ICML 2024 Papers

Large Language Models are prone to biased predictions and hallucinations, underlining the paramount importance of understanding their model-internal reasoning process. However, achieving faithful attributions for the entirety of a black-box transformer model and maintaining computational efficiency is an unsolved challenge. By extending the Layer-wise Relevance Propagation attribution method to handle attention layers, we address these challenges effectively. While partial solutions exist, our method is the first to faithfully and holistically attribute not only input but also latent representations of transformer models with the computational efficiency similar to a single backward pass. Through extensive evaluations against existing methods on LLaMa 2, Mixtral 8x7b, Flan-T5 and vision transformer architectures, we demonstrate that our proposed approach surpasses alternative methods in terms of faithfulness and enables the understanding of latent representations, opening up the door for concept-based explanations. We provide an LRP library at https://github.com/rachtibat/LRP-eXplains-Transformers.

Harnessing transformer-based models, visual tracking has made substantial strides. However, the sluggish performance of current trackers limits their practicality on devices with constrained computational capabilities, especially for real-time unmanned aerial vehicle (UAV) tracking. Addressing this challenge, we introduce AVTrack, an adaptive computation framework tailored to selectively activate transformer blocks for real-time UAV tracking in this work. Our novel Activation Module (AM) dynamically optimizes ViT architecture, selectively engaging relevant components and enhancing inference efficiency without compromising much tracking performance. Moreover, we bolster the effectiveness of ViTs, particularly in addressing challenges arising from extreme changes in viewing angles commonly encountered in UAV tracking, by learning view-invariant representations through mutual information maximization. Extensive experiments on five tracking benchmarks affirm the effectiveness and versatility of our approach, positioning it as a state-of-the-art solution in visual tracking. Code is released at: https://github.com/wuyou3474/AVTrack.

The Partial Credit Model (PCM) of Andrich (1978) and Masters (1982) is a fundamental model within the psychometric literature with wide-ranging modern applications. It models the integer-valued response that a subject gives to an item where there is a natural notion of monotonic progress between consecutive response values, such as partial scores on a test and customer ratings of a product. In this paper, we introduce a novel, time-efficient and accurate statistical spectral algorithm for inference under the PCM model. We complement our algorithmic contribution with in-depth non-asymptotic statistical analysis, the first of its kind in the literature. We show that the spectral algorithm enjoys the optimal error guarantee under three different metrics, all under reasonable sampling assumptions. We leverage the efficiency of the spectral algorithm to propose a novel EM-based algorithm for learning mixtures of PCMs. We perform comprehensive experiments on synthetic and real-life datasets covering education testing, recommendation systems, and financial investment applications. We show that the proposed spectral algorithm is competitive with previously introduced algorithms in terms of accuracy while being orders of magnitude faster.

To leverage LLMs for visual synthesis, traditional methods convert raster image information into discrete grid tokens through specialized visual modules, while disrupting the model’s ability to capture the true semantic representation of visual scenes. This paper posits that an alternative representation of images, vector graphics, can effectively surmount this limitation by enabling a more natural and semantically coherent segmentation of the image information. Thus, we introduce StrokeNUWA, a pioneering work exploring a better visual representation "stroke" tokens on vector graphics, which is inherently visual semantics rich, naturally compatible with LLMs, and highly compressed. Equipped with stroke tokens, StrokeNUWA can significantly surpass traditional LLM-based and optimization-based methods across various metrics in the vector graphic generation task. Besides, StrokeNUWA achieves up to a $94\times$ speedup in inference over the speed of prior methods with an exceptional SVG code compression ratio of 6.9%.

Planning is a fundamental property of human intelligence. Reasoning about asynchronous plans is challenging since it requires sequential and parallel planning to optimize time costs. Can large language models (LLMs) succeed at this task? Here, we present the first large-scale study investigating this question. We find that a representative set of closed and open-source LLMs, including GPT-4 and LLaMA-2, behave poorly when not supplied with illustrations about the task-solving process in our benchmark AsyncHow. We propose a novel technique called *Plan Like a Graph* (PLaG) that combines graphs with natural language prompts and achieves state-of-the-art results. We show that although PLaG can boost model performance, LLMs still suffer from drastic degradation when task complexity increases, highlighting the limits of utilizing LLMs for simulating digital devices. We see our study as an exciting step towards using LLMs as efficient autonomous agents. Our code and data are available at https://github.com/fangru-lin/graph-llm-asynchow-plan.

Class-incremental learning (CIL) aims to train a model to learn new classes from non-stationary data streams without forgetting old ones. In this paper, we propose a new kind of connectionist model by tailoring neural unit dynamics that adapt the behavior of neural networks for CIL. In each training session, it introduces a supervisory mechanism to guide network expansion whose growth size is compactly commensurate with the intrinsic complexity of a newly arriving task. This constructs a near-minimal network while allowing the model to expand its capacity when cannot sufficiently hold new classes. At inference time, it automatically reactivates the required neural units to retrieve knowledge and leaves the remaining inactivated to prevent interference. We name our model AutoActivator, which is effective and scalable. To gain insights into the neural unit dynamics, we theoretically analyze the model’s convergence property via a universal approximation theorem on learning sequential mappings, which is under-explored in the CIL community. Experiments show that our method achieves strong CIL performance in rehearsal-free and minimal-expansion settings with different backbones.

We propose Embodied AI (E-AI) as the next fundamental step in the pursuit of Artificial General Intelligence (AGI), juxtaposing it against current AI advancements, particularly Large Language Models (LLMs). We traverse the evolution of the embodiment concept across diverse fields (philosophy, psychology, neuroscience, and robotics) to highlight how E-AI distinguishes itself from the classical paradigm of static learning. By broadening the scope of E-AI, we introduce a theoretical framework based on cognitive architectures, emphasizing perception, action, memory, and learning as essential components of an embodied agent. This framework is aligned with Friston’s active inference principle, offering a comprehensive approach to E-AI development. Despite the progress made in the field of AI, substantial challenges, such as the formulation of a novel AI learning theory and the innovation of advanced hardware, persist. Our discussion lays down a foundational guideline for future E-AI research. Highlighting the importance of creating E-AI agents capable of seamless communication, collaboration, and coexistence with humans and other intelligent entities within real-world environments, we aim to steer the AI community towards addressing the multifaceted challenges and seizing the opportunities that lie ahead in the quest for AGI.

Large Language Models (LLMs) demonstrate remarkable proficiency in comprehending and handling text-based tasks. Many efforts are being made to transfer these attributes to video modality, which are termed Video-LLMs. However, existing Video-LLMs can only capture the coarse-grained semantics and are unable to effectively handle tasks related to comprehension or localization of specific video segments. In light of these challenges, we propose Momentor, a Video-LLM capable of accomplishing fine-grained temporal understanding tasks. To support the training of Momentor, we design an automatic data generation engine to construct Moment-10M, a large-scale video instruction dataset with segment-level instruction data. We train Momentor on Moment-10M, enabling it to perform segment-level reasoning and localization. Zero-shot evaluations on several tasks demonstrate that Momentor excels in fine-grained temporally grounded comprehension and localization.

Compressing high-capability Large Language Models (LLMs) has emerged as a favored strategy for resource-efficient inferences. While state-of-the-art (SoTA) compression methods boast impressive advancements in preserving benign task performance, the potential risks of compression in terms of safety and trustworthiness have been largely neglected. This study conducts the first, thorough evaluation of **three (3) leading LLMs** using **five (5) SoTA compression techniques** across **eight (8) trustworthiness dimensions**. Our experiments highlight the intricate interplay between compression and trustworthiness, revealing some interesting patterns. We find that quantization is currently a more effective approach than pruning in achieving efficiency and trustworthiness simultaneously. For instance, a 4-bit quantized model retains the trustworthiness of its original counterpart, but model pruning significantly degrades trustworthiness, even at 50% sparsity. Moreover, employing quantization within a moderate bit range could unexpectedly improve certain trustworthiness dimensions such as ethics and fairness. Conversely, extreme quantization to very low bit levels (3 bits) tends to reduce trustworthiness significantly. This increased risk cannot be uncovered by looking at benign performance alone, in turn, mandating comprehensive trustworthiness evaluation in practice. These findings culminate in practical recommendations for simultaneously achieving high utility, efficiency, and trustworthiness in LLMs. Code and models are available at https://decoding-comp-trust.github.io.

Deep equilibrium (DEQ) models are widely recognized as a memory efficient alternative to standard neural networks, achieving state-of-the-art performance in language modeling and computer vision tasks. These models solve a fixed point equation instead of explicitly computing the output, which sets them apart from standard neural networks. However, existing DEQ models often lack formal guarantees of the existence and uniqueness of the fixed point, and the convergence of the numerical scheme used for computing the fixed point is not formally established. As a result, DEQ models are potentially unstable in practice. To address these drawbacks, we introduce a novel class of DEQ models called positive concave deep equilibrium (pcDEQ) models. Our approach, which is based on nonlinear Perron-Frobenius theory, enforces nonnegative weights and activation functions that are concave on the positive orthant. By imposing these constraints, we can easily ensure the existence and uniqueness of the fixed point without relying on additional complex assumptions commonly found in the DEQ literature, such as those based on monotone operator theory in convex analysis. Furthermore, the fixed point can be computed with the standard fixed point algorithm, and we provide theoretical guarantees of its geometric convergence, which, in particular, simplifies the training process. Experiments demonstrate the competitiveness of our pcDEQ models against other implicit models.

Abstract visual reasoning is a characteristically human ability, allowing the identification of relational patterns that are abstracted away from object features, and the systematic generalization of those patterns to unseen problems. Recent work has demonstrated strong systematic generalization in visual reasoning tasks involving multi-object inputs, through the integration of slot-based methods used for extracting object-centric representations coupled with strong inductive biases for relational abstraction. However, this approach was limited to problems containing a single rule, and was not scalable to visual reasoning problems containing a large number of objects. Other recent work proposed Abstractors, an extension of Transformers that incorporates strong relational inductive biases, thereby inheriting the Transformer's scalability and multi-head architecture, but it has yet to be demonstrated how this approach might be applied to multi-object visual inputs. Here we combine the strengths of the above approaches and propose Slot Abstractors, an approach to abstract visual reasoning that can be scaled to problems involving a large number of objects and multiple relations among them. The approach displays state-of-the-art performance across four abstract visual reasoning tasks, as well as an abstract reasoning task involving real-world images.

Post-training quantization of Large Language Models (LLMs) is challenging. In this work, we introduce **L**ow-rank **Q**uantization **E**rror **R**eduction (LQER), which combines quantization and low-rank approximation to recover the model capability. LQER leverages an activation-induced scale matrix to drive the singular value distribution of quantization error towards a desirable distribution, which enables nearly-lossless W4A8 quantization on various LLMs and downstream tasks without the need for knowledge distillation, grid search, or gradient-based iterative optimization. Unlike existing methods, the computation pattern of LQER eliminates the need for specialized Scatter and Gather processes to collect high-precision weights from irregular memory locations. Our W4A8 LLMs achieve near-lossless performance on six popular downstream tasks, while using $1.36 \times$ fewer hardware resources than the leading state-of-the-art method. We will open-source our framework at [https://github.com/ChengZhang-98/lqer](https://github.com/ChengZhang-98/lqer)

Multimodal Large Language Models (MLLMs) have gained significant attention recently, showing remarkable potential in artificial general intelligence. However, assessing the utility of MLLMs presents considerable challenges, primarily due to the absence multimodal benchmarks that align with human preferences. Drawing inspiration from the concept of LLM-as-a-Judge within LLMs, this paper introduces a novel benchmark, termed MLLM-as-a-Judge, to assess the ability of MLLMs in assisting judges across diverse modalities, encompassing three distinct tasks: Scoring Evaluation, Pair Comparison, and Batch Ranking. Our study reveals that, while MLLMs demonstrate remarkable human-like discernment in Pair Comparisons, there is a significant divergence from human preferences in Scoring Evaluation and Batch Ranking tasks. Furthermore, a closer examination reveals persistent challenges in the evaluative capacities of LLMs, including diverse biases, hallucinatory responses, and inconsistencies in judgment, even in advanced models such as GPT-4V. These findings emphasize the pressing need for enhancements and further research efforts to be undertaken before regarding MLLMs as fully reliable evaluators. In light of this, we advocate for additional efforts dedicated to supporting the continuous development within the domain of MLLM functioning as judges. The code and dataset are publicly available at our project homepage: https://mllm-judge.github.io/.

Adversarial training has achieved substantial performance in defending image retrieval against adversarial examples. However, existing studies in deep metric learning (DML) still suffer from two major limitations: *weak adversary* and *model collapse*. In this paper, we address these two limitations by proposing **C**ollapse-**A**ware **TRI**plet **DE**coupling (**CA-TRIDE**). Specifically, TRIDE yields a stronger adversary by spatially decoupling the perturbation targets into the anchor and the other candidates. Furthermore, CA prevents the consequential model collapse, based on a novel metric, collapseness, which is incorporated into the optimization of perturbation. We also identify two drawbacks of the existing robustness metric in image retrieval and propose a new metric for a more reasonable robustness evaluation. Extensive experiments on three datasets demonstrate that CA-TRIDE outperforms existing defense methods in both conventional and new metrics. *Codes are available at https://github.com/michaeltian108/CA-TRIDE.*

This paper introduces AIM, a collection of vision models pre-trained with an autoregressive objective. These models are inspired by their textual counterparts, i.e., Large Language Models (LLMs), and exhibit similar scaling properties. Specifically, we highlight two key findings: (1) the performance of the visual features scale with both the model capacity and the quantity of data, (2) the value of the objective function correlates with the performance of the model on downstream tasks. We illustrate the practical implication of these findings by pre-training a 7 billion parameter AIM on 2 billion images, that achieves 84.0% on ImageNet-1k with a frozen trunk. Interestingly, even at this scale, we observe no sign of saturation in performance, suggesting that AIM potentially represents a new frontier for training large-scale vision models. The pre-training of AIM is similar to the pre-training of LLMs, and does not require any image-specific strategy to stabilize the training at scale.

Watermarking techniques offer a promising way to identify machine-generated content via embedding covert information into the contents generated from language models. A challenge in the domain lies in preserving the distribution of original generated content after watermarking. Our research extends and improves upon existing watermarking framework, placing emphasis on the importance of a Distribution-Preserving (DiP) watermark. Contrary to the current strategies, our proposed DiPmark simultaneously preserves the original token distribution during watermarking (distribution-preserving), is detectable without access to the language model API and prompts (accessible), and is provably robust to moderate changes of tokens (resilient). DiPmark operates by selecting a random set of tokens prior to the generation of a word, then modifying the token distribution through a distribution-preserving reweight function to enhance the probability of these selected tokens during the sampling process. Extensive empirical evaluation on various language models and tasks demonstrates our approach's distribution-preserving property, accessibility, and resilience, making it a effective solution for watermarking tasks that demand impeccable quality preservation.

Uncertainty decomposition refers to the task of decomposing the total uncertainty of a predictive model into aleatoric (data) uncertainty, resulting from inherent randomness in the data-generating process, and epistemic (model) uncertainty, resulting from missing information in the model's training data. In large language models (LLMs) specifically, identifying sources of uncertainty is an important step toward improving reliability, trustworthiness, and interpretability, but remains an important open research question. In this paper, we introduce an uncertainty decomposition framework for LLMs, called input clarification ensembling, which can be applied to any pre-trained LLM. Our approach generates a set of clarifications for the input, feeds them into an LLM, and ensembles the corresponding predictions. We show that, when aleatoric uncertainty arises from ambiguity or under-specification in LLM inputs, this approach makes it possible to factor an (un-clarified) LLM's predictions into separate aleatoric and epistemic terms, using a decomposition similar to the one employed by Bayesian neural networks. Empirical evaluations demonstrate that input clarification ensembling provides accurate and reliable uncertainty quantification on several language processing tasks. Code and data are available at https://github.com/UCSB-NLP-Chang/llm_uncertainty.

Autoformalization involves automatically translating informal math into formal theorems and proofs that are machine-verifiable. Euclidean geometry provides an interesting and controllable domain for studying autoformalization. In this paper, we introduce a neuro-symbolic framework for autoformalizing Euclidean geometry, which combines domain knowledge, SMT solvers, and large language models (LLMs). One challenge in Euclidean geometry is that informal proofs rely on diagrams, leaving gaps in texts that are hard to formalize. To address this issue, we use theorem provers to fill in such diagrammatic information automatically, so that the LLM only needs to autoformalize the explicit textual steps, making it easier for the model. We also provide automatic semantic evaluation for autoformalized theorem statements. We construct LeanEuclid, an autoformalization benchmark consisting of problems from Euclid's Elements and the UniGeo dataset formalized in the Lean proof assistant. Experiments with GPT-4 and GPT-4V show the capability and limitations of state-of-the-art LLMs on autoformalizing geometry problems. The data and code are available at https://github.com/loganrjmurphy/LeanEuclid.

We present an approach for estimating the fraction of text in a large corpus which is likely to be substantially modified or produced by a large language model (LLM). Our maximum likelihood model leverages expert-written and AI-generated reference texts to accurately and efficiently examine real-world LLM-use at the corpus level. We apply this approach to a case study of scientific peer review in AI conferences that took place after the release of ChatGPT: *ICLR* 2024, *NeurIPS* 2023, *CoRL* 2023 and *EMNLP* 2023. Our results suggest that between 6.5% and 16.9% of text submitted as peer reviews to these conferences could have been substantially modified by LLMs, i.e. beyond spell-checking or minor writing updates. The circumstances in which generated text occurs offer insight into user behavior: the estimated fraction of LLM-generated text is higher in reviews which report lower confidence, were submitted close to the deadline, and from reviewers who are less likely to respond to author rebuttals. We also observe corpus-level trends in generated text which may be too subtle to detect at the individual level, and discuss the implications of such trends on peer review. We call for future interdisciplinary work to examine how LLM use is changing our information and knowledge practices.