Zilong Zheng

Generating ultra-long sequences with large language models (LLMs) has become increasingly crucial but remains a highly time-intensive task, particularly for sequences up to 100K tokens. While traditional speculative decoding methods exist, simply extending their generation limits fails to accelerate the process and can be detrimental. Through an in-depth analysis, we identify three major challenges hindering efficient generation: frequent model reloading, dynamic key-value (KV) management and repetitive generation. To address these issues, we introduce TOKENSWIFT, a novel framework designed to substantially accelerate the generation process of ultra-long sequences while maintaining the target model's inherent quality. Experimental results demonstrate that TOKENSWIFT achieves over 3 times speedup across models of varying scales (1.5B, 7B, 8B, 14B) and architectures (MHA, GQA). This acceleration translates to hours of time savings for ultra-long sequence generation, establishing TOKENSWIFT as a scalable and effective solution at unprecedented lengths. Code can be found at this URL.

Model collapse in synthetic data indicates that iterative training on self-generated data leads to a gradual decline in performance. With the proliferation of AI models, synthetic data will fundamentally reshape the web data ecosystem. Future GPT-{n} models will inevitably be trained on a blend of synthetic and human-produced data. In this paper, we focus on two questions: what is the impact of synthetic data on language model training, and how to synthesize data without model collapse? We first pre-train language models across different proportions of synthetic data, revealing a negative correlation between the proportion of synthetic data and model performance. We further conduct statistical analysis on synthetic data to uncover distributional shift phenomenon and over-concentration of n-gram features. Inspired by the above findings, we propose token editing on human-produced data to obtain semi-synthetic data. As a proof of concept, we theoretically demonstrate that token-level editing can prevent model collapse, as the test error is constrained by a finite upper bound. We conduct extensive experiments on pre-training from scratch, continual pre-training, and supervised fine-tuning. The results validate our theoretical proof that token-level editing improves data quality and enhances model performance.

Developing AI agents capable of interacting with open-world environments to solve diverse tasks is a compelling challenge. However, evaluating such open-ended agents remains difficult, with current benchmarks facing scalability limitations. To address this, we introduce Minecraft Universe (MCU), a comprehensive evaluation framework set within the open-world video game Minecraft. MCU incorporates three key components: (1) an expanding collection of 3,452 composable atomic tasks that encompasses 11 major categories and 41 subcategories of challenges; (2) a task composition mechanism capable of generating infinite diverse tasks with varying difficulty; and (3) a general evaluation framework that achieves 91.5% alignment with human ratings for open-ended task assessment. Empirical results reveal that even state-of-the-art foundation agents struggle with the increasing diversity and complexity of tasks. These findings highlight the necessity of MCU as a robust benchmark to drive progress in AI agent development within open-ended environments.

How to align large language models (LLMs) with user preferences from a static general dataset has been frequently studied. However, user preferences are usually personalized, changing, and diverse. This leads to the problem that the actual user preferences often do not coincide with those trained by the model developers in the practical use of LLMs. Since we cannot collect enough data and retrain for every demand, researching efficient real-time preference adaptation methods based on the backbone LLMs during test time is important. To this end, we introduce Amulet, a novel, training-free framework that formulates the decoding process of every token as a separate online learning problem with the guidance of simple user-provided prompts, thus enabling real-time optimization to satisfy users' personalized preferences. To reduce the computational cost brought by this optimization process for each token, we additionally provide a closed-form solution for each iteration step of the optimization process, thereby reducing the computational time cost to a negligible level. The detailed experimental results demonstrate that Amulet can achieve significant performance improvements in rich settings with combinations of different LLMs, datasets, and user preferences, while maintaining acceptable computational efficiency.

Knowledge editing techniques have emerged as essential tools for updating the factual knowledge of large language models (LLMs) and multimodal models (LMMs), allowing them to correct outdated or inaccurate information without retraining from scratch. However, existing benchmarks for multimodal knowledge editing primarily focus on entity-level knowledge represented as simple triplets, which fail to capture the complexity of real-world multimodal information. To address this issue, we introduce MMKE-Bench, a comprehensive MultiModal Knowledge Editing Benchmark, designed to evaluate the ability of LMMs to edit diverse visual knowledge in real-world scenarios. MMKE-Bench addresses these limitations by incorporating three types of editing tasks: visual entity editing, visual semantic editing, and user-specific editing. Besides, MMKE-Bench uses free-form natural language to represent and edit knowledge, offering a more flexible and effective format. The benchmark consists of 2,940 pieces of knowledge and 7,229 images across 110 fine-grained types, with evaluation questions automatically generated and human-verified. We assess five state-of-the-art knowledge editing methods on three prominent LMMs, revealing that no method excels across all criteria, and that visual and user-specific edits are particularly challenging. MMKE-Bench sets a new standard for evaluating the robustness of multimodal knowledge editing techniques, driving progress in this rapidly evolving field.

The existing fine-tuning paradigm for language models is brittle in knowledge editing scenarios, where the model must incorporate new information without extensive retraining. This brittleness often results in overfitting, reduced performance, and unnatural language generation. To address this, we propose Consistent In-Context Editing (ICE), a novel approach that leverages the model's in-context learning capability to tune toward a contextual distribution rather than a one-hot target. ICE introduces a straightforward optimization framework that includes both a target and a procedure, enhancing the robustness and effectiveness of gradient-based tuning methods. We provide analytical insights into ICE across four critical aspects of knowledge editing: accuracy, locality, generalization, and linguistic quality, showing its advantages. Experimental results across four datasets confirm the effectiveness of ICE and demonstrate its potential for continual editing, ensuring that updated information is incorporated while preserving the integrity of the model.

The rapid advancement of multi-modal language models (MLLMs) like GPT-4o has propelled the development of Omni language models, designed to process and proactively respond to continuous streams of multi-modal data. Despite their potential, evaluating their real-world interactive capabilities in streaming video contexts remains a formidable challenge. In this work, we introduce OmniMMI, a comprehensive multi-modal interaction benchmark tailored for OmniLLMs in streaming video contexts. OmniMMI encompasses over 1,121 real-world interactive videos and 2,290 questions, addressing two critical yet underexplored challenges in existing video benchmarks: streaming video understanding and proactive reasoning, across six distinct subtasks. Moreover, we propose a novel framework, Multi-modal Multiplexing Modeling (M4), designed to enhance real-time interactive reasoning with minimum finetuning on pre-trained MLLMs. Extensive experimental results reveal that the existing MLLMs fall short in interactive streaming understanding, particularly struggling with proactive tasks and multi-turn queries. Our proposed M4, though lightweight, demonstrates a significant improvement in handling proactive tasks and real-time interactions.

Recent advances in large language models (LLMs) have made them indispensable, raising significant concerns over managing their safety. Automated red teaming offers a promising alternative to the labor-intensive and error-prone manual probing for vulnerabilities, providing more consistent and scalable safety evaluations. However, existing approaches often compromise diversity by focusing on maximizing attack success rate. Additionally, methods that decrease the cosine similarity from historical embeddings with semantic diversity rewards lead to novelty stagnation as history grows. To address these issues, we introduce DiveR-CT, which relaxes conventional constraints on the objective and semantic reward, granting greater freedom for the policy to enhance diversity. Our experiments demonstrate DiveR-CT's marked superiority over baselines by 1) generating data that perform better in various diversity metrics across different attack success rate levels, 2) better-enhancing resiliency in blue team models through safety tuning based on collected data, 3) allowing dynamic control of objective weights for reliable and controllable attack success rates, and 4) reducing susceptibility to reward overoptimization. Project details and code can be found at https://andrewzh112.github.io/#diverct.

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 (>> 4K) and struggle to effectively utilize information from the middle part of the context. To address these issues, we propose Continuity-Relativity indExing with gAussian Middle (CREAM), which interpolates positional encodings by manipulating position indices. Apart from being simple, 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 CREAM successfully extends LLMs to the target length for both Base and Chat versions of Llama2-7B with "Never Miss A Beat".

Accommodating long sequences efficiently in autoregressive Transformers, especially within an extended context window, poses significant challenges due to the quadratic computational complexity and substantial KV memory requirements inherent in self-attention mechanisms. In this work, we introduce SPARSEK Attention, a novel sparse attention mechanism designed to overcome these computational and memory obstacles while maintaining performance. Our approach integrates a scoring network and a differentiable top-k mask operator, SPARSEK, to select a constant number of KV pairs for each query, thereby enabling gradient-based optimization. As a result, SPARSEK Attention offers linear time complexity and constant memory footprint during generation. Experimental results reveal that SPARSEK Attention outperforms previous sparse attention methods and provides significant speed improvements during both training and inference, particularly in language modeling and downstream tasks. Furthermore, our method can be seamlessly integrated into pre-trained Large Language Models (LLMs) with minimal fine-tuning, offering a practical solution for effectively managing long-range dependencies in diverse applications.

A prerequisite for social coordination is bidirectional communication between teammates, each playing two roles simultaneously: as receptive listeners and expressive speakers. For robots working with humans in complex situations with multiple goals that differ in importance, failure to fulfill the expectation of either role could undermine group performance due to misalignment of values between humans and robots. Specifically, a robot needs to serve as an effective listener to infer human users’ intents from instructions and feedback and as an expressive speaker to explain its decision processes to users. Here, we investigate how to foster effective bidirectional human-robot communications in the context of value alignment—collaborative robots and users form an aligned understanding of the importance of possible task goals. We propose an explainable artificial intelligence (XAI) system in which a group of robots predicts users’ values by taking in situ feedback into consideration while communicating their decision processes to users through explanations. To learn from human feedback, our XAI system integrates a cooperative communication model for inferring human values associated with multiple desirable goals. To be interpretable to humans, the system simulates human mental dynamics and predicts optimal explanations using graphical models. We conducted psychological experiments to examine the core components of the proposed computational framework. Our results show that real-time human-robot mutual understanding in complex cooperative tasks is achievable with a learning model based on bidirectional communication. We believe that this interaction framework can shed light on bidirectional value alignment in communicative XAI systems and, more broadly, in future human-machine teaming systems. An explainable artificial intelligence collaboration framework enables in situ bidirectional human-robot value alignment.

Exploiting internal statistics of a single natural image has long been recognized as a significant research paradigm where the goal is to learn the internal distribution of patches within the image without relying on external training data. Different from prior works that model such a distribution implicitly with a top-down latent variable model (e.g., generator), this paper proposes to explicitly represent the statistical distribution within a single natural image by using an energy-based generative framework, where a pyramid of energy functions, each parameterized by a bottom-up deep neural network, are used to capture the distributions of patches at different resolutions. Meanwhile, a coarse-to-fine sequential training and sampling strategy is presented to train the model efficiently. Besides learning to generate random samples from white noise, the model can learn in parallel with a self-supervised task (e.g., recover the input image from its corrupted version), which can further improve the descriptive power of the learned model. The proposed model is simple and natural in that it does not require an auxiliary model (e.g., discriminator) to assist the training. Besides, it also unifies internal statistics learning and image generation in a single framework. Experimental results presented on various image generation and manipulation tasks, including super-resolution, image editing, harmonization, style transfer, etc., have demonstrated the effectiveness of our model for internal learning.

We propose a novel model to address the task of Visual Dialog which exhibits complex dialog structures. To obtain a reasonable answer based on the current question and the dialog history, the underlying semantic dependencies between dialog entities are essential. In this paper, we explicitly formalize this task as inference in a graphical model with partially observed nodes and unknown graph structures (relations in dialog). The given dialog entities are viewed as the observed nodes. The answer to a given question is represented by a node with missing value. We first introduce an Expectation Maximization algorithm to infer both the underlying dialog structures and the missing node values (desired answers). Based on this, we proceed to propose a differentiable graph neural network (GNN) solution that approximates this process. Experiment results on the VisDial and VisDial-Q datasets show that our model outperforms comparative methods. It is also observed that our method can infer the underlying dialog structure for better dialog reasoning.

This paper proposes a 3D shape descriptor network, which is a deep convolutional energy-based model, for modeling volumetric shape patterns. The maximum likelihood training of the model follows an “analysis by synthesis” scheme and can be interpreted as a mode seeking and mode shifting process. The model can synthesize 3D shape patterns by sampling from the probability distribution via MCMC such as Langevin dynamics. The model can be used to train a 3D generator network via MCMC teaching. The conditional version of the 3D shape descriptor net can be used for 3D object recovery and 3D object super-resolution. Experiments demonstrate that the proposed model can generate realistic 3D shape patterns and can be useful for 3D shape analysis.