scene animation, allowing users to control object appearance and dynamics using natural language commands. As users interact with the environment, the virtual world progressively unfolds with increasing 3D detail, delivering a dynamic, immersive, and visually coherent exploration experience. NeoWorld significantly outperforms existing 2D and depth-layered 2.5D methods on the WorldScore benchmark.

Rowan Zellers, Jiasen Lu, Ximing Lu et al.

As humans, we navigate a multimodal world, building a holistic understanding from all our senses. We introduce MERLOT Reserve, a model that represents videos jointly over time -- through a new training objective that learns from audio, subtitles, and video frames. Given a video, we replace snippets of text and audio with a MASK token; the model learns by choosing the correct masked-out snippet. Our objective learns faster than alternatives, and performs well at scale: we pretrain on 20 million YouTube videos. Empirical results show that MERLOT Reserve learns strong multimodal representations. When finetuned, it sets state-of-the-art on Visual Commonsense Reasoning (VCR), TVQA, and Kinetics-600; outperforming prior work by 5%, 7%, and 1.5% respectively. Ablations show that these tasks benefit from audio pretraining -- even VCR, a QA task centered around images (without sound). Moreover, our objective enables out-of-the-box prediction, revealing strong multimodal commonsense understanding. In a fully zero-shot setting, our model obtains competitive results on four video tasks, even outperforming supervised approaches on the recently proposed Situated Reasoning (STAR) benchmark. We analyze why audio enables better vision-language representations, suggesting significant opportunities for future research. We conclude by discussing ethical and societal implications of multimodal pretraining.

Yanpeng Zhao, Jack Hessel, Youngjae Yu et al.

Machines that can represent and describe environmental soundscapes have practical potential, e.g., for audio tagging and captioning systems. Prevailing learning paradigms have been relying on parallel audio-text data, which is, however, scarcely available on the web. We propose VIP-ANT that induces \textbf{A}udio-\textbf{T}ext alignment without using any parallel audio-text data. Our key idea is to share the image modality between bi-modal image-text representations and bi-modal image-audio representations; the image modality functions as a pivot and connects audio and text in a tri-modal embedding space implicitly. In a difficult zero-shot setting with no paired audio-text data, our model demonstrates state-of-the-art zero-shot performance on the ESC50 and US8K audio classification tasks, and even surpasses the supervised state of the art for Clotho caption retrieval (with audio queries) by 2.2\% R@1. We further investigate cases of minimal audio-text supervision, finding that, e.g., just a few hundred supervised audio-text pairs increase the zero-shot audio classification accuracy by 8\% on US8K. However, to match human parity on some zero-shot tasks, our empirical scaling experiments suggest that we would need about $2^{21} \approx 2M$ supervised audio-caption pairs. Our work opens up new avenues for learning audio-text connections with little to no parallel audio-text data.

Songlin Yang, Yanpeng Zhao, Kewei Tu

Neural lexicalized PCFGs (L-PCFGs) have been shown effective in grammar induction. However, to reduce computational complexity, they make a strong independence assumption on the generation of the child word and thus bilexical dependencies are ignored. In this paper, we propose an approach to parameterize L-PCFGs without making implausible independence assumptions. Our approach directly models bilexical dependencies and meanwhile reduces both learning and representation complexities of L-PCFGs. Experimental results on the English WSJ dataset confirm the effectiveness of our approach in improving both running speed and unsupervised parsing performance.

Yanpeng Zhao, Ivan Titov

Compound probabilistic context-free grammars (C-PCFGs) have recently established a new state of the art for unsupervised phrase-structure grammar induction. However, due to the high space and time complexities of chart-based representation and inference, it is difficult to investigate C-PCFGs comprehensively. In this work, we rely on a fast implementation of C-PCFGs to conduct an evaluation complementary to that of~\citet{kim-etal-2019-compound}. We start by analyzing and ablating C-PCFGs on English treebanks. Our findings suggest that (1) C-PCFGs are data-efficient and can generalize to unseen sentence/constituent lengths; and (2) C-PCFGs make the best use of sentence-level information in generating preterminal rule probabilities. We further conduct a multilingual evaluation of C-PCFGs. The experimental results show that the best configurations of C-PCFGs, which are tuned on English, do not always generalize to morphology-rich languages.

Yanpeng Zhao, Ivan Titov

Exploiting visual groundings for language understanding has recently been drawing much attention. In this work, we study visually grounded grammar induction and learn a constituency parser from both unlabeled text and its visual groundings. Existing work on this task (Shi et al., 2019) optimizes a parser via Reinforce and derives the learning signal only from the alignment of images and sentences. While their model is relatively accurate overall, its error distribution is very uneven, with low performance on certain constituents types (e.g., 26.2% recall on verb phrases, VPs) and high on others (e.g., 79.6% recall on noun phrases, NPs). This is not surprising as the learning signal is likely insufficient for deriving all aspects of phrase-structure syntax and gradient estimates are noisy. We show that using an extension of probabilistic context-free grammar model we can do fully-differentiable end-to-end visually grounded learning. Additionally, this enables us to complement the image-text alignment loss with a language modeling objective. On the MSCOCO test captions, our model establishes a new state of the art, outperforming its non-grounded version and, thus, confirming the effectiveness of visual groundings in constituency grammar induction. It also substantially outperforms the previous grounded model, with largest improvements on more `abstract' categories (e.g., +55.1% recall on VPs).

Yanpeng Zhao, Ivan Titov

Semantic role labeling (SRL) is an NLP task involving the assignment of predicate arguments to types, called semantic roles. Though research on SRL has primarily focused on verbal predicates and many resources available for SRL provide annotations only for verbs, semantic relations are often triggered by other linguistic constructions, e.g., nominalizations. In this work, we investigate a transfer scenario where we assume role-annotated data for the source verbal domain but only unlabeled data for the target nominal domain. Our key assumption, enabling the transfer between the two domains, is that selectional preferences of a role (i.e., preferences or constraints on the admissible arguments) do not strongly depend on whether the relation is triggered by a verb or a noun. For example, the same set of arguments can fill the Acquirer role for the verbal predicate `acquire' and its nominal form `acquisition'. We approach the transfer task from the variational autoencoding perspective. The labeler serves as an encoder (predicting role labels given a sentence), whereas selectional preferences are captured in the decoder component (generating arguments for the predicting roles). Nominal roles are not labeled in the training data, and the learning objective instead pushes the labeler to assign roles predictive of the arguments. Sharing the decoder parameters across the domains encourages consistency between labels predicted for both domains and facilitates the transfer. The method substantially outperforms baselines, such as unsupervised and `direct transfer' methods, on the English CoNLL-2009 dataset.

Yanpeng Zhao, Wei Bi, Deng Cai et al.

Language style transfer is the problem of migrating the content of a source sentence to a target style. In many of its applications, parallel training data are not available and source sentences to be transferred may have arbitrary and unknown styles. First, each sentence is encoded into its content and style latent representations. Then, by recombining the content with the target style, we decode a sentence aligned in the target domain. To adequately constrain the encoding and decoding functions, we couple them with two loss functions. The first is a style discrepancy loss, enforcing that the style representation accurately encodes the style information guided by the discrepancy between the sentence style and the target style. The second is a cycle consistency loss, which ensures that the transferred sentence should preserve the content of the original sentence disentangled from its style. We validate the effectiveness of our model in three tasks: sentiment modification of restaurant reviews, dialog response revision with a romantic style, and sentence rewriting with a Shakespearean style.