ICLR 2025 论文列表

Understanding and explaining differences between audio recordings is crucial for fields like audio forensics, quality assessment, and audio generation. This involves identifying and describing audio events, acoustic scenes, signal characteristics, and their emotional impact on listeners. This paper stands out as the first work to comprehensively study the task of explaining audio differences and then propose benchmark, baselines for the task. First, we present two new datasets for audio difference explanation derived from the AudioCaps and Clotho audio captioning datasets. Using Large Language Models (LLMs), we generate three levels of difference explanations: (1) concise descriptions of audio events and objects, (2) brief sentences about audio events, acoustic scenes, and signal properties, and (3) comprehensive explanations that include semantics and listener emotions. For the baseline, we use prefix tuning where audio embeddings from two audio files are used to prompt a frozen language model. Our empirical analysis and ablation studies reveal that the naive baseline struggles to distinguish perceptually similar sounds and generate detailed tier 3 explanations. To address these limitations, we propose ADIFF, which introduces a cross-projection module, position captioning, and a three-step training process to enhance the model’s ability to produce detailed explanations. We evaluate our model using objective metrics and human evaluation and show our model enhancements lead to significant improvements in performance over naive baseline and SoTA Audio-Language Model (ALM) Qwen Audio. Lastly, we conduct multiple ablation studies to study the effects of cross-projection, language model parameters, position captioning, third stage fine-tuning, and present our findings. Our benchmarks, findings, and strong baseline pave the way for nuanced and human-like explanations of audio differences.

Recently, human-computer interaction with various modalities has shown promising applications, like GPT-4o and Gemini. Meanwhile, multimodal representation models have emerged as the foundation for these versatile multimodal understanding and generation pipeline. Models like CLIP, CLAP and ImageBind can map their specialized modalities into respective joint spaces. To construct a high-quality omni representation space that can be shared and expert in any modality, we propose to merge these advanced models into a unified space in scale. With this insight, we present \textbf{OmniBind}, advanced multimodal joint representation models via fusing knowledge of 14 pre-trained spaces, which support 3D, audio, image, video and language inputs. To alleviate the interference between different knowledge sources in integrated space, we dynamically assign weights to different spaces by learning routers with two objectives: cross-modal overall alignment and language representation decoupling. Notably, since binding and routing spaces only require lightweight networks, OmniBind is extremely training-efficient. Extensive experiments demonstrate the versatility and superiority of OmniBind as an omni representation model, highlighting its great potential for diverse applications, such as any-query and composable multimodal understanding.

Zero-shot medical detection can further improve detection performance without relying on annotated medical images even upon the fine-tuned model, showing great clinical value. Recent studies leverage grounded vision-language models (GLIP) to achieve this by using detailed disease descriptions as prompts for the target disease name during the inference phase. However, these methods typically treat prompts as equivalent context to the target name, making it difficult to assign specific disease knowledge based on visual information, leading to a coarse alignment between images and target descriptions. In this paper, we propose StructuralGLIP, which introduces an auxiliary branch to encode prompts into a latent knowledge bank layer-by-layer, enabling more context-aware and fine-grained alignment. Specifically, in each layer, we select highly similar features from both the image representation and the knowledge bank, forming structural representations that capture nuanced relationships between image patches and target descriptions. These features are then fused across modalities to further enhance detection performance. Extensive experiments demonstrate that StructuralGLIP achieves a +4.1\% AP improvement over prior state-of-the-art methods across seven zero-shot medical detection benchmarks, and consistently improves fine-tuned models by +3.2\% AP on endoscopy image datasets.

Information retrieval has transitioned from standalone systems into essential components across broader applications, with indexing efficiency, cost-effectiveness, and freshness becoming increasingly critical yet often overlooked. In this paper, we introduce SemI-parametric Disentangled Retrieval (SiDR), a bi-encoder retrieval framework that decouples retrieval index from neural parameters to enable efficient, low-cost, and parameter-agnostic indexing for emerging use cases. Specifically, in addition to using embeddings as indexes like existing neural retrieval methods, SiDR supports a non-parametric tokenization index for search, achieving BM25-like indexing complexity with significantly better effectiveness. Our comprehensive evaluation across 16 retrieval benchmarks demonstrates that SiDR outperforms both neural and term-based retrieval baselines under the same indexing workload: (i) When using an embedding-based index, SiDR exceeds the performance of conventional neural retrievers while maintaining similar training complexity; (ii) When using a tokenization-based index, SiDR drastically reduces indexing cost and time, matching the complexity of traditional term-based retrieval, while consistently outperforming BM25 on all in-domain datasets; (iii) Additionally, we introduce a late parametric mechanism that matches BM25 index preparation time while outperforming other neural retrieval baselines in effectiveness.

Illumination degradation image restoration (IDIR) techniques aim to improve the visibility of degraded images and mitigate the adverse effects of deteriorated illumination. Among these algorithms, diffusion-based models (DM) have shown promising performance but are often burdened by heavy computational demands and pixel misalignment issues when predicting the image-level distribution. To tackle these problems, we propose to leverage DM within a compact latent space to generate concise guidance priors and introduce a novel solution called Reti-Diff for the IDIR task. Specifically, Reti-Diff comprises two significant components: the Retinex-based latent DM (RLDM) and the Retinex-guided transformer (RGformer). RLDM is designed to acquire Retinex knowledge, extracting reflectance and illumination priors to facilitate detailed reconstruction and illumination correction. RGformer subsequently utilizes these compact priors to guide the decomposition of image features into their respective reflectance and illumination components. Following this, RGformer further enhances and consolidates these decomposed features, resulting in the production of refined images with consistent content and robustness to handle complex degradation scenarios. Extensive experiments demonstrate that Reti-Diff outperforms existing methods on three IDIR tasks, as well as downstream applications.

The site conditions that make astronomical observatories in space and on the ground so desirable---cold and dark---demand a physical remoteness that leads to limited data transmission capabilities. Such transmission limitations directly bottleneck the amount of data acquired and in an era of costly modern observatories, any improvements in lossless data compression has the potential scale to billions of dollars worth of additional science that can be accomplished on the same instrument. Traditional lossless methods for compressing astrophysical data are manually designed. Neural data compression, on the other hand, holds the promise of learning compression algorithms end-to-end from data and outperforming classical techniques by leveraging the unique spatial, temporal, and wavelength structures of astronomical images. This paper introduces [AstroCompress](https://huggingface.co/AstroCompress): a neural compression challenge for astrophysics data, featuring four new datasets (and one legacy dataset) with 16-bit unsigned integer imaging data in various modes: space-based, ground-based, multi-wavelength, and time-series imaging. We provide code to easily access the data and benchmark seven lossless compression methods (three neural and four non-neural, including all practical state-of-the-art algorithms). Our results on lossless compression indicate that lossless neural compression techniques can enhance data collection at observatories, and provide guidance on the adoption of neural compression in scientific applications. Though the scope of this paper is restricted to lossless compression, we also comment on the potential exploration of lossy compression methods in future studies.

We introduce CHAMP, a novel method for learning sequence-to-sequence, multi-hypothesis 3D human poses from 2D keypoints by leveraging a conditional distribution with a diffusion model. To predict a single output 3D pose sequence, we generate and aggregate multiple 3D pose hypotheses. For better aggregation results, we develop a method to score these hypotheses during training, effectively integrating conformal prediction into the learning process. This process results in a differentiable conformal predictor that is trained end-to-end with the 3D pose estimator. Post-training, the learned scoring model is used as the conformity score, and the 3D pose estimator is combined with a conformal predictor to select the most accurate hypotheses for downstream aggregation. Our results indicate that using a simple mean aggregation on the conformal prediction-filtered hypotheses set yields competitive results. When integrated with more sophisticated aggregation techniques, our method achieves state-of-the-art performance across various metrics and datasets while inheriting the probabilistic guarantees of conformal prediction.

Comprehensive evaluation of Multimodal Large Language Models (MLLMs) has recently garnered widespread attention in the research community. However, we observe that existing benchmarks present several common barriers that make it difficult to measure the significant challenges that models face in the real world, including: 1) small data scale leads to a large performance variance; 2) reliance on model-based annotations results in restricted data quality; 3) insufficient task difficulty, especially caused by the limited image resolution. To tackle these issues, we introduce MME-RealWorld. Specifically, we collect more than $300$ K images from public datasets and the Internet, filtering $13,366$ high-quality images for annotation. This involves the efforts of professional $25$ annotators and $7$ experts in MLLMs, contributing to $29,429$ question-answer pairs that cover $43$ subtasks across $5$ real-world scenarios, extremely challenging even for humans. As far as we know, **MME-RealWorld is the largest manually annotated benchmark to date, featuring the highest resolution and a targeted focus on real-world applications**. We further conduct a thorough evaluation involving $29$ prominent MLLMs, such as GPT-4o, Gemini 1.5 Pro, and Claude 3.5 Sonnet. Our results show that even the most advanced models struggle with our benchmarks, where none of them reach 60\% accuracy. The challenges of perceiving high-resolution images and understanding complex real-world scenarios remain urgent issues to be addressed. The data and evaluation code are released in our Project Page.

Lifelong user behavior sequences are crucial for capturing user interests and predicting user responses in modern recommendation systems. A two-stage paradigm is typically adopted to handle these long sequences: a subset of relevant behaviors is first searched from the original long sequences via an attention mechanism in the first stage and then aggregated with the target item to construct a discriminative representation for prediction in the second stage. In this work, we identify and characterize, for the first time, a neglected deficiency in existing long-sequence recommendation models: a single set of embeddings struggles with learning both attention and representation, leading to interference between these two processes. Initial attempts to address this issue with some common methods (e.g., linear projections---a technique borrowed from language processing) proved ineffective, shedding light on the unique challenges of recommendation models. To overcome this, we propose the Decoupled Attention and Representation Embeddings (DARE) model, where two distinct embedding tables are initialized and learned separately to fully decouple attention and representation. Extensive experiments and analysis demonstrate that DARE provides more accurate searches of correlated behaviors and outperforms baselines with AUC gains up to 9‰ on public datasets and notable improvements on Tencent's advertising platform. Furthermore, decoupling embedding spaces allows us to reduce the attention embedding dimension and accelerate the search procedure by 50\% without significant performance impact, enabling more efficient, high-performance online serving. Code in PyTorch for experiments, including model analysis, is available at https://github.com/thuml/DARE.

Large Language Models (LLMs) have demonstrated improved generation performance by incorporating externally retrieved knowledge, a process known as retrieval-augmented generation (RAG). Despite the potential of this approach, existing studies evaluate RAG effectiveness by 1) assessing retrieval and generation components jointly, which obscures retrieval's distinct contribution, or 2) examining retrievers using traditional metrics such as NDCG, which creates a gap in understanding retrieval's true utility in the overall generation process. To address the above limitations, in this work, we introduce an automatic evaluation method that measures retrieval quality through the lens of information gain within the RAG framework. Specifically, we propose Semantic Perplexity (SePer), a metric that captures the LLM's internal belief about the correctness of the retrieved information. We quantify the utility of retrieval by the extent to which it reduces semantic perplexity post-retrieval. Extensive experiments demonstrate that SePer not only aligns closely with human preferences but also offers a more precise and efficient evaluation of retrieval utility across diverse RAG scenarios.

Building interactable replicas of articulated objects is a key challenge in computer vision. Existing methods often fail to effectively integrate information across different object states, limiting the accuracy of part-mesh reconstruction and part dynamics modeling, particularly for complex multi-part articulated objects. We introduce ArtGS, a novel approach that leverages 3D Gaussians as a flexible and efficient representation to address these issues. Our method incorporates canonical Gaussians with coarse-to-fine initialization and updates for aligning articulated part information across different object states, and employs a skinning-inspired part dynamics modeling module to improve both part-mesh reconstruction and articulation learning. Extensive experiments on both synthetic and real-world datasets, including a new benchmark for complex multi-part objects, demonstrate that ArtGS achieves state-of-the-art performance in joint parameter estimation and part mesh reconstruction. Our approach significantly improves reconstruction quality and efficiency, especially for multi-part articulated objects. Additionally, we provide comprehensive analyses of our design choices, validating the effectiveness of each component to highlight potential areas for future improvement.

We present RASO, a foundation model designed to Recognize Any Surgical Object, offering robust open-set recognition capabilities across a broad range of surgical procedures and object classes, in both surgical images and videos. RASO leverages a novel weakly-supervised learning framework that generates tag-image-text pairs automatically from large-scale unannotated surgical lecture videos, significantly reducing the need for manual annotations. Our scalable data generation pipeline gathers 2,200 surgical procedures and produces 3.6 million tag annotations across 2,066 unique surgical tags. Our experiments show that RASO achieves improvements of 2.9 mAP, 4.5 mAP, 10.6 mAP, and 7.2 mAP on four standard surgical benchmarks respectively in zero-shot settings, and surpasses state-of-the-art models in supervised surgical action recognition tasks. We will open-source our code, model, and dataset to facilitate further research.

Implicit functions such as Neural Radiance Fields (NeRFs), occupancy networks, and signed distance functions (SDFs) have become pivotal in computer vision for reconstructing detailed object shapes from sparse views. Achieving optimal performance with these models can be challenging due to the extreme sparsity of inputs and distribution shifts induced by data corruptions. To this end, large, noise-free synthetic datasets can serve as shape priors to help models fill in gaps, but the resulting reconstructions must be approached with caution. Uncertainty estimation is crucial for assessing the quality of these reconstructions, particularly in identifying areas where the model is uncertain about the parts it has inferred from the prior. In this paper, we introduce Dropsembles, a novel method for uncertainty estimation in tuned implicit functions. We demonstrate the efficacy of our approach through a series of experiments, starting with toy examples and progressing to a real-world scenario. Specifically, we train a Convolutional Occupancy Network on synthetic anatomical data and test it on low-resolution MRI segmentations of the lumbar spine. Our results show that Dropsembles achieve the accuracy and calibration levels of deep ensembles but with significantly less computational cost.

Vision Language Models (VLMs) have recently been leveraged to generate robotic actions, forming Vision-Language-Action (VLA) models. However, directly adapting a pretrained VLM for robotic control remains challenging, particularly when constrained by a limited number of robot demonstrations. In this work, we introduce LLaRA: Large Language and Robotics Assistant, a framework that formulates robot action policy as visuo-textual conversations and enables an efficient transfer of a pretrained VLM into a powerful VLA, motivated by the success of visual instruction tuning in Computer Vision. First, we present an automated pipeline to generate conversation-style instruction tuning data for robots from existing behavior cloning datasets, aligning robotic actions with image pixel coordinates. Further, we enhance this dataset in a self-supervised manner by defining six auxiliary tasks, without requiring any additional action annotations. We show that a VLM finetuned with a limited amount of such datasets can produce meaningful action decisions for robotic control. Through experiments across multiple simulated and real-world tasks, we demonstrate that LLaRA achieves state-of-the-art performance while preserving the generalization capabilities of large language models. The code, datasets, and pretrained models are available at https://github.com/LostXine/LLaRA.

Unlike traditional vision-only models, vision language models (VLMs) offer an intuitive way to access visual content through language prompting by combining a large language model (LLM) with a vision encoder. However, both the LLM and the vision encoder come with their own set of biases, cue preferences, and shortcuts, which have been rigorously studied in uni-modal models. A timely question is how such (potentially misaligned) biases and cue preferences behave under multi-modal fusion in VLMs. As a first step towards a better understanding, we investigate a particularly well-studied vision-only bias - the texture vs. shape bias and the dominance of local over global information. As expected, we find that VLMs inherit this bias to some extent from their vision encoders. Surprisingly, the multi-modality alone proves to have important effects on the model behavior, i.e., the joint training and the language querying change the way visual cues are processed. While this direct impact of language-informed training on a model's visual perception is intriguing, it raises further questions on our ability to actively steer a model's output so that its prediction is based on particular visual cues of the user's choice. Interestingly, VLMs have an inherent tendency to recognize objects based on shape information, which is different from what a plain vision encoder would do. Further active steering towards shape-based classifications through language prompts is however limited. In contrast, active VLM steering towards texture-based decisions through simple natural language prompts is often more successful.

Scene image editing is crucial for entertainment, photography, and advertising design. Existing methods solely focus on either 2D individual object or 3D global scene editing. This results in a lack of a unified approach to effectively control and manipulate scenes at the 3D level with different levels of granularity. In this work, we propose 3DitScene, a novel and unified scene editing framework leveraging language-guided disentangled Gaussian Splatting that enables seamless editing from 2D to 3D, allowing precise control over scene composition and individual objects. We first incorporate 3D Gaussians that are refined through generative priors and optimization techniques. Language features from CLIP then introduce semantics into 3D geometry for object disentanglement. With the disentangled Gaussians, 3DitScene allows for manipulation at both the global and individual levels, revolutionizing creative expression and empowering control over scenes and objects. Experimental results demonstrate the effectiveness and versatility of 3DitScene in scene image editing.

Enabling intelligent agents to comprehend and interact with 3D environments through natural language is crucial for advancing robotics and human-computer interaction. A fundamental task in this field is ego-centric 3D visual grounding, where agents locate target objects in real-world 3D spaces based on verbal descriptions. However, this task faces two significant challenges: (1) loss of fine-grained visual semantics due to sparse fusion of point clouds with ego-centric multi-view images, (2) limited textual semantic context due to arbitrary language descriptions. We propose DenseGrounding, a novel approach designed to address these issues by enhancing both visual and textual semantics. For visual features, we introduce the Hierarchical Scene Semantic Enhancer, which retains dense semantics by capturing fine-grained global scene features and facilitating cross-modal alignment. For text descriptions, we propose a Language Semantic Enhancer that leverage large language models to provide rich context and diverse language descriptions with additional context during model training. Extensive experiments show that DenseGrounding significantly outperforms existing methods in overall accuracy, achieving improvements of **5.81%** and **7.56%** when trained on the comprehensive full training dataset and smaller mini subset, respectively, further advancing the SOTA in ego-centric 3D visual grounding. Our method also achieves **1st place** and receives **Innovation Award** in the 2024 Autonomous Grand Challenge Multi-view 3D Visual Grounding Track, validating its effectiveness and robustness.

State-of-the-art retrieval models typically address a straightforward search scenario, in which retrieval tasks are fixed (e.g., finding a passage to answer a specific question) and only a single modality is supported for both queries and retrieved results. This paper introduces techniques for advancing information retrieval with multimodal large language models (MLLMs), enabling a broader search scenario, termed universal multimodal retrieval, where multiple modalities and diverse retrieval tasks are accommodated. To this end, we first study fine-tuning an MLLM as a bi-encoder retriever on 10 datasets with 16 retrieval tasks. Our empirical results show that the fine-tuned MLLM retriever is capable of understanding challenging queries, composed of both text and image, but it underperforms compared to a smaller CLIP retriever in cross-modal retrieval tasks due to the modality bias exhibited by MLLMs. To address the issue, we propose modality-aware hard negative mining to mitigate the modality bias exhibited by MLLM retrievers. Second, we propose continuously fine-tuning the universal multimodal retriever to enhance its text retrieval capability while preserving multimodal retrieval capability. As a result, our model, MM-Embed, achieves state-of-the-art performance on the multimodal retrieval benchmark M-BEIR, which spans multiple domains and tasks, while also surpassing the state-of-the-art text retrieval model, NV-Embed-v1, on the MTEB retrieval benchmark. Finally, we explore prompting the off-the-shelf MLLMs as zero-shot rerankers to refine the ranking of the candidates from the multimodal retriever. We find that, through prompt-and-reranking, MLLMs can further improve multimodal retrieval when the user queries (e.g., text-image composed queries) are more complex and challenging to understand. These findings also pave the way for advancing universal multimodal retrieval in the future. We release the model weights at: https://huggingface.co/nvidia/MM-Embed.

Multimodal Large Language Models (MLLMs) have achieved remarkable success in vision understanding, reasoning, and interaction. However, the inference computation and memory increase progressively with the generation of output tokens during decoding, directly affecting the efficacy of MLLMs. Existing methods attempt to reduce the vision context redundancy to achieve efficient MLLMs. Unfortunately, the efficiency benefits of the vision context reduction in the prefill stage gradually diminish during the decoding stage. To address this problem, we proposed a dynamic vision-language context sparsification framework Dynamic-LLaVA, which dynamically reduces the redundancy of vision context in the prefill stage and decreases the memory and computation overhead of the generated language context during decoding. Dynamic-LLaVA designs a tailored sparsification inference scheme for different inference modes, i.e., prefill, decoding with and without KV cache, to achieve efficient inference of MLLMs. In practice, Dynamic-LLaVA can reduce computation consumption by $\sim$75\% in the prefill stage. Meanwhile, throughout the entire generation process of MLLMs, Dynamic-LLaVA reduces the $\sim$50\% computation consumption under decoding without KV cache, while saving $\sim$50\% GPU memory overhead when decoding with KV cache, due to the vision-language context sparsification. Extensive experiments also demonstrate that Dynamic-LLaVA achieves efficient inference for MLLMs with negligible understanding and generation ability degradation or even performance gains compared to the full-context inference baselines. Code is available at https://github.com/Osilly/dynamic_llava.