ICLR 2025 论文列表

We introduce Latent Action Pretraining for general Action models (LAPA), the first unsupervised method for pretraining Vision-Language-Action (VLA) models without ground-truth robot action labels. Existing Vision-Language-Action models require action labels typically collected by human teleoperators during pretraining, which significantly limits possible data sources and scale. In this work, we propose a method to learn from internet-scale videos that do not have robot action labels. We first train an action quantization model leveraging VQ-VAE-based objective to learn discrete latent actions between image frames, then pretrain a latent VLA model to predict these latent actions from observations and task descriptions, and finally finetune the VLA on small-scale robot manipulation data to map from latent to robot actions. Experimental results demonstrate that our method significantly outperforms existing techniques that train robot manipulation policies from large-scale videos. Furthermore, it outperforms the state-of-the-art VLA model trained with robotic action labels on real-world manipulation tasks that require language conditioning, generalization to unseen objects, and semantic generalization to unseen instructions. Training only on human manipulation videos also shows positive transfer, opening up the potential for leveraging web-scale data for robotics foundation models.

Pre-trained multi-modal Vision-Language Models like CLIP are widely used off-the-shelf for a variety of applications. In this paper, we show that the common practice of individually exploiting the text or image encoders of these powerful multi-modal models is highly suboptimal for intra-modal tasks like image-to-image retrieval. We argue that this is inherently due to the CLIP-style inter-modal contrastive loss that does not enforce any intra-modal constraints, leading to what we call intra-modal misalignment. To demonstrate this, we leverage two optimization-based modality inversion techniques that map representations from their input modality to the complementary one without any need for auxiliary data or additional trained adapters. We empirically show that, in the intra-modal tasks of image-to-image and text-to-text retrieval, approaching these tasks inter-modally significantly improves performance with respect to intra-modal baselines on more than fifteen datasets. Additionally, we demonstrate that approaching a native inter-modal task (e.g. zero-shot image classification) intra-modally decreases performance, further validating our findings. Finally, we show that incorporating an intra-modal term in the pre-training objective or narrowing the modality gap between the text and image feature embedding spaces helps reduce the intra-modal misalignment. The code is publicly available at: https://github.com/miccunifi/Cross-the-Gap.

The rapid advancements in Vision-Language Models (VLMs) have shown great potential in tackling mathematical reasoning tasks that involve visual context. Unlike humans who can reliably apply solution steps to similar problems with minor modifications, we found that state-of-the-art VLMs like GPT-4o can consistently fail in these scenarios, revealing limitations in their mathematical reasoning capabilities. In this paper, we investigate the **mathematical reasoning robustness** in VLMs and evaluate how well these models perform under different variants of the same question, such as changes in visual numerical values or function graphs. While several vision-based math benchmarks have been developed to assess VLMs' problem-solving capabilities, these benchmarks contain only static sets of problems and cannot easily evaluate mathematical reasoning robustness. To fill this gap, we introduce **DynaMath**, a dynamic visual math benchmark designed for in-depth assessment of VLMs. **DynaMath** includes 501 high-quality, multi-topic *seed* questions, *each represented as a Python program*. Those programs are carefully designed and annotated to enable the automatic generation of a much larger set of *concrete* questions, including many different types of visual and textual variations. **DynaMath** allows us to evaluate the generalization ability of VLMs, by assessing their performance under varying input conditions of a seed question. We evaluated 14 state-of-the-art VLMs with 5,010 generated concrete questions (10 per seed question). Our results show that the worst-case model accuracy, defined as the percentage of correctly answered seed questions in all 10 variants, is significantly lower than the average-case accuracy. In addition, many models show high consistency in answering these questions -- the incorrectness of a certain variant of a seed question is not only due to inherent randomness. Our analysis emphasizes the need to study the robustness of VLMs' reasoning abilities, and **DynaMath** provides valuable insights to guide the development of more reliable models for mathematical reasoning.

Large Language Models (LLMs) have demonstrated strong reasoning capabilities in solving complex problems. However, current approaches primarily enhance reasoning through the elaboration of thoughts while neglecting the diversity of reasoning types. LLMs typically employ deductive reasoning, proceeding step-by-step from given conditions, which limits their exploration during problem-solving. Our analysis reveals that certain problems are exclusively solvable through specific reasoning strategies like inductive, abductive, or analogical reasoning. However, incorporating diverse reasoning approaches presents two key challenges: identifying the appropriate reasoning type for each problem and exploiting this approach during problem-solving. Therefore, we propose the TypedThinker that predicts suitable reasoning types based on the problem and their previous effectiveness and provides relevant demonstrations to guide LLMs in applying these strategies. Experimental results show significant improvements across multiple benchmarks, with performance gains of 3.4\% for Mistral 7B, 6.5\% for LLaMA3 8B, and 7\% for Qwen 2 7B on logical and mathematical reasoning tasks. TypedThinker enhances LLM reasoning without requiring knowledge distillation from larger models. It can be integrated into more advanced systems like GPT-4o or specialized models like MetaMath to diversify their reasoning approaches and improve their problem-solving capabilities.

Vision-language models (VLMs) are known to be susceptible to distribution shifts between pre-training data and test data, and test-time adaptation (TTA) methods for VLMs have been proposed to mitigate the detrimental impact of the distribution shifts. However, the existing methods solely rely on the internal knowledge encoded within the model parameters, which are constrained to pre-training data. To complement the limitation of the internal knowledge, we propose **Retrieval-Augmented-TTA (RA-TTA)** for adapting VLMs to test distribution using **external** knowledge obtained from a web-scale image database. By fully exploiting the bi-modality of VLMs, RA-TTA **adaptively** retrieves proper external images for each test image to refine VLMs' predictions using the retrieved external images, where fine-grained **text descriptions** are leveraged to extend the granularity of external knowledge. Extensive experiments on 17 datasets demonstrate that the proposed RA-TTA outperforms the state-of-the-art methods by 3.01-9.63\% on average.

Existing multimodal retrieval benchmarks primarily focus on evaluating whether models can retrieve and utilize external textual knowledge for question answering. However, there are scenarios where retrieving visual information is either more beneficial or easier to access than textual data. In this paper, we introduce a multimodal retrieval-augmented generation benchmark, MRAG-Bench, in which we systematically identify and categorize scenarios where visually augmented knowledge is better than textual knowledge, for instance, more images from varying viewpoints. MRAG-Bench consists of 16,130 images and 1,353 human-annotated multiple-choice questions across 9 distinct scenarios. With MRAG-Bench, we conduct an evaluation of 10 open-source and 4 proprietary large vision-language models (LVLMs). Our results show that all LVLMs exhibit greater improvements when augmented with images compared to textual knowledge, confirming that MRAG-Bench is vision-centric. Additionally, we conduct extensive analysis with MRAG-Bench, which offers valuable insights into retrieval-augmented LVLMs. Notably, the top-performing model, GPT-4o, faces challenges in effectively leveraging retrieved knowledge, achieving only a 5.82\% improvement with ground-truth information, in contrast to a 33.16\% improvement observed in human participants. These findings highlight the importance of MRAG-Bench in encouraging the community to enhance LVLMs' ability to utilize retrieved visual knowledge more effectively.

Multimodal models or Vision Language Models (VLMs) have reshaped the paradigm in machine learning, offering zero-shot capabilities that require no additional training when adapted to new classification tasks. However, despite their advancements, spurious correlations still exist in VLMs. Existing approaches to tackle this issue often require target label annotations, contradicting the principle of zero-shot classification, or they primarily focus on a single modality, risking misalignment between text and image modalities. Others rely on extensive domain knowledge or large language models (LLMs) to characterize spurious features, making the performance sensitive to the generated prompts and undermining zero-shot capability. In response, we propose a new solution that tackles spurious correlations in VLMs within the zero-shot setting. Our approach utilizes a translation operation that preserves the latent space distribution to address issues of spurious correlations. In particular, our method is grounded in and inspired by a theoretical analysis, which identifies that the optimal translation directions are along the spurious vector. As VLMs unify two modalities, we compute spurious vectors from the text prompts and guide the translation for image embeddings, aligning the requirements for the fusion of different modalities in VLMs. We conducted experiments on benchmark datasets, which have shown significant improvements in worst-group accuracy. Additionally, our visualizations of VLMs further demonstrate the effectiveness of this intervention.

We study reward models for long-horizon manipulation by learning from action-free videos and language instructions, which we term the visual-instruction correlation (VIC) problem. Existing VIC methods face challenges in learning rewards for long-horizon tasks due to their lack of sub-stage awareness, difficulty in modeling task complexities, and inadequate object state estimation. To address these challenges, we introduce VICtoR, a novel hierarchical VIC reward model capable of providing effective reward signals for long-horizon manipulation tasks. Trained solely on primitive motion demonstrations, VICtoR effectively provides precise reward signals for long-horizon tasks by assessing task progress at various stages using a novel stage detector and motion progress evaluator. We conducted extensive experiments in both simulated and real-world datasets. The results suggest that VICtoR outperformed the best existing methods, achieving a 43% improvement in success rates for long-horizon tasks. Our project page can be found at https://cmlab-victor.github.io/cmlab-vicotor.github.io/.

Large Language Models (LLMs) have been successful in mathematical reasoning tasks such as formal theorem proving when integrated with interactive proof assistants like Lean. Existing approaches involve training or fine-tuning an LLM on a specific dataset to perform well on particular domains, such as undergraduate-level mathematics. These methods struggle with generalizability to advanced mathematics. A fundamental limitation is that these approaches operate on static domains, failing to capture how mathematicians often work across multiple domains and projects simultaneously or cyclically. We present LeanAgent, a novel lifelong learning framework for formal theorem proving that continuously generalizes to and improves on ever-expanding mathematical knowledge without forgetting previously learned knowledge. LeanAgent introduces several key innovations, including a curriculum learning strategy that optimizes the learning trajectory in terms of mathematical difficulty, a dynamic database for efficient management of evolving mathematical knowledge, and progressive training to balance stability and plasticity. LeanAgent successfully generates formal proofs for 155 theorems across 23 diverse Lean repositories where formal proofs were previously missing, many from advanced mathematics. It performs significantly better than the static LLM baseline, proving challenging theorems in domains like abstract algebra and algebraic topology while showcasing a clear progression of learning from basic concepts to advanced topics. In addition, we analyze LeanAgent's superior performance on key lifelong learning metrics. LeanAgent achieves exceptional scores in stability and backward transfer, where learning new tasks improves performance on previously learned tasks. This emphasizes LeanAgent's continuous generalizability and improvement, explaining its superior theorem-proving performance.

When learning an input-output mapping from very few examples, is it better to first infer a latent function that explains the examples, or is it better to directly predict new test outputs, e.g. using a neural network? We study this question on ARC by training neural models for \emph{induction} (inferring latent functions) and \emph{transduction} (directly predicting the test output for a given test input). We train on synthetically generated variations of Python programs that solve ARC training tasks. We find inductive and transductive models solve different kinds of test problems, despite having the same training problems and sharing the same neural architecture: Inductive program synthesis excels at precise computations, and at composing multiple concepts, while transduction succeeds on fuzzier perceptual concepts. Ensembling them approaches human-level performance on ARC.

The Segment Anything model (SAM) has shown a generalized ability to group image pixels into patches, but applying it to semantic-aware segmentation still faces major challenges. This paper presents SAM-CP, a simple approach that establishes two types of composable prompts beyond SAM and composes them for versatile segmentation. Specifically, given a set of classes (in texts) and a set of SAM patches, the Type-I prompt judges whether a SAM patch aligns with a text label, and the Type-II prompt judges whether two SAM patches with the same text label also belong to the same instance. To decrease the complexity in dealing with a large number of semantic classes and patches, we establish a unified framework that calculates the affinity between (semantic and instance) queries and SAM patches, and then merges patches with high affinity to the query. Experiments show that SAM-CP achieves semantic, instance, and panoptic segmentation in both open and closed domains. In particular, it achieves state-of-the-art performance in open-vocabulary segmentation. Our research offers a novel and generalized methodology for equipping vision foundation models like SAM with multi-grained semantic perception abilities. Codes are released on https://github.com/ucas-vg/SAM-CP.

Large Multimodal Models (LMMs) such as LLaVA have shown strong performance in visual-linguistic reasoning. These models first embed images into a fixed large number of visual tokens and then feed them into a Large Language Model (LLM). However, this design causes an excessive number of tokens for dense visual scenarios such as high-resolution images and videos, leading to great inefficiency. While token pruning/merging methods do exist, they produce a single length output for each image and do not afford flexibility in trading off information density v.s. efficiency. Inspired by the concept of Matryoshka Dolls, we propose : Matryoshka Multimodal Models, which learns to represent visual content as nested sets of visual tokens that capture information across multiple coarse-to-fine granularities. Our approach offers several unique benefits for LMMs: (1) One can explicitly control the visual granularity per test instance during inference, e.g. , adjusting the number of tokens used to represent an image based on the anticipated complexity or simplicity of the content; (2) provides a framework for analyzing the granularity needed for existing datasets, where we find that COCO-style benchmarks only need around 9 visual tokens to obtain accuracy similar to that of using all 576 tokens; (3) Our approach provides a foundation to explore the best trade-off between performance and visual token length at sample level, where our investigation reveals that a large gap exists between the oracle upper bound and current fixed-scale representations.

Diffusion models for Text-to-Image (T2I) conditional generation have recently achieved tremendous success. Yet, aligning these models with user’s intentions still involves a laborious trial-and-error process, and this challenging alignment problem has attracted considerable attention from the research community. In this work, instead of relying on fine-grained linguistic analyses of prompts, human annotation, or auxiliary vision-language models, we use Mutual Information (MI) to guide model alignment. In brief, our method uses self-supervised fine-tuning and relies on a point-wise MI estimation between prompts and images to create a synthetic fine-tuning set for improving model alignment. Our analysis indicates that our method is superior to the state-of-the-art, yet it only requires the pre-trained denoising network of the T2I model itself to estimate MI, and a simple fine-tuning strategy that improves alignment while maintaining image quality. Code available at https://github.com/Chao0511/mitune.

Understanding the structure and function of circuits is crucial for electronic design automation (EDA). Circuits can be formulated as And-Inverter graphs (AIGs), enabling efficient implementation of representation learning through graph neural networks (GNNs). Masked modeling paradigms have been proven effective in graph representation learning. However, masking augmentation to original circuits will destroy their logical equivalence, which is unsuitable for circuit representation learning. Moreover, existing masked modeling paradigms often prioritize structural information at the expense of abstract information such as circuit function. To address these limitations, we introduce MGVGA, a novel constrained masked modeling paradigm incorporating masked gate modeling (MGM) and Verilog-AIG alignment (VGA). Specifically, MGM preserves logical equivalence by masking gates in the latent space rather than in the original circuits, subsequently reconstructing the attributes of these masked gates. Meanwhile, large language models (LLMs) have demonstrated an excellent understanding of the Verilog code functionality. Building upon this capability, VGA performs masking operations on original circuits and reconstructs masked gates under the constraints of equivalent Verilog codes, enabling GNNs to learn circuit functions from LLMs. We evaluate MGVGA on various logic synthesis tasks for EDA and show the superior performance of MGVGA compared to previous state-of-the-art methods. Our code is available at https://github.com/wuhy68/MGVGA.

Rating-based human evaluation has become an essential tool to accurately evaluate the impressive performance of large language models (LLMs). However, current rating systems suffer from several important limitations: first, they fail to account for biases that significantly influence evaluation results, second, they require large and expensive preference datasets to obtain accurate ratings, and third, they do not facilitate meaningful comparisons of model ratings across different tasks. To address these issues, we introduce Polyrating, an expressive and flexible rating system based on maximum a posteriori estimation that enables a more nuanced and thorough analysis of model performance at lower costs. Polyrating can detect and quantify biases affecting human preferences, ensuring fairer model comparisons. Further, Polyrating can reduce the cost of human evaluations by up to $41$% for new models and up to $77$% for new tasks by leveraging existing benchmark scores. Lastly, Polyrating enables direct comparisons of ratings across different tasks, providing a comprehensive understanding of an LLMs' strengths, weaknesses, and relative performance across different applications.

The advent of real-time large multimodal models (LMMs) like GPT-4o has sparked considerable interest in efficient LMMs. LMM frameworks typically encode visual inputs into vision tokens (continuous representations) and integrate them and textual instructions into the context of large language models (LLMs), where large-scale parameters and numerous context tokens (predominantly vision tokens) result in substantial computational overhead. Previous efforts towards efficient LMMs always focus on replacing the LLM backbone with smaller models, while neglecting the crucial issue of token quantity. In this paper, we introduce LLaVA-Mini, an efficient LMM with minimal vision tokens. To achieve a high compression ratio of vision tokens while preserving visual information, we first analyze how LMMs understand vision tokens and find that most vision tokens only play a crucial role in the early layers of LLM backbone, where they mainly fuse visual information into text tokens. Building on this finding, LLaVA-Mini introduces modality pre-fusion to fuse visual information into text tokens in advance, thereby facilitating the extreme compression of vision tokens fed to LLM backbone into one token. LLaVA-Mini is a unified large multimodal model that can support the understanding of images, high-resolution images, and videos in an efficient manner. Experiments across 11 image-based and 7 video-based benchmarks demonstrate that LLaVA-Mini outperforms LLaVA-v1.5 with just 1 vision token instead of 576. Efficiency analyses reveal that LLaVA-Mini can reduce FLOPs by 77%, deliver low-latency responses within 40 milliseconds, and process over 10,000 frames of video on the GPU hardware with 24GB of memory.

Video Multimodal Large Language Models~(MLLMs) have shown remarkable capability of understanding the video semantics on various downstream tasks. Despite the advancements, there is still a lack of systematic research on visual context representation, which refers to the scheme to select frames from a video and further select the tokens from a frame. In this paper, we explore the design space for visual context representation, and aim to improve the performance of video MLLMs by finding more effective representation schemes. Firstly, we formulate the task of visual context representation as a constrained optimization problem, and model the language modeling loss as a function of the number of frames and the number of embeddings (or tokens) per frame, given the maximum visual context window size. Then, we explore the scaling effects in frame selection and token selection respectively, and fit the corresponding function curve by conducting extensive empirical experiments. We examine the effectiveness of typical selection strategies and present empirical findings to determine the two factors. Furthermore, we study the joint effect of frame selection and token selection, and derive the optimal formula for determining the two factors. We demonstrate that the derived optimal settings show alignment with the best-performed results of empirical experiments. The data and code are available at: https://github.com/RUCAIBox/Opt-Visor.

Visual perspective taking (VPT) is the ability to perceive and reason about the perspectives of others. It is an essential feature of human intelligence, which develops over the first decade of life and requires an ability to process the 3D structure of visual scenes. A growing number of reports have indicated that deep neural networks (DNNs) become capable of analyzing 3D scenes after training on large image datasets. We investigated if this emergent ability for 3D analysis in DNNs is sufficient for VPT with the 3D perception challenge (3D-PC): a novel benchmark for 3D perception in humans and DNNs. The 3D-PC is comprised of three 3D-analysis tasks posed within natural scene images: (i.) a simple test of object depth order, (ii.) a basic VPT task (VPT-basic), and (iii.) a more challenging version of VPT (VPT-perturb) designed to limit the effectiveness of "shortcut" visual strategies. We tested human participants (N=33) and linearly probed or text-prompted over 300 DNNs on the challenge and found that nearly all of the DNNs approached or exceeded human accuracy in analyzing object depth order. Surprisingly, DNN accuracy on this task correlated with their object recognition performance. In contrast, there was an extraordinary gap between DNNs and humans on VPT-basic. Humans were nearly perfect, whereas most DNNs were near chance. Fine-tuning DNNs on VPT-basic brought them close to human performance, but they, unlike humans, dropped back to chance when tested on VPT-perturb. Our challenge demonstrates that the training routines and architectures of today's DNNs are well-suited for learning basic 3D properties of scenes and objects but are ill-suited for reasoning about these properties like humans do. We release our 3D-PC datasets and code to help bridge this gap in 3D perception between humans and machines.

Learned image compression (LIC) has demonstrated superior rate-distortion (R-D) performance compared to traditional codecs, but is challenged by training inefficiency that could incur more than two weeks to train a state-of-the-art model from scratch. Existing LIC methods overlook the slow convergence caused by compacting energy in learning nonlinear transforms. In this paper, we first reveal that such energy compaction consists of two components, \emph{i.e.}, feature decorrelation and uneven energy modulation. On such basis, we propose a linear auxiliary transform (AuxT) to disentangle energy compaction in training nonlinear transforms. The proposed AuxT obtains coarse approximation to achieve efficient energy compaction such that distribution fitting with the nonlinear transforms can be simplified to fine details. We then develop wavelet-based linear shortcuts (WLSs) for AuxT that leverages wavelet-based downsampling and orthogonal linear projection for feature decorrelation and subband-aware scaling for uneven energy modulation. AuxT is lightweight and plug-and-play to be integrated into diverse LIC models to address the slow convergence issue. Experimental results demonstrate that the proposed approach can accelerate training of LIC models by 2 times and simultaneously achieves an average 1\% BD-rate reduction. To our best knowledge, this is one of the first successful attempt that can significantly improve the convergence of LIC with comparable or superior rate-distortion performance.