NeurIPS 2024 论文列表

The proliferation of open-source Large Language Models (LLMs) from various institutions has highlighted the urgent need for comprehensive evaluation methods. However, current evaluation platforms, such as the widely recognized HuggingFace open LLM leaderboard, neglect a crucial aspect -- uncertainty, which is vital for thoroughly assessing LLMs. To bridge this gap, we introduce a new benchmarking approach for LLMs that integrates uncertainty quantification. Our examination involves nine LLMs (LLM series) spanning five representative natural language processing tasks. Our findings reveal that: I) LLMs with higher accuracy may exhibit lower certainty; II) Larger-scale LLMs may display greater uncertainty compared to their smaller counterparts; and III) Instruction-finetuning tends to increase the uncertainty of LLMs. These results underscore the significance of incorporating uncertainty in the evaluation of LLMs. Our implementation is available at https://github.com/smartyfh/LLM-Uncertainty-Bench.

Current post-training quantization methods for LLMs compress the weights down to 4-bits, with moderate to low degradation in accuracy. However, further reducing the number of bits or accelerating the network while avoiding large accuracy drops, especially for smaller, sub 7B models, remains an actively researched and open problem. To address this, in this work, we introduce Quantization with Binary Bases (QBB), a new approach for low-bit quantization that effectively removes (nearly) all multiplications, reducing the implementation to summations. Our novel approach works by decomposing the original weights into a set of binary (1-bit) matrices using an iterative process. For a given layer, starting from a weight matrix, we first construct an initial approximation using an analytical solution, where each new binary matrix, paired with a scaling vector, approximates the residual error of the previous estimation. Secondly, using gradient descent and a progressive learning curriculum, we find the optimal set of binary matrices and scaling vectors that minimize the $\ell_2$ distance between the produced approximation and original weights. Thirdly, as previous steps are input agnostic, we holistically optimize the scaling vectors alone, calibrating them in student-teacher fashion, with the teacher providing both the data, by autoregressive generation starting from a random token, and the target logits. When evaluated across multiple LLM families, our approach matches and outperforms all prior works, setting a new state-of-the-art result using a summation-only based approach.

It is commonplace to produce application-specific models by fine-tuning large pre-trained models using a small bespoke dataset. The widespread availability of foundation model checkpoints on the web poses considerable risks, including the vulnerability to backdoor attacks. In this paper, we unveil a new vulnerability: the privacy backdoor attack. This black-box privacy attack aims to amplify the privacy leakage that arises when fine-tuning a model: when a victim fine-tunes a backdoored model, their training data will be leaked at a significantly higher rate than if they had fine-tuned a typical model. We conduct extensive experiments on various datasets and models, including both vision-language models (CLIP) and large language models, demonstrating the broad applicability and effectiveness of such an attack. Additionally, we carry out multiple ablation studies with different fine-tuning methods and inference strategies to thoroughly analyze this new threat. Our findings highlight a critical privacy concern within the machine learning community and call for a re-evaluation of safety protocols in the use of open-source pre-trained models.

Recent advances in Video Large Language Models (Video-LLMs) have demonstrated their great potential in general-purpose video understanding. To verify the significance of these models, a number of benchmarks have been proposed to diagnose their capabilities in different scenarios. However, existing benchmarks merely evaluate models through video-level question-answering, lacking fine-grained event-level assessment and task diversity. To fill this gap, we introduce E.T. Bench (Event-Level & Time-Sensitive Video Understanding Benchmark), a large-scale and high-quality benchmark for open-ended event-level video understanding. Categorized within a 3-level task taxonomy, E.T. Bench encompasses 7.3K samples under 12 tasks with 7K videos (251.4h total length) under 8 domains, providing comprehensive evaluations. We extensively evaluated 8 Image-LLMs and 12 Video-LLMs on our benchmark, and the results reveal that state-of-the-art models for coarse-level (video-level) understanding struggle to solve our fine-grained tasks, e.g., grounding event-of-interests within videos, largely due to the short video context length, improper time representations, and lack of multi-event training data. Focusing on these issues, we further propose a strong baseline model, E.T. Chat, together with an instruction-tuning dataset E.T. Instruct 164K tailored for fine-grained event-level understanding. Our simple but effective solution demonstrates superior performance in multiple scenarios.

Large Language Models (LLMs) often do not perform well on queries that require the aggregation of information across texts. To better evaluate this setting and facilitate modeling efforts, we introduce TACT - Text And Calculations through Tables, a dataset crafted to evaluate LLMs' reasoning and computational abilities using complex instructions. TACT contains challenging instructions that demand stitching information scattered across one or more texts, and performing complex integration on this information to generate the answer. We construct this dataset by leveraging an existing dataset of texts and their associated tables. For each such tables, we formulate new queries, and gather their respective answers. We demonstrate that all contemporary LLMs perform poorly on this dataset, achieving an accuracy below 38%. To pinpoint the difficulties and thoroughly dissect the problem, we analyze model performance across three components: table-generation, Pandas command-generation, and execution. Unexpectedly, we discover that each component presents substantial challenges for current LLMs. These insights lead us to propose a focused modeling framework, which we refer to as _IE as a tool_. Specifically, we propose to add "tools" for each of the above steps, and implement each such tool with few-shot prompting. This approach shows an improvement over existing prompting techniques, offering a promising direction for enhancing model capabilities in these tasks.

We present the Brain Treebank, a large-scale dataset of electrophysiological neural responses, recorded from intracranial probes while 10 subjects watched one or more Hollywood movies. Subjects watched on average 2.6 Hollywood movies, for an average viewing time of 4.3 hours, and a total of 43 hours. The audio track for each movie was transcribed with manual corrections. Word onsets were manually annotated on spectrograms of the audio track for each movie. Each transcript was automatically parsed and manually corrected into the universal dependencies (UD) formalism, assigning a part of speech to every word and a dependency parse to every sentence. In total, subjects heard over 38,000 sentences (223,000 words), while they had on average 168 electrodes implanted. This is the largest dataset of intracranial recordings featuring grounded naturalistic language, one of the largest English UD treebanks in general, and one of only a few UD treebanks aligned to multimodal features. We hope that this dataset serves as a bridge between linguistic concepts, perception, and their neural representations. To that end, we present an analysis of which electrodes are sensitive to language features while also mapping out a rough time course of language processing across these electrodes. The Brain Treebank is available at https://BrainTreebank.dev/

Most jailbreak papers claim the jailbreaks they propose are highly effective, often boasting near-100% attack success rates. However, it is perhaps more common than not for jailbreak developers to substantially exaggerate the effectiveness of their jailbreaks. We suggest this problem arises because jailbreak researchers lack a standard, high-quality benchmark for evaluating jailbreak performance, leaving researchers to create their own. To create a benchmark, researchers must choose a dataset of forbidden prompts to which a victim model will respond, along with an evaluation method that scores the harmfulness of the victim model’s responses. We show that existing benchmarks suffer from significant shortcomings and introduce the StrongREJECT benchmark to address these issues. StrongREJECT's dataset contains prompts that victim models must answer with specific, harmful information, while its automated evaluator measures the extent to which a response gives useful information to forbidden prompts. In doing so, the StrongREJECT evaluator achieves state-of-the-art agreement with human judgments of jailbreak effectiveness. Notably, we find that existing evaluation methods significantly overstate jailbreak effectiveness compared to human judgments and the StrongREJECT evaluator. We describe a surprising and novel phenomenon that explains this discrepancy: jailbreaks bypassing a victim model’s safety fine-tuning tend to reduce its capabilities. Together, our findings underscore the need for researchers to use a high-quality benchmark, such as StrongREJECT, when developing new jailbreak attacks. We release the StrongREJECT code and data at https://strong-reject.readthedocs.io/.

Recently, interpretable machine learning has re-explored concept bottleneck models (CBM). An advantage of this model class is the user's ability to intervene on predicted concept values, affecting the downstream output. In this work, we introduce a method to perform such concept-based interventions on *pretrained* neural networks, which are not interpretable by design, only given a small validation set with concept labels. Furthermore, we formalise the notion of *intervenability* as a measure of the effectiveness of concept-based interventions and leverage this definition to fine-tune black boxes. Empirically, we explore the intervenability of black-box classifiers on synthetic tabular and natural image benchmarks. We focus on backbone architectures of varying complexity, from simple, fully connected neural nets to Stable Diffusion. We demonstrate that the proposed fine-tuning improves intervention effectiveness and often yields better-calibrated predictions. To showcase the practical utility of our techniques, we apply them to deep chest X-ray classifiers and show that fine-tuned black boxes are more intervenable than CBMs. Lastly, we establish that our methods are still effective under vision-language-model-based concept annotations, alleviating the need for a human-annotated validation set.

Embodied AI agents require a fine-grained understanding of the physical world mediated through visual and language inputs. Such capabilities are difficult to learn solely from task-specific data. This has led to the emergence of pre-trained vision-language models as a tool for transferring representations learned from internet-scale data to downstream tasks and new domains. However, commonly used contrastively trained representations such as in CLIP have been shown to fail at enabling embodied agents to gain a sufficiently fine-grained scene understanding—a capability vital for control. To address this shortcoming, we consider representations from pre-trained text-to-image diffusion models, which are explicitly optimized to generate images from text prompts and as such, contain text-conditioned representations that reflect highly fine-grained visuo-spatial information. Using pre-trained text-to-image diffusion models, we construct Stable Control Representations which allow learning downstream control policies that generalize to complex, open-ended environments. We show that policies learned using Stable Control Representations are competitive with state-of-the-art representation learning approaches across a broad range of simulated control settings, encompassing challenging manipulation and navigation tasks. Most notably, we show that Stable Control Representations enable learning policies that exhibit state-of-the-art performance on OVMM, a difficult open-vocabulary navigation benchmark.

This paper introduces Chain-of-Sight, a vision-language bridge module that accelerates the pre-training of Multimodal Large Language Models (MLLMs). Our approach employs a sequence of visual resamplers that capture visual details at various spacial scales. This architecture not only leverages global and local visual contexts effectively, but also facilitates the flexible extension of visual tokens through a compound token scaling strategy, allowing up to a 16x increase in the token count post pre-training. Consequently, Chain-of-Sight requires significantly fewer visual tokens in the pre-training phase compared to the fine-tuning phase. This intentional reduction of visual tokens during pre-training notably accelerates the pre-training process, cutting down the wall-clock training time by $\sim$73\%. Empirical results on a series of vision-language benchmarks reveal that the pre-train acceleration through Chain-of-Sight is achieved without sacrificing performance, matching or surpassing the standard pipeline of utilizing all visual tokens throughout the entire training process. Further scaling up the number of visual tokens for pre-training leads to stronger performances, competitive to existing approaches in a series of benchmarks.

Large Vision-Language Models (LVLMs) are capable of handling diverse data types such as imaging, text, and physiological signals, and can be applied in various fields. In the medical field, LVLMs have a high potential to offer substantial assistance for diagnosis and treatment. Before that, it is crucial to develop benchmarks to evaluate LVLMs' effectiveness in various medical applications. Current benchmarks are often built upon specific academic literature, mainly focusing on a single domain, and lacking varying perceptual granularities. Thus, they face specific challenges, including limited clinical relevance, incomplete evaluations, and insufficient guidance for interactive LVLMs. To address these limitations, we developed the GMAI-MMBench, the most comprehensive general medical AI benchmark with well-categorized data structure and multi-perceptual granularity to date. It is constructed from 284 datasets across 38 medical image modalities, 18 clinical-related tasks, 18 departments, and 4 perceptual granularities in a Visual Question Answering (VQA) format. Additionally, we implemented a lexical tree structure that allows users to customize evaluation tasks, accommodating various assessment needs and substantially supporting medical AI research and applications. We evaluated 50 LVLMs, and the results show that even the advanced GPT-4o only achieves an accuracy of 53.96\%, indicating significant room for improvement. Moreover, we identified five key insufficiencies in current cutting-edge LVLMs that need to be addressed to advance the development of better medical applications. We believe that GMAI-MMBench will stimulate the community to build the next generation of LVLMs toward GMAI.

A fundamental objective in robot manipulation is to enable models to comprehend visual scenes and execute actions. Although existing Vision-Language-Action (VLA) models for robots can handle a range of basic tasks, they still face challenges in two areas: (1) insufficient reasoning ability to tackle complex tasks, and (2) high computational costs for VLA model fine-tuning and inference. The recently proposed state space model (SSM) known as Mamba demonstrates promising capabilities in non-trivial sequence modeling with linear inference complexity. Inspired by this, we introduce RoboMamba, an end-to-end robotic VLA model that leverages Mamba to deliver both robotic reasoning and action capabilities, while maintaining efficient fine-tuning and inference. Specifically, we first integrate the vision encoder with Mamba, aligning visual tokens with language embedding through co-training, empowering our model with visual common sense and robotic-related reasoning. To further equip RoboMamba with SE(3) pose prediction abilities, we explore an efficient fine-tuning strategy with a simple policy head. We find that once RoboMamba possesses sufficient reasoning capability, it can acquire manipulation skills with minimal fine-tuning parameters (0.1\% of the model) and time. In experiments, RoboMamba demonstrates outstanding reasoning capabilities on general and robotic evaluation benchmarks. Meanwhile, our model showcases impressive pose prediction results in both simulation and real-world experiments, achieving inference speeds 3 times faster than existing VLA models.

Recent approaches to vision-language tasks are built on the remarkable capabilities of large vision-language models (VLMs). These models excel in zero-shot and few-shot learning, enabling them to learn new tasks without parameter updates. However, their primary challenge lies in their design, which primarily accommodates 2D input, thus limiting their effectiveness for medical images, particularly radiological images like MRI and CT, which are typically 3D. To bridge the gap between state-of-the-art 2D VLMs and 3D medical image data, we developed an innovative, one-pass, unsupervised representative slice selection method called Vote-MI, which selects representative 2D slices from 3D medical imaging. To evaluate the effectiveness of vote-MI when implemented with VLMs, we introduce BrainMD, a robust, multimodal dataset comprising 2,453 annotated 3D MRI brain scans with corresponding textual radiology reports and electronic health records. Based on BrainMD, we further develop two benchmarks, BrainMD-select (including the most representative 2D slice of 3D image) and BrainBench (including various vision-language downstream tasks). Extensive experiments on the BrainMD dataset and its two corresponding benchmarks demonstrate that our representative selection method significantly improves performance in zero-shot and few-shot learning tasks. On average, Vote-MI achieves a 14.6\% and 16.6\% absolute gain for zero-shot and few-shot learning, respectively, compared to randomly selecting examples. Our studies represent a significant step toward integrating AI in medical imaging to enhance patient care and facilitate medical research. We hope this work will serve as a foundation for data selection as vision-language models are increasingly applied to new tasks.

Concept Bottleneck Models (CBMs) provide interpretable prediction by introducing an intermediate Concept Bottleneck Layer (CBL), which encodes human-understandable concepts to explain models' decision. Recent works proposed to utilize Large Language Models and pre-trained Vision-Language Models to automate the training of CBMs, making it more scalable and automated. However, existing approaches still fall short in two aspects: First, the concepts predicted by CBL often mismatch the input image, raising doubts about the faithfulness of interpretation. Second, it has been shown that concept values encode unintended information: even a set of random concepts could achieve comparable test accuracy to state-of-the-art CBMs. To address these critical limitations, in this work, we propose a novel framework called Vision-Language-Guided Concept Bottleneck Model (VLG-CBM) to enable faithful interpretability with the benefits of boosted performance. Our method leverages off-the-shelf open-domain grounded object detectors to provide visually grounded concept annotation, which largely enhances the faithfulness of concept prediction while further improving the model performance. In addition, we propose a new metric called Number of Effective Concepts (NEC) to control the information leakage and provide better interpretability. Extensive evaluations across five standard benchmarks show that our method, VLG-CBM, outperforms existing methods by at least 4.27\% and up to 51.09\% on *Accuracy at NEC=5* (denoted as ANEC-5), and by at least 0.45\% and up to 29.78\% on *average accuracy* (denoted as ANEC-avg), while preserving both faithfulness and interpretability of the learned concepts as demonstrated in extensive experiments.

Vision-Language Models (VLMs) excel in generating textual responses from visual inputs, but their versatility raises security concerns. This study takes the first step in exposing VLMs’ susceptibility to data poisoning attacks that can manipulate responses to innocuous, everyday prompts. We introduce Shadowcast, a stealthy data poisoning attack where poison samples are visually indistinguishable from benign images with matching texts. Shadowcast demonstrates effectiveness in two attack types. The first is a traditional Label Attack, tricking VLMs into misidentifying class labels, such as confusing Donald Trump for Joe Biden. The second is a novel Persuasion Attack, leveraging VLMs’ text generation capabilities to craft persuasive and seemingly rational narratives for misinformation, such as portraying junk food as healthy. We show that Shadowcast effectively achieves the attacker’s intentions using as few as 50 poison samples. Crucially, the poisoned samples demonstrate transferability across different VLM architectures, posing a significant concern in black-box settings. Moreover, Shadowcast remains potent under realistic conditions involving various text prompts, training data augmentation, and image compression techniques. This work reveals how poisoned VLMs can disseminate convincing yet deceptive misinformation to everyday, benign users, emphasizing the importance of data integrity for responsible VLM deployments. Our code is available at: https://github.com/umd-huang-lab/VLM-Poisoning.

Recently, Large Language Models (LLMs) make remarkable evolutions in language understanding and generation. Following this, various benchmarks for measuring all kinds of capabilities of LLMs have sprung up. In this paper, we challenge the reasoning and understanding abilities of LLMs by proposing a FaLlacy Understanding Benchmark (FLUB) containing cunning texts that are easy for humans to understand but difficult for models to grasp. Specifically, the cunning texts that FLUB focuses on mainly consist of the tricky, humorous, and misleading texts collected from the real internet environment. And we design three tasks with increasing difficulty in the FLUB benchmark to evaluate the fallacy understanding ability of LLMs. Based on FLUB, we investigate the performance of multiple representative and advanced LLMs, reflecting our FLUB is challenging and worthy of more future study. Interesting discoveries and valuable insights are achieved in our extensive experiments and detailed analyses. We hope that our benchmark can encourage the community to improve LLMs' ability to understand fallacies. Our data and codes are available at https://github.com/THUKElab/FLUB.

The rapid development of large language and vision models (LLVMs) has been driven by advances in visual instruction tuning. Recently, open-source LLVMs have curated high-quality visual instruction tuning datasets and utilized additional vision encoders or multiple computer vision models in order to narrow the performance gap with powerful closed-source LLVMs. These advancements are attributed to multifaceted information required for diverse capabilities, including fundamental image understanding, real-world knowledge about common-sense and non-object concepts (e.g., charts, diagrams, symbols, signs, and math problems), and step-by-step procedures for solving complex questions. Drawing from the multifaceted information, we present a new efficient LLVM, Mamba-based traversal of rationales (Meteor), which leverages multifaceted rationale to enhance understanding and answering capabilities. To embed lengthy rationales containing abundant information, we employ the Mamba architecture, capable of processing sequential data with linear time complexity. We introduce a new concept of traversal of rationale that facilitates efficient embedding of rationale. Subsequently, the backbone multimodal language model (MLM) is trained to generate answers with the aid of rationale. Through these steps, Meteor achieves significant improvements in vision language performances across multiple evaluation benchmarks requiring diverse capabilities, without scaling up the model size or employing additional vision encoders and computer vision models.

Existing Multimodal Large Language Models (MLLMs) follow the paradigm that perceives visual information by aligning visual features with the input space of Large Language Models (LLMs) and concatenating visual tokens with text tokens to form a unified sequence input for LLMs. These methods demonstrate promising results on various vision-language tasks but are limited by the high computational effort due to the extended input sequence resulting from the involvement of visual tokens. In this paper, instead of input space alignment, we propose a novel parameter space alignment paradigm that represents visual information as model weights. For each input image, we use a vision encoder to extract visual features, convert features into perceptual weights, and merge the perceptual weights with LLM's weights. In this way, the input of LLM does not require visual tokens, which reduces the length of the input sequence and greatly improves efficiency. Following this paradigm, we propose VLoRA with the perceptual weights generator. The perceptual weights generator is designed to convert visual features to perceptual weights with low-rank property, exhibiting a form similar to LoRA. The experimental results show that our VLoRA achieves comparable performance on various benchmarks for MLLMs, while significantly reducing the computational costs for both training and inference. Code and models are released at \url{https://github.com/FeipengMa6/VLoRA}.

Multimodal foundation models serve numerous applications at the intersection of vision and language. Still, despite being pretrained on extensive data, they become outdated over time. To keep models updated, research into continual pretraining mainly explores scenarios with either (1) infrequent, indiscriminate updates on large-scale new data, or (2) frequent, sample-level updates. However, practical model deployment often operates in the gap between these two limit cases, as real-world applications demand adaptation to specific subdomains, tasks or concepts --- spread over the entire, varying life cycle of a model. In this work, we complement current perspectives on continual pretraining through a research test bed and offer comprehensive guidance for effective continual model updates in such scenarios. We first introduce FoMo-in-Flux, a continual multimodal pretraining benchmark with realistic compute constraints and practical deployment requirements, constructed over 63 datasets with diverse visual and semantic coverage. Using FoMo-in-Flux, we explore the complex landscape of practical continual pretraining through multiple perspectives: (1) data mixtures and stream orderings that emulate real-world deployment settings, (2) methods ranging from simple fine-tuning and traditional continual learning strategies to parameter-efficient updates and model merging, (3) meta-learning-rate schedules and mechanistic design choices, and (4) model and compute scaling. Together, our insights provide a practitioner's guide to continual multimodal pretraining for real-world deployment. Benchmark and code is provided here: https://github.com/ExplainableML/fomo_in_flux.