The autoregressive nature of conventional large language models (LLMs) inherently limits inference speed, as tokens are generated sequentially. While speculative and parallel decoding techniques attempt to mitigate this, they face limitations: either relying on less accurate smaller models for generation or failing to fully leverage the base LLM's representations. We introduce a novel architecture, Tandem transformers, to address these issues. This architecture uniquely combines (1) a small autoregressive model and (2) a large model operating in block mode (processing multiple tokens simultaneously). The small model's predictive accuracy is substantially enhanced by granting it attention to the large model's richer representations. On the PaLM2 pretraining dataset, a tandem of PaLM2-Bison and PaLM2-Gecko demonstrates a 3.3% improvement in next-token prediction accuracy over a standalone PaLM2-Gecko, offering a 1.16x speedup compared to a PaLM2-Otter model with comparable downstream performance. We further incorporate the tandem model within the speculative decoding (SPEED) framework where the large model validates tokens from the small model. This ensures that the Tandem of PaLM2-Bison and PaLM2-Gecko achieves substantial speedup (around 1.14x faster than using vanilla PaLM2-Gecko in SPEED) while maintaining identical downstream task accuracy.

Backdoor attacks are commonly executed by contaminating training data, such that a trigger can activate predetermined harmful effects during the test phase. In this work, we present AnyDoor, a test-time backdoor attack against multimodal large language models (MLLMs), which involves injecting the backdoor into the textual modality using adversarial test images (sharing the same universal perturbation), without requiring access to or modification of the training data. AnyDoor employs similar techniques used in universal adversarial attacks, but distinguishes itself by its ability to decouple the timing of setup and activation of harmful effects. In our experiments, we validate the effectiveness of AnyDoor against popular MLLMs such as LLaVA-1.5, MiniGPT-4, InstructBLIP, and BLIP-2, as well as provide comprehensive ablation studies. Notably, because the backdoor is injected by a universal perturbation, AnyDoor can dynamically change its backdoor trigger prompts/harmful effects, exposing a new challenge for defending against backdoor attacks. Our project page is available at https://sail-sg.github.io/AnyDoor/.

Foundation models have recently expanded into robotics after excelling in computer vision and natural language processing. The models are accessible in two ways: open-source or paid, closed-source options. Users with access to both face a problem when deciding between effective yet costly closed-source models and free but less powerful open-source alternatives. We call it the model selection problem. Existing supervised-learning methods are impractical due to the high cost of collecting extensive training data from closed-source models. Hence, we focus on the online learning setting where algorithms learn while collecting data, eliminating the need for large pre-collected datasets. We thus formulate a user-centric online model selection problem and propose a novel solution that combines an open-source encoder to output context and an online learning algorithm that processes this context. The encoder distills vast data distributions into low-dimensional features, i.e., the context, without additional training. The online learning algorithm aims to maximize a composite reward that includes model performance, execution time, and costs based on the context extracted from the data. It results in an improved trade-off between selecting open-source and closed-source models compared to non-contextual methods, as validated by our theoretical analysis. Experiments across language-based robotic tasks such as Waymo Open Dataset, ALFRED, and Open X-Embodiment demonstrate real-world applications of the solution. The results show that the solution significantly improves the task success rate by up to 14%.

A multimodal large language model (MLLM) agent can receive instructions, capture images, retrieve histories from memory, and decide which tools to use. Nonetheless, red-teaming efforts have revealed that adversarial images/prompts can jailbreak an MLLM and cause unaligned behaviors. In this work, we report an even more severe safety issue in multi-agent environments, referred to as infectious jailbreak. It entails the adversary simply jailbreaking a single agent, and without any further intervention from the adversary, (almost) all agents will become infected exponentially fast and exhibit harmful behaviors. To validate the feasibility of infectious jailbreak, we simulate multi-agent environments containing up to one million LLaVA-1.5 agents, and employ randomized pair-wise chat as a proof-of-concept instantiation for multi-agent interaction. Our results show that feeding an (infectious) adversarial image into the memory of any randomly chosen agent is sufficient to achieve infectious jailbreak. Finally, we derive a simple principle for determining whether a defense mechanism can provably restrain the spread of infectious jailbreak, but how to design a practical defense that meets this principle remains an open question to investigate. Our project page is available at https://sail-sg.github.io/Agent-Smith/.

Parameter-efficient tuning (PEFT) techniques like low-rank adaptation (LoRA) offer training efficiency on Large Language Models, but their impact on model performance remains limited. Recent efforts integrate LoRA and Mixture-of-Experts (MoE) to improve the performance of PEFT methods. Despite promising results, research on improving the efficiency of LoRA with MoE is still in its early stages. Recent studies have shown that experts in the MoE architecture have different strengths and also exhibit some redundancy. Does this statement also apply to parameter-efficient MoE? In this paper, we introduce a novel parameter-efficient MoE method, \textit{\textbf{M}oE-L\textbf{o}RA with \textbf{L}ayer-wise Expert \textbf{A}llocation (MoLA)} for Transformer-based models, where each model layer has the flexibility to employ a varying number of LoRA experts. We investigate several architectures with varying layer-wise expert configurations. Experiments on six well-known NLP and commonsense QA benchmarks demonstrate that MoLA achieves equal or superior performance compared to all baselines. We find that allocating more LoRA experts to higher layers further enhances the effectiveness of models with a certain number of experts in total. With much fewer parameters, this allocation strategy outperforms the setting with the same number of experts in every layer. This work can be widely used as a plug-and-play parameter-efficient tuning approach for various applications. The code is available at https://github.com/GCYZSL/MoLA.

Robotic planning algorithms direct agents to perform actions within diverse environments to accomplish a task. Large Language Models (LLMs) like PaLM 2, GPT-3.5, and GPT-4 have revolutionized this domain, using their embedded real-world knowledge to tackle complex tasks involving multiple agents and objects. This paper introduces an innovative planning algorithm that integrates LLMs into the robotics context, enhancing task-focused execution and success rates. Key to our algorithm is a closed-loop feedback which provides real-time environmental states and error messages, crucial for refining plans when discrepancies arise. The algorithm draws inspiration from the human neural system, emulating its brain-body architecture by dividing planning across two LLMs in a structured, hierarchical fashion. Our method not only surpasses baselines within the VirtualHome Environment, registering a notable 35% average increase in task-oriented success rates, but achieves an impressive execution score of 85%, approaching the human-level benchmark of 94%. Moreover, effectiveness of the algorithm in real robot scenarios is shown using a realistic physics simulator and the Franka Research 3 Arm.

Background: Colonoscopy, a crucial diagnostic tool in gastroenterology, depends heavily on superior bowel preparation. ChatGPT, a large language model with emergent intelligence which also exhibits potential in medical applications. This study aims to assess the accuracy and consistency of ChatGPT in using the Boston Bowel Preparation Scale (BBPS) for colonoscopy assessment. Methods: We retrospectively collected 233 colonoscopy images from 2020 to 2023. These images were evaluated using the BBPS by 3 senior endoscopists and 3 novice endoscopists. Additionally, ChatGPT also assessed these images, having been divided into three groups and undergone specific Fine-tuning. Consistency was evaluated through two rounds of testing. Results: In the initial round, ChatGPT's accuracy varied between 48.93% and 62.66%, trailing the endoscopists' accuracy of 76.68% to 77.83%. Kappa values for ChatGPT was between 0.52 and 0.53, compared to 0.75 to 0.87 for the endoscopists. Conclusion: While ChatGPT shows promise in bowel preparation scoring, it currently does not match the accuracy and consistency of experienced endoscopists. Future research should focus on in-depth Fine-tuning.

Transformer-based models have dominated natural language processing and other areas in the last few years due to their superior (zero-shot) performance on benchmark datasets. However, these models are poorly understood due to their complexity and size. While probing-based methods are widely used to understand specific properties, the structures of the representation space are not systematically characterized; consequently, it is unclear how such models generalize and overgeneralize to new inputs beyond datasets. In this paper, based on a new gradient descent optimization method, we are able to explore the embedding space of a commonly used vision-language model. Using the Imagenette dataset, we show that while the model achieves over 99\% zero-shot classification performance, it fails systematic evaluations completely. Using a linear approximation, we provide a framework to explain the striking differences. We have also obtained similar results using a different model to support that our results are applicable to other transformer models with continuous inputs. We also propose a robust way to detect the modified images.

As Large Language Models (LLMs) become more proficient, their misuse in large-scale viral disinformation campaigns is a growing concern. This study explores the capability of ChatGPT to generate unconditioned claims about the war in Ukraine, an event beyond its knowledge cutoff, and evaluates whether such claims can be differentiated by human readers and automated tools from human-written ones. We compare war-related claims from ClaimReview, authored by IFCN-registered fact-checkers, and similar short-form content generated by ChatGPT. We demonstrate that ChatGPT can produce realistic, target-specific disinformation cheaply, fast, and at scale, and that these claims cannot be reliably distinguished by humans or existing automated tools.

Generative AI is changing the way developers interact with software systems, providing services that can produce and deliver new content, crafted to satisfy the actual needs of developers. For instance, developers can ask for new code directly from within their IDEs by writing natural language prompts, and integrated services based on generative AI, such as Copilot, immediately respond to prompts by providing ready-to-use code snippets. Formulating the prompt appropriately, and incorporating the useful information while avoiding any information overload, can be an important factor in obtaining the right piece of code. The task of designing good prompts is known as prompt engineering. In this paper, we systematically investigate the influence of eight prompt features on the style and the content of prompts, on the level of correctness, complexity, size, and similarity to the developers' code of the generated code. We specifically consider the task of using Copilot with 124,800 prompts obtained by systematically combining the eight considered prompt features to generate the implementation of 200 Java methods. Results show how some prompt features, such as the presence of examples and the summary of the purpose of the method, can significantly influence the quality of the result.