ICLR 2025 论文列表

The utility of synthetic data to enhance pretraining data quality and hence to improve downstream task accuracy has been widely explored in recent large language models (LLMs). Yet, these approaches fall inadequate in complex, multi-hop and mathematical reasoning tasks as the synthetic data typically fails to add complementary knowledge to the existing raw corpus. In this work, we propose a novel large-scale and diverse Math Informed syNthetic Dialogue (MIND) generation method that improves the mathematical reasoning ability of LLMs. Specifically, using MIND, we generate synthetic conversations based on OpenWebMath (OWM), resulting in a new math corpus, MIND-OWM. Our experiments with different conversational settings reveal that incorporating knowledge gaps between dialog participants is essential for generating high-quality math data. We further identify an effective way to format and integrate synthetic and raw data during pretraining to maximize the gain in mathematical reasoning, emphasizing the need to restructure raw data rather than use it as-is. Compared to pretraining just on raw data, a model pretrained on MIND-OWM shows significant boost in mathematical reasoning (GSM8K: +13.42%, MATH: +2.30%), including superior performance in specialized knowledge (MMLU: +4.55%, MMLU-STEM: +4.28%) and general purpose reasoning tasks (GENERAL REASONING: +2.51%).

We introduce MuirBench, a comprehensive benchmark that focuses on robust multi-image understanding capabilities of multimodal LLMs. MuirBench consists of 12 diverse multi-image tasks (e.g., scene understanding, ordering) that involve 10 categories of multi-image relations (e.g., multiview, temporal relations). Comprising 11,264 images and 2,600 multiple-choice questions, MuirBench is created in a pairwise manner, where each standard instance is paired with an unanswerable variant that has minimal semantic differences, in order for a reliable assessment. Evaluated upon 20 recent multi-modal LLMs, our results reveal that even the best-performing models like GPT-4o and Gemini Pro find it challenging to solve MuirBench, achieving 68.0% and 49.3% in accuracy. Open-source multimodal LLMs trained on single images can hardly generalize to multi-image questions, hovering below 33.3% in accuracy. These results highlight the importance of MuirBench in encouraging the community to develop multimodal LLMs that can look beyond a single image, suggesting potential pathways for future improvements.

Training and serving long-context large language models (LLMs) incurs substantial overhead. To address this, two critical steps are often required: a pretrained LLM typically undergoes a separate stage for context length extension by training on long-context data, followed by architectural modifications to reduce the overhead of KV cache during serving. This paper argues that integrating length extension with a GPU-friendly KV cache reduction architecture not only reduces training overhead during length extension, but also achieves better long-context performance. This leads to our proposed LongGen, which finetunes a pretrained LLM into an efficient architecture during length extension. LongGen builds on three key insights: (1) Sparse attention patterns, such as window attention (attending to recent tokens), attention sink (initial ones), and blockwise sparse attention (strided token blocks) are well-suited for building efficient long-context models, primarily due to their GPU-friendly memory access patterns, enabling efficiency gains not just theoretically but in practice as well. (2) It is essential for the model to have direct access to all tokens. A hybrid architecture with 1/3 full attention layers and 2/3 efficient ones achieves a balanced trade-off between efficiency and long-context performance. (3) Lightweight training on 5B long-context data is sufficient to extend the hybrid model's context length from 4K to 128K. We evaluate LongGen on both Llama-2 7B and Llama-2 70B, demonstrating its effectiveness across different scales. During training with 128K-long contexts, LongGen achieves 1.55x training speedup and reduces wall-clock time by 36%, compared to a full-attention baseline. During inference, LongGen reduces KV cache memory by 62%, achieving 1.67x prefilling speedup and 1.41x decoding speedup. Compared to baselines that apply KV-cache reduction techniques to full-attention long-context LLMs, LongGen achieves substantially stronger performance not only on the Needle-in-a-Haystack retrieval task, but also on more challenging long-context reasoning tasks, including BABILong and RULER.

As online video content rapidly grows, the task of text-video retrieval (TVR) becomes increasingly important. A key challenge in TVR is the information asymmetry between video and text: videos are inherently richer in information, while their textual descriptions often capture only fragments of this complexity. This paper introduces a novel, data-centric framework to bridge this gap by enriching textual representations to better match the richness of video content. During training, videos are segmented into event-level clips and captioned to ensure comprehensive coverage. During retrieval, a large language model (LLM) generates semantically diverse queries to capture a broader range of possible matches. To enhance retrieval efficiency, we propose a query selection mechanism that identifies the most relevant and diverse queries, reducing computational cost while improving accuracy. Our method achieves state-of-the-art results across multiple benchmarks, demonstrating the power of data-centric approaches in addressing information asymmetry in TVR. This work paves the way for new research focused on leveraging data to improve cross-modal retrieval.

Recent advances in text-to-image generation have achieved impressive aesthetic quality, making these models usable for both personal and commercial purposes. However, in the fields of marketing and advertising, images are often created to be more engaging, as reflected in user behaviors such as increasing clicks, likes, and purchases, in addition to being aesthetically pleasing. To this end, we introduce the challenge of optimizing the image generation process for improved viewer engagement. In order to study image engagement and utility in real-world marketing scenarios, we collect *EngagingImageNet*, the first large-scale dataset of images, along with associated user engagement metrics. Further, we find that existing image evaluation metrics like aesthetics, CLIPScore, PickScore, ImageReward, *etc.* are unable to capture viewer engagement. To address the lack of reliable metrics for assessing image utility, we use the *EngagingImageNet* dataset to train *EngageNet*, an engagement-aware Vision Language Model (VLM) that predicts viewer engagement of images by leveraging contextual information about the tweet content, enterprise details, and posting time. We then explore methods to enhance the engagement of text-to-image models, making initial strides in this direction. These include conditioning image generation on improved prompts, supervised fine-tuning of stable diffusion on high-performing images, and reinforcement learning to align stable diffusion with *EngageNet*-based reward signals, all of which lead to the generation of images with higher viewer engagement. Finally, we propose the *Engagement Arena*, to benchmark text-to-image models based on their ability to generate engaging images, using *EngageNet* as the evaluator, thereby encouraging the research community to measure further advances in the engagement of text-to-image modeling. These contributions provide a new pathway for advancing utility-driven image generation, with significant implications for the commercial application of image generation. We have released our code and dataset on [behavior-in-the-wild.github.io/image-engagement](https://behavior-in-the-wild.github.io/image-engagement).

The frequent retrieval of Key-Value (KV) cache data has emerged as a significant factor contributing to the inefficiency of the inference process in large language models. Previous research has demonstrated that a small subset of critical KV cache tokens largely influences attention outcomes, leading to methods that either employ fixed sparsity patterns or dynamically select critical tokens based on the query. While dynamic sparse patterns have proven to be more effective, they introduce significant computational overhead, as critical tokens must be reselected for each self-attention computation. In this paper, we reveal substantial similarities in KV cache token criticality across neighboring queries, layers, and heads. Motivated by this insight, we propose HShare, a hierarchical KV sharing framework. HShare facilitates the sharing of critical KV cache token indices across layers, heads, and queries, which significantly reduces the computational overhead associated with query-aware dynamic token sparsity. In addition, we introduce a greedy algorithm that dynamically determines the optimal layer-level and head-level sharing configuration for the decoding phase. We evaluate the effectiveness and efficiency of HShare across various tasks using three models: LLaMA2-7b, LLaMA3-70b, and Mistral-7b. Experimental results demonstrate that HShare achieves competitive accuracy with different sharing ratios, while delivering up to an $8.6\times$ speedup in self-attention operations and a $2.7\times$ improvement in end-to-end throughput compared with FlashAttention2 and GPT-fast respectively. The source code is publicly available at ~\url{https://github.com/wuhuaijin/HShare}.

Recent advancements in multimodal models highlight the value of rewritten captions for improving performance, yet key challenges remain. For example, while synthetic captions often provide superior quality and image-text alignment, it is not clear whether they can fully replace AltTexts: the role of synthetic captions and their interaction with original web-crawled AltTexts in pre-training is still not well understood. Moreover, different multimodal foundation models may have unique preferences for specific caption formats, but efforts to identify the optimal captions for each model remain limited. In this work, we propose a novel, controllable, and scalable captioning pipeline designed to generate diverse caption formats tailored to various multimodal models. By examining short synthetic captions (SSC) and descriptive synthetic captions (DSC) as case studies, we systematically explore their effects and interactions with AltTexts across models such as CLIP, multimodal LLMs, and diffusion models. Our findings reveal that a hybrid approach that keeps both synthetic captions and AltTexts can outperform the use of synthetic captions alone, improving both alignment and performance, with each model demonstrating preferences for particular caption formats. This comprehensive analysis provides valuable insights into optimizing captioning strategies, thereby advancing the pre-training of multimodal foundation models.

Transformer and its variants have shown great potential for various vision tasks in recent years, including image classification, object detection and segmentation. Meanwhile, recent studies also reveal that with proper architecture design, convolutional networks (ConvNets) also achieve competitive performance with transformers. However, no prior methods have explored to utilize pure convolution to build a Transformer-style Decoder module, which is essential for Encoder-Decoder architecture like Detection Transformer (DETR). To this end, in this paper we explore whether we could build query-based detection and segmentation framework with ConvNets instead of sophisticated transformer architecture. We propose a novel mechanism dubbed InterConv to perform interaction between object queries and image features via convolutional layers. Equipped with the proposed InterConv, we build Detection ConvNet (DECO), which is composed of a backbone and convolutional encoder-decoder architecture. We compare the proposed DECO against prior detectors on the challenging COCO benchmark. Despite its simplicity, our DECO achieves competitive performance in terms of detection accuracy and running speed. Specifically, with the ResNet-18 and ResNet-50 backbone, our DECO achieves $40.5\%$ and $47.8\%$ AP with $66$ and $34$ FPS, respectively. The proposed method is also evaluated on the segment anything task, demonstrating similar performance and higher efficiency. We hope the proposed method brings another perspective for designing architectures for vision tasks. Codes are available at \url{https://github.com/xinghaochen/DECO} and \url{https://github.com/mindspore-lab/models/tree/master/research/huawei-noah/DECO}.

Understanding and manipulating the causal generation mechanisms in language models is essential for controlling their behavior. Previous work has primarily relied on techniques such as representation surgery---e.g., model ablations or manipulation of linear subspaces tied to specific concepts---to intervene on these models. To understand the impact of interventions precisely, it is useful to examine counterfactuals---e.g., how a given sentence would have appeared had it been generated by the model following a specific intervention. We highlight that counterfactual reasoning is conceptually distinct from interventions, as articulated in Pearl's causal hierarchy. Based on this observation, we propose a framework for generating true string counterfactuals by reformulating language models as a structural equation model using the Gumbel-max trick, which we called Gumbel counterfactual generation. This reformulation allows us to model the joint distribution over original strings and their counterfactuals resulting from the same instantiation of the sampling noise. We develop an algorithm based on hindsight Gumbel sampling that allows us to infer the latent noise variables and generate counterfactuals of observed strings. Our experiments demonstrate that the approach produces meaningful counterfactuals while at the same time showing that commonly used intervention techniques have considerable undesired side effects.

In this paper, we aim to build an adversarially robust zero-shot image classifier that can accurately and efficiently classify unseen examples while defending against unforeseen adversarial attacks, addressing critical challenges in real-world safety-sensitive scenarios. To achieve this, we focus on two key challenges: zero-shot classification and defense against unforeseen attacks. We ground our work on CLIP, a vision-language pre-trained model to perform zero-shot classification. To defend against unforeseen attacks, we adopt a purification approach, as it is independent of specific attack types. We then define a purification risk as the KL divergence between the joint distributions of the purification and attack process. The derived lower bound of purification risk inspires us to explore purification in CLIP's multi-modal latent space. We propose a CLIP-based purification method called CLIPure, which has two variants: _CLIPure-Diff_, which models image likelihood with a generative process of its latent vector, and _CLIPure-Cos_, which models the likelihood based on the similarity between embeddings of the image and a blank template. As far as we know, CLIPure is the first purification method in latent space and _CLIPure-Cos_ is the first purification method not relying on generative models, substantially improving defense efficiency. Extensive experimental results show that the robustness achieved by CLIPure is within a small gap of clean accuracy, outperforming SOTA robustness by a large margin, e.g., from 71.7\% to **91.1\%** on CIFAR10, from 59.6\% to **72.6\%** on ImageNet, and **108\%** relative improvements of average robustness on the 13 datasets over previous SOTA, with only 14\% extra inference cost and no additional training.

Aligning vision and language concepts at a finer level remains an essential topic of multimodal large language models (MLLMs), particularly for tasks such as referring and grounding. Existing methods, such as *proxy encoding* and *geometry encoding* genres, incorporate additional syntax to encode spatial information, imposing extra burdens when communicating between language with vision modules. In this study, we propose ClawMachine, offering a new methodology that explicitly notates each entity using **token collectives**—groups of visual tokens that collaboratively represent higher-level semantics. A hybrid perception mechanism is also explored to perceive and understand scenes from both discrete and continuous spaces. Our method unifies the prompt and answer of visual referential tasks without using additional syntax. By leveraging a joint vision-language vocabulary, ClawMachine integrates referring and grounding in an auto-regressive manner, demonstrating great potential with scaled up pre-training data. Experiments show that ClawMachine achieves superior performance on scene-level and referential understanding tasks with higher efficiency. It also exhibits the potential to integrate multi-source information for complex visual reasoning, which is beyond the capability of many MLLMs. Our code is available at https://github.com/martian422/ClawMachine.

Embedding models play a crucial role in a variety of downstream tasks, including semantic similarity, information retrieval, and clustering. While there has been a surge of interest in developing universal text embedding models that generalize across tasks (e.g., MTEB), progress in learning universal multimodal embedding models has been comparatively slow, despite their importance and practical applications. In this work, we explore the potential of building universal multimodal embeddings capable of handling a broad range of downstream tasks. Our contributions are twofold: (1) we propose MMEB (Massive Multimodal Embedding Benchmark), which covers four meta-tasks (classification, visual question answering, multimodal retrieval, and visual grounding) and 36 datasets, including 20 training datasets and 16 evaluation datasets spanning both in-distribution and out-of-distribution tasks, and (2) VLM2Vec (Vision-Language Model → Vector), a contrastive training framework that transforms any vision-language model into an embedding model through contrastive training on MMEB. Unlike previous models such as CLIP and BLIP, which encode text and images independently without task-specific guidance, VLM2Vec can process any combination of images and text while incorporating task instructions to generate a fixed-dimensional vector. We develop a series of VLM2Vec models based on state-of-the-art VLMs, including Phi-3.5-V, LLaVA-1.6, and Qwen2-VL, and evaluate them on MMEB’s benchmark. With LoRA tuning, VLM2Vec achieves a 10% to 20% improvement over existing multimodal embedding models on MMEB’s evaluation sets. Our findings reveal that VLMs are surprisingly strong embedding models.

Speculative sampling is a promising approach to accelerate the decoding stage for Large Language Models (LLMs). Recent advancements that leverage target LLM's contextual information, such as hidden states and KV cache, have shown significant practical improvements. However, these approaches suffer from inconsistent context between training and decoding. We also observe another discrepancy between the training and decoding objectives in existing speculative sampling methods. In this work, we propose a solution named HArmonized Speculative Sampling (HASS) that learns harmonized representations to address these issues. HASS accelerates the decoding stage without adding inference overhead through harmonized objective distillation and harmonized context alignment. Experiments on four LLaMA models demonstrate that HASS achieves 2.81x-4.05x wall-clock time speedup ratio averaging across three datasets, surpassing EAGLE-2 by 8%-20%. The code is available at https://github.com/HArmonizedSS/HASS.

Multimodal Large Language Models (MLLMs) are thriving through continuous fine-tuning by LLMs. Driven by the law that "scale is everything", MLLMs expand their training sets during version iterations. In this paper, we propose a large-scale training data engine built around an evaluating-exploring-evolving (E3) loop. Evaluating the data provides insights into its characteristics. Exploring quality rules helps identify which data enhances training. Together, these processes facilitate the systematic evolution of new, high-quality data. With the E3 loop, we introduce ZooProbe, an efficient data engine for MLLMs. First, the problem of data expansion is formalized as a tree of sampling and growth. ZooProbe introduces a small-scale model *zoo* to obtain comprehensive evaluations for child datasets. From multiple perspectives, visual, textual, and multimodal models cover over 50 dimensions of intrinsic and meta attributes, such as object and topic distribution, and higher-level properties, like annotation quality and scene complexity. ZooProbe constructs based on A$^\star$ search, modeling the heuristic function as a quality estimate from data evaluation results. It dynamically explores the rule of data quality based on the model state of the *probe* datasets. Additionally, it evolves new targeted data with identified high-quality rules. We also develop an extra heuristic quality ranker with the data utilized and discarded during the expansion. Our experiments show that ZooProbe significantly breaks the scaling law in multimodal instruction fine-tuning at scales of 260$k$ and below. ZooProbe generates high-quality data that accelerates MLLM training and enhances performance, automating the evolution of large-scale training data.

The Transformer architecture is widely deployed in many popular and impactful Large Language Models. At its core is the attention mechanism for calculating correlations between pairs of tokens. Performing an attention computation takes quadratic time in the input size, and had become the time bottleneck for transformer operations. In order to circumvent this, researchers have used a variety of approaches, including designing heuristic algorithms for performing attention computations faster, and proposing alternatives to the attention mechanism which can be computed more quickly. For instance, state space models such as Mamba were designed to replace attention with an almost linear time alternative. In this paper, we prove that any such approach cannot perform important tasks that Transformer is able to perform (assuming a popular conjecture from fine-grained complexity theory). We focus on document similarity tasks, where one is given as input many documents and would like to find a pair which is (approximately) the most similar. We prove that Transformer is able to perform this task, and we prove that this task cannot be performed in truly subquadratic time by any algorithm. Thus, any model which can be evaluated in subquadratic time – whether because of subquadratic-time heuristics for attention, faster attention replacements like Mamba, or any other reason – cannot perform this task. In other words, in order to perform tasks that (implicitly or explicitly) involve document similarity, one may as well use Transformer and cannot avoid its quadratic running time.

The Sparse Mixture of Experts (SMoE) has been widely employed to enhance the efficiency of training and inference for Transformer-based foundational models, yielding promising results. However, the performance of SMoE heavily depends on the choice of hyper-parameters, such as the number of experts and the number of experts to be activated (referred to as top-$k$), resulting in significant computational overhead due to the extensive model training by searching over various hyper-parameter configurations. As a remedy, we introduce the Dynamic Mixture of Experts (DynMoE) technique. DynMoE incorporates (1) a novel gating method that enables each token to automatically determine the number of experts to activate. (2) An adaptive process automatically adjusts the number of experts during training. Extensive numerical results across Vision, Language, and Vision-Language tasks demonstrate the effectiveness of our approach to achieve competitive performance compared to GMoE for vision and language tasks, and MoE-LLaVA for vision-language tasks, while maintaining efficiency by activating fewer parameters. Our code is available at \url{https://github.com/LINs-lab/DynMoE}.

With the growing capabilities of large language models (LLMs), they are increasingly applied in areas like intelligent customer service, code generation, and knowledge management. Natural language (NL) prompts act as the ``APIs'' for human-LLM interaction. To improve prompt quality, best practices for prompt engineering (PE) have been developed, including writing guidelines and templates. Building on this, we propose Controlled NL for Prompt (CNL-P), which not only incorporates PE best practices but also draws on key principles from software engineering (SE). CNL-P introduces precise grammar structures and strict semantic norms, further eliminating NL's ambiguity, allowing for a declarative but structured and accurate expression of user intent. This helps LLMs better interpret and execute the prompts, leading to more consistent and higher-quality outputs. We also introduce an NL2CNL-P conversion tool based on LLMs, enabling users to write prompts in NL, which are then transformed into CNL-P format, thus lowering the learning curve of CNL-P. In particular, we develop a linting tool that checks CNL-P prompts for syntactic and semantic accuracy, applying static analysis techniques to NL for the first time. Extensive experiments demonstrate that CNL-P enhances the quality of LLM responses through the novel and organic synergy of PE and SE. We believe that CNL-P can bridge the gap between emerging PE and traditional SE, laying the foundation for a new programming paradigm centered around NL. To further demonstrate the effectiveness of CNL-P, we develop the AI-native CNL-P IDE, UGAiForge that leverages CNL-P to enable users to Think, Describe, and Build with AI and for AI. For detailed information, please refer https://ugaiforge.ai.

Protecting the intellectual property of open-source Large Language Models (LLMs) is very important, because training LLMs costs extensive computational resources and data. Therefore, model owners and third parties need to identify whether a suspect model is a subsequent development of the victim model. To this end, we propose a training-free REEF to identify the relationship between the suspect and victim models from the perspective of LLMs' feature representations. Specifically, REEF computes and compares the centered kernel alignment similarity between the representations of a suspect model and a victim model on the same samples. This training-free REEF does not impair the model's general capabilities and is robust to sequential fine-tuning, pruning, model merging, and permutations. In this way, REEF provides a simple and effective way for third parties and models' owners to protect LLMs' intellectual property together. Our code is publicly accessible at https://github.com/AI45Lab/REEF.

Transformers are widely used as generic backbones in computer vision, despite initially introduced for natural language processing. Recently, the Long Short-Term Memory (LSTM) has been extended to a scalable and performant architecture - the xLSTM - which overcomes long-standing LSTM limitations via exponential gating and parallelizable matrix memory structure. In this paper, we introduce Vision-LSTM (ViL), an adaption of the xLSTM building blocks to computer vision. ViL comprises a stack of xLSTM blocks where odd blocks process the sequence of patch tokens from top to bottom while even blocks go from bottom to top. ViL achieves strong performances on classification, transfer learning and segmentation tasks as well as a beneficial pre-training cost-to-performance trade-off. Experiments show that ViL holds promise to be further deployed as new generic backbone for computer vision architectures.