ICLR 2025 论文列表

As large language models (LLMs) grow in sophistication, some of their capabilities surpass human abilities, making it essential to ensure their alignment with human values and intentions, i.e., Superalignment. This superalignment challenge is particularly critical for complex tasks, as annotations provided by humans, as weak supervisors, may be overly simplistic, incomplete, or incorrect. Previous work has demonstrated the potential of training a strong model using the weak dataset generated by a weak model as weak supervision. However, these studies have been limited to a single capability. In this work, we conduct extensive experiments to investigate weak to strong generalization for LLMs with multi-capabilities. The experiments reveal that different capabilities tend to remain relatively independent in this generalization, and the effectiveness of weak supervision is significantly impacted by the quality and diversity of the weak datasets. Moreover, the self-bootstrapping of the strong model leads to performance degradation due to its overconfidence and the limited diversity of its generated dataset. To address these issues, we proposed a novel training framework using reward models to select valuable data, thereby providing weak supervision for strong model training. In addition, we propose a two-stage training method on both weak and selected datasets to train the strong model. Experimental results demonstrate our method significantly improves the weak to strong generalization with multi-capabilities.

We consider multi-draft speculative sampling, where the proposal sequences are sampled independently from different draft models. At each step, a token-level draft selection scheme takes a list of valid tokens as input and produces an output token whose distribution matches that of the target model. Previous works have demonstrated that the optimal scheme (which maximizes the probability of accepting one of the input tokens) can be cast as a solution to a linear program. In this work we show that the optimal scheme can be decomposed into a two-step solution: in the first step an importance sampling (IS) type scheme is used to select one intermediate token; in the second step (single-draft) speculative sampling is applied to generate the output token. For the case of two identical draft models we further 1) establish a necessary and sufficient condition on the distributions of the target and draft models for the acceptance probability to equal one and 2) provide an explicit expression for the optimal acceptance probability. Our theoretical analysis also motives a new class of token-level selection schemes based on weighted importance sampling. Our experimental results demonstrate consistent improvements in the achievable block efficiency and token rates over baseline schemes in a number of scenarios.

In this paper, we identify a common yet challenging issue in image-text matching, i.e., clone negatives: negative image-text pairs that semantically resemble positive pairs, leading to ambiguous and sub-optimal matching outcomes. To tackle this issue, we propose Adaptive Contrastive Learning (AdaCL), which introduces two margin parameters along with a modulating anchor to dynamically strengthen the compactness between positives and mitigate the influence of clone negatives. The modulating anchor is selected based on the distribution of negative samples without the need for explicit training, allowing for progressive tuning and advanced in-batch supervision. Extensive experiments across several tasks demonstrate the effectiveness of AdaCL in image-text matching. Furthermore, we extend AdaCL to weakly-supervised image-text matching by replacing human-annotated descriptions with automatically generated captions, thereby increasing the number of potential clone negatives. AdaCL maintains robustness in this setting, alleviating the reliance on crowd-sourced annotations and laying a foundation for scalable vision-language contrastive learning.

In this work, we present ALLaM: Arabic Large Language Model, a series of large language models to support the ecosystem of Arabic Language Technologies (ALT). ALLaM is carefully trained, considering the values of language alignment and transferability of knowledge at scale. The models are based on an autoregressive decoder-only architecture and are pretrained on a mixture of Arabic and English texts. We illustrate how the second-language acquisition via vocabulary expansion can help steer a language model towards a new language without any major catastrophic forgetting in English. Furthermore, we highlight the effectiveness of using translation data and the process of knowledge encoding within the language model's latent space. Finally, we show that effective alignment with human preferences can significantly enhance the performance of a large language model (LLM) compared to less aligned models of a larger scale. Our methodology enables us to achieve state-of-the-art performance in various Arabic benchmarks, including MMLU Arabic, ACVA, and Arabic Exams. Our aligned models improve both in Arabic and English from its base aligned models.

Multi-modal Large Language Models (MLLMs) have recently showcased superior proficiency in general visual scenarios. However, we identify their mathematical capabilities remain under-explored with three areas to be improved: visual encoding of math diagrams, diagram-language alignment, and chain-of-thought (CoT) reasoning. This draws forth an urgent demand for an effective training paradigm and a large-scale, comprehensive dataset with detailed CoT rationales, which is challenging to collect and costly to annotate manually. To tackle this issue, we propose MAVIS, a MAthematical VISual instruction tuning pipeline for MLLMs, featuring an automatic data engine to efficiently create mathematical visual datasets. We design the data generation process to be entirely independent of human intervention or GPT API usage, while ensuring the diagram-caption correspondence, question-answer correctness, and CoT reasoning quality. With this approach, we curate two datasets, MAVIS-Caption (558K diagram-caption pairs) and MAVIS-Instruct (834K visual math problems with CoT rationales), and propose four progressive stages for training MLLMs from scratch. First, we utilize MAVIS-Caption to fine-tune a math-specific vision encoder (CLIP-Math) through contrastive learning, tailored for improved diagram visual encoding. Second, we also leverage MAVIS-Caption to align the CLIP-Math with a large language model (LLM) by a projection layer, enhancing vision-language alignment in mathematical domains. Third, we adopt MAVIS-Instruct to perform the instruction tuning for robust problem-solving skills, and term the resulting model as MAVIS-7B. Fourth, we apply Direct Preference Optimization (DPO) to enhance the CoT capabilities of our model, further refining its step-wise reasoning performance. On various mathematical benchmarks, our MAVIS-7B achieves leading results among open-source MLLMs, e.g., surpassing other 7B models by +9.3% and the second-best LLaVA-NeXT (110B) by +6.9%, demonstrating the effectiveness of our method.

Low-light image enhancement poses a significant challenge due to the limited information captured by image sensors in low-light environments. Despite recent improvements in deep learning models, the lack of paired training datasets remains a significant obstacle. Therefore, unsupervised methods have emerged as a promising solution. In this work, we focus on the strength of curve-adjustment-based approaches to tackle unsupervised methods. The majority of existing unsupervised curve-adjustment approaches iteratively estimate higher order curve parameters to enhance the exposure of images while efficiently preserving the details of the images. However, the convergence of the enhancement procedure cannot be guaranteed, leading to sensitivity to the number of iterations and limited performance. To address this problem, we consider the iterative curve-adjustment update process as a dynamic system and formulate it as a Neural Ordinary Differential Equations (NODE) for the first time, and this allows us to learn a continuous dynamics of the latent image. The strategy of utilizing NODE to leverage continuous dynamics in iterative methods enhances unsupervised learning and aids in achieving better convergence compared to discrete-space approaches. Consequently, we achieve state-of-the-art performance in unsupervised low-light image enhancement across various benchmark datasets.

Large Language Models (LLM) are increasingly trained on data generated by other LLMs, either because generated text and images become part of the pre-training corpus, or because synthetized data is used as a replacement for expensive human-annotation. This raises concerns about *model collapse*, a drop in model performance when their training sets include generated data. Considering that it is easier for both humans and machines to tell between good and bad examples than to generate high-quality samples, we investigate the use of verification on synthesized data to prevent model collapse. We provide a theoretical characterization using Gaussian mixtures, linear classifiers, and linear verifiers to derive conditions with measurable proxies to assess whether the verifier can effectively select synthesized data that leads to optimal performance. We experiment with two practical tasks -- computing matrix eigenvalues with transformers and news summarization with LLMs -- which both exhibit model collapse when trained on generated data, and show that verifiers, even imperfect ones, can indeed be harnessed to prevent model collapse and that our proposed proxy measure strongly correlates with performance.

Language-based object detection (LOD) aims to align visual objects with language expressions. A large amount of paired data is utilized to improve LOD model generalizations. During the training process, recent studies leverage vision-language models (VLMs) to automatically generate human-like expressions for visual objects, facilitating training data scaling up. In this process, we observe that VLM hallucinations bring inaccurate object descriptions (e.g., object name, color, and shape) to deteriorate VL alignment quality. To reduce VLM hallucinations, we propose an agentic workflow controlled by an LLM to re-align language to visual objects via adaptively adjusting image and text prompts. We name this workflow Real-LOD, which includes planning, tool use, and reflection steps. Given an image with detected objects and VLM raw language expressions, Real-LOD reasons its state automatically and arranges action based on our neural symbolic designs (i.e., planning). The action will adaptively adjust the image and text prompts and send them to VLMs for object re-description (i.e., tool use). Then, we use another LLM to analyze these refined expressions for feedback (i.e., reflection). These steps are conducted in a cyclic form to gradually improve language descriptions for re-aligning to visual objects. We construct a dataset that contains a tiny amount of 0.18M images with re-aligned language expression and train a prevalent LOD model to surpass existing LOD methods by around 50% on the standard benchmarks. Our Real-LOD workflow, with automatic VL refinement, reveals a potential to preserve data quality along with scaling up data quantity, which further improves LOD performance from a data-alignment perspective.

We introduce WildBench, an automated evaluation framework designed to benchmark large language models (LLMs) using challenging, real-world user queries. WildBench consists of 1,024 tasks carefully selected from over one million human-chatbot conversation logs. For automated evaluation with WildBench, we have developed two metrics, WB-Reward and WB-Score, which are computable using advanced LLMs such as GPT-4-turbo. WildBench evaluation uses task-specific checklists to evaluate model outputs systematically and provides structured explanations that justify the scores and comparisons, resulting in more reliable and interpretable automatic judgments. WB-Reward employs fine-grained pairwise comparisons between model responses, generating five potential outcomes: much better, slightly better, slightly worse, much worse, or a tie. Unlike previous evaluations that employed a single baseline model, we selected three baseline models at varying performance levels to ensure a comprehensive pairwise evaluation. Additionally, we propose a simple method to mitigate length bias, by converting outcomes of “slightly better/worse” to “tie” if the winner response exceeds the loser one by more than K characters. WB-Score evaluates the quality of model outputs individually, making it a fast and cost-efficient evaluation metric. WildBench results demonstrate a strong correlation with the human-voted Elo ratings from Chatbot Arena on hard tasks. Specifically, WB-Reward achieves a Pearson correlation of 0.98 with top-ranking models. Additionally, WB-Score reaches 0.95, surpassing both ArenaHard’s 0.91 and AlpacaEval2.0’s 0.89 for length-controlled win rates, as well as the 0.87 for regular win rates.

As a core component of Transformer-like deep architectures, a feed-forward network (FFN) for channel mixing is responsible for learning features of each token. Recent works show channel mixing can be enhanced by increasing computational burden or can be slimmed at the sacrifice of performance. Although some efforts have been made, existing works are still struggling to solve the paradox of performance and complexity trade-offs. In this paper, we propose an Asymmetric Factorized Bilinear Operation (AFBO) to replace FFN of vision transformer (ViT), which attempts to efficiently explore rich statistics of token features for achieving better performance and complexity trade-off. Specifically, our AFBO computes second-order statistics via a spatial-channel factorized bilinear operation for feature learning, which replaces a simple linear projection in FFN and enhances the feature learning ability of ViT by modeling second-order correlation among token features. Furthermore, our AFBO presents two structured-sparsity channel mapping strategies, namely Grouped Cross Channel Mapping (GCCM) and Overlapped Cycle Channel Mapping (OCCM). They decompose bilinear operation into grouped channel features by considering information interaction between groups, significantly reducing computational complexity while guaranteeing model performance. Finally, our AFBO is built with GCCM and OCCM in an asymmetric way, aiming to achieve a better trade-off. Note that our AFBO is model-agnostic, which can be flexibly integrated with existing ViTs. Experiments are conducted with twenty ViTs on various tasks, and the results show our AFBO is superior to its counterparts while improving existing ViTs in terms of generalization and robustness.

Recent advancements in Multimodal Large Language Models (MLLMs) have accelerated the development of Graphical User Interface (GUI) agents capable of automating complex tasks across digital platforms. However, precise GUI element grounding remains a key challenge for accurate interaction and generalization. In this work, we present an effective GUI grounding framework, which includes an automated data collection engine that gathers extensive GUI screenshots and annotations to ensure broad generalization. We also propose a lightweight and flexible GUI grounding module designed to efficiently localize UI elements by pre-training on the collected data, and introduce a novel method to integrate this module with MLLMs for the effective execution of GUI tasks. Our approach demonstrates superior performance in task accuracy and adaptability, as validated by benchmarks such as ScreenSpot, MiniWob, AITW, and Mind2Web.

Egocentric video-language pretraining is a crucial step in advancing the understanding of hand-object interactions in first-person scenarios. Despite successes on existing testbeds, we find that current EgoVLMs can be easily misled by simple modifications, such as changing the verbs or nouns in interaction descriptions, with models struggling to distinguish between these changes. This raises the question: "Do EgoVLMs truly understand hand-object interactions?'' To address this question, we introduce a benchmark called $\textbf{EgoHOIBench}$, revealing the performance limitation of current egocentric models when confronted with such challenges. We attribute this performance gap to insufficient fine-grained supervision and the greater difficulty EgoVLMs experience in recognizing verbs compared to nouns. To tackle these issues, we propose a novel asymmetric contrastive objective named $\textbf{EgoNCE++}$. For the video-to-text objective, we enhance text supervision by generating negative captions using large language models or leveraging pretrained vocabulary for HOI-related word substitutions. For the text-to-video objective, we focus on preserving an object-centric feature space that clusters video representations based on shared nouns. Extensive experiments demonstrate that EgoNCE++ significantly enhances EgoHOI understanding, leading to improved performance across various EgoVLMs in tasks such as multi-instance retrieval, action recognition, and temporal understanding. Our code is available at https://github.com/xuboshen/EgoNCEpp.

Visual Language Models (VLMs) like GPT-4V have broadened the scope of LLM applications, yet they face significant challenges in accurately processing visual details, particularly in scientific diagrams. This paper explores the necessity of meticulous visual detail collection and region decomposition for enhancing the performance of VLMs in scientific diagram analysis. We propose a novel approach that combines traditional computer vision techniques with VLMs to systematically decompose diagrams into discernible visual elements and aggregate essential metadata. Our method employs techniques in OpenCV library to identify and label regions, followed by a refinement process using shape detection and region merging algorithms, which are particularly suited to the structured nature of scientific diagrams. This strategy not only improves the granularity and accuracy of visual information processing but also extends the capabilities of VLMs beyond their current limitations. We validate our approach through a series of experiments that demonstrate enhanced performance in diagram analysis tasks, setting a new standard for integrating visual and language processing in a multimodal context.

In the image acquisition process, various forms of degradation, including noise, blur, haze, and rain, are frequently introduced. These degradations typically arise from the inherent limitations of cameras or unfavorable ambient conditions. To recover clean images from their degraded versions, numerous specialized restoration methods have been developed, each targeting a specific type of degradation. Recently, all-in-one algorithms have garnered significant attention by addressing different types of degradations within a single model without requiring the prior information of the input degradation type. However, most methods purely operate in the spatial domain and do not delve into the distinct frequency variations inherent to different degradation types. To address this gap, we propose an adaptive all-in-one image restoration network based on frequency mining and modulation. Our approach is motivated by the observation that different degradation types impact the image content on different frequency subbands, thereby requiring different treatments for each restoration task. Specifically, we first mine low- and high-frequency information from the input features, guided by the adaptively decoupled spectra of the degraded image. The extracted features are then modulated by a bidirectional operator to facilitate interactions between different frequency components. Finally, the modulated features are merged into the original input for a progressively guided restoration. With this approach, the model achieves adaptive reconstruction by accentuating the informative frequency subbands according to different input degradations. Extensive experiments demonstrate that the proposed method, AdaIR, achieves state-of-the-art performance on different image restoration tasks, including image denoising, dehazing, deraining, motion deblurring, and low-light image enhancement. The code is available at https://github.com/c-yn/AdaIR.

Recent advancements in large language models (LLMs) have sparked optimism about their potential to accelerate scientific discovery, with a growing number of works proposing research agents that autonomously generate and validate new ideas. Despite this, no evaluations have shown that LLM systems can take the very first step of producing novel, expert-level ideas, let alone perform the entire research process. We address this by establishing an experimental design that evaluates research idea generation while controlling for confounders and performs the first comparison between expert NLP researchers and an LLM ideation agent. By recruiting over 100 NLP researchers to write novel ideas and blind reviews of both LLM and human ideas, we obtain the first statistically significant conclusion on current LLM capabilities for research ideation: we find LLM-generated ideas are judged as more novel (p < 0.05) than human expert ideas while being judged slightly weaker on feasibility. Studying our agent baselines closely, we identify open problems in building and evaluating research agents, including failures of LLM self-evaluation and their lack of diversity in generation.

Training and fine-tuning large language models (LLMs) come with challenges related to memory and computational requirements due to the increasing size of the model weights and the optimizer states. To tackle these challenges, various techniques have been developed, such as low-rank adaptation (LoRA), which involves introducing a parallel trainable low-rank matrix to the fixed pre-trained weights at each layer. However, these methods often fall short compared to the full-rank weight training approach, as they restrict the parameter search to a low-rank subspace. This limitation can disrupt training dynamics and may require a full-rank warm start to mitigate the impact. In this paper, we introduce a new method inspired by a phenomenon we formally prove: as training progresses, the rank of the estimated layer gradients gradually decreases and asymptotically approaches rank one. Leveraging this, our approach involves adaptively reducing the rank of the gradients during Adam optimization steps, using an efficient online-updating low-rank projections rule. We further present a randomized-svd scheme for efficiently finding the projection matrix. Our technique enables full-parameter fine-tuning with adaptive low-rank gradient updates, significantly reducing overall memory requirements during training compared to state-of-the-art methods while improving model performance in both pretraining and fine-tuning. Finally, we provide a convergence analysis of our method and demonstrate its merits for training and fine-tuning language and biological foundation models.

The inherent quadratic complexity of the attention mechanism in transformer models has driven the research community to explore alternative architectures with sub-quadratic complexity, such as state-space models. Mamba has established itself as a leading model within this emerging paradigm, achieving state-of-the-art results in various language modeling benchmarks. However, despite its impressive performance, Mamba's effectiveness is limited by its pre-training context length, resulting in a pronounced degradation when the model is tasked with handling longer contexts. Our investigation reveals that Mamba's inability to generalize effectively to long contexts is primarily due to the out-of-distribution (OOD) discretization steps. To address this critical limitation, we introduce _**MambaExtend**_, a novel framework designed to significantly enhance the context extension capabilities of Mamba. Specifically, MambaExtend leverages a _**training-free**_ approach to calibrate _only_ the scaling factors of discretization modules for different layers. We demonstrate both gradient-based and gradient-free zeroth-order optimization to learn the optimal scaling factors for each Mamba layer, requiring orders of magnitude fewer updates as opposed to the parameter fine-tuning-based alternatives. Using this approach, we achieve a training-free context extension of up to 32x, expanding the context from 2k to 64k tokens with minimal increases in perplexity. In contrast to existing fine-tuning methods, MambaExtend selectively calibrates the scaling factors, requiring up to $\mathbf{5.42 * 10^6} \times$ fewer parameter updates and incurring up to $\mathbf{3.87} \times$ lower peak memory usage, while delivering comparable or superior long-context performance across multiple tasks. Codes and checkpoints are available here$^1$.

Prompt learning has demonstrated promising results in fine-tuning pre-trained multimodal models. However, the performance improvement is limited when applied to more complex and fine-grained tasks. The reason is that most existing methods directly optimize the parameters involved in the prompt generation process through loss backpropagation, which constrains the richness and specificity of the prompt representations. In this paper, we propose Diffusion-Driven Prompt Generator (Diff-Prompt), aiming to use the diffusion model to generate rich and fine-grained prompt information for complex downstream tasks. Specifically, our approach consists of three stages. In the first stage, we train a Mask-VAE to compress the masks into latent space. In the second stage, we leverage an improved Diffusion Transformer (DiT) to train a prompt generator in the latent space, using the masks for supervision. In the third stage, we align the denoising process of the prompt generator with the pre-trained model in the semantic space, and use the generated prompts to fine-tune the model. We conduct experiments on a complex pixel-level downstream task, referring expression comprehension, and compare our method with various parameter-efficient fine-tuning approaches. Diff-Prompt achieves a maximum improvement of 8.87 in R@1 and 14.05 in R@5 compared to the foundation model and also outperforms other state-of-the-art methods across multiple metrics. The experimental results validate the effectiveness of our approach and highlight the potential of using generative models for prompt generation. Code is available at https://github.com/Kelvin-ywc/diff-prompt.

Language plays a vital role in the realm of human motion. Existing methods have largely depended on CLIP text embeddings for motion generation, yet they fall short in effectively aligning language and motion due to CLIP’s pretraining on static image-text pairs. This work introduces LaMP, a novel Language-Motion Pretraining model, which transitions from a language-vision to a more suitable language-motion latent space. It addresses key limitations by generating motion-informative text embeddings, significantly enhancing the relevance and semantics of generated motion sequences. With LaMP, we advance three key tasks: text-to-motion generation, motion-text retrieval, and motion captioning through aligned language-motion representation learning. For generation, LaMP instead of CLIP provides the text condition, and an autoregressive masked prediction is designed to achieve mask modeling without rank collapse in transformers. For retrieval, motion features from LaMP’s motion transformer interact with query tokens to retrieve text features from the text transformer, and vice versa. For captioning, we finetune a large language model with the language-informative motion features to develop a strong motion captioning model. In addition, we introduce the LaMP-BertScore metric to assess the alignment of generated motions with textual descriptions. Extensive experimental results on multiple datasets demonstrate substantial improvements over previous methods across all three tasks. Project page: https://aigc3d.github.io/LaMP