ICLR 2025 Papers

*Visual reprogramming* (VR) reuses pre-trained vision models for downstream image classification tasks by adding trainable noise patterns to inputs. When applied to vision-language models (e.g., CLIP), existing VR approaches follow the same pipeline used in vision models (e.g., ResNet, ViT), where ground-truth class labels are inserted into fixed text templates to guide the optimization of VR patterns. This label-based approach, however, overlooks the rich information and diverse attribute-guided textual representations that CLIP can exploit, which may lead to the misclassification of samples. In this paper, we propose ***Attr**ibute-based **V**isual **R**eprogramming* (AttrVR) for CLIP, utilizing ***des**criptive **attr**ibutes* (DesAttrs) and ***dist**inctive **attr**ibutes* (DistAttrs), which respectively represent common and unique feature descriptions for different classes. Besides, as images of the same class may reflect different attributes after VR, AttrVR iteratively refines patterns using the $k$-nearest DesAttrs and DistAttrs for each image sample, enabling more dynamic and sample-specific optimization. Theoretically, AttrVR is shown to reduce intra-class variance and increase inter-class separation. Empirically, it achieves superior performance in 12 downstream tasks for both ViT-based and ResNet-based CLIP. The success of AttrVR facilitates more effective integration of VR from unimodal vision models into vision-language models. Our code is available at https://github.com/tmlr-group/AttrVR.

As trajectories sampled by policies used by reinforcement learning (RL) and generative flow networks (GFlowNets) grow longer, credit assignment and exploration become more challenging, and the long planning horizon hinders mode discovery and generalization. The challenge is particularly pronounced in entropy-seeking RL methods, such as generative flow networks, where the agent must learn to sample from a structured distribution and discover multiple high-reward states, each of which take many steps to reach. To tackle this challenge, we propose an approach to incorporate the discovery of action abstractions, or high-level actions, into the policy optimization process. Our approach involves iteratively extracting action subsequences commonly used across many high-reward trajectories and `chunking' them into a single action that is added to the action space. In empirical evaluation on synthetic and real-world environments, our approach demonstrates improved sample efficiency performance in discovering diverse high-reward objects, especially on harder exploration problems. We also observe that the abstracted high-order actions are potentially interpretable, capturing the latent structure of the reward landscape of the action space. This work provides a cognitively motivated approach to action abstraction in RL and is the first demonstration of hierarchical planning in amortized sequential sampling.

Model merging has emerged as a powerful technique for combining task-specific vision models into a unified and multi-functional model. Previous methods represented by task arithmetic, have demonstrated effectiveness and scalability in this domain. When large vision-language models (LVLMs) arise with model size scaling up, this design becomes challenging to fuse different instruction-tuned LVLMs for generalization enhancement. The large scale and multi-modal nature of LVLMs present unique obstacles, including constructing reusable and modular components to accommodate the multi-component architecture of LVLMs and the requirement for dynamic fusion based on multi-modal input tokens. To address these challenges, we propose the \textbf{RE}cipe \textbf{ME}rging \textbf{DY}namics (REMEDY) method, a scalable and flexible paradigm for model merging in LVLMs. We first define reusable modules termed \textit{recipes} including the projector and shallow LLM layers, enhancing visual-language understanding. Then, we introduce a modality-aware allocator dynamically generates weights in a one-shot manner based on input relevance to existing recipes, enabling efficient cross-modal knowledge integration. REMEDY thus offers an adaptive solution for LVLMs to tackle both seen (i.e., multi-task learning) and unseen (i.e., zero-shot generalization) tasks. Experimental results demonstrate that our method consistently improves performance on both seen and unseen tasks, underscoring the effectiveness of REMEDY in diverse multi-modal scenarios.

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.

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.

In autonomous driving, end-to-end planners directly utilize raw sensor data, enabling them to extract richer scene features and reduce information loss compared to traditional planners. This raises a crucial research question: how can we develop better scene feature representations to fully leverage sensor data in end-to-end driving? Self-supervised learning methods show great success in learning rich feature representations in NLP and computer vision. Inspired by this, we propose a novel self-supervised learning approach using the LAtent World model (LAW) for end-to-end driving. LAW predicts future latent scene features based on current features and ego trajectories. This self-supervised task can be seamlessly integrated into perception-free and perception-based frameworks, improving scene feature learning while optimizing trajectory prediction. LAW achieves state-of-the-art performance across multiple benchmarks, including real-world open-loop benchmark nuScenes, NAVSIM, and simulator-based closed-loop benchmark CARLA. The code will be released.

Action anticipation models require an understanding of temporal action patterns and dependencies to predict future actions from previous events. The key challenges arise from the vast number of possible action sequences, given the flexibility in action ordering and the interleaving of multiple goals. Since only a subset of such action sequences are present in action anticipation datasets, there is an inherent ordering bias in them. Another challenge is the presence of noisy input to the models due to erroneous action recognition or other upstream tasks. This paper addresses these challenges by introducing a novel data augmentation strategy that separately augments observed action sequences and next actions. To address biased action ordering, we introduce a grammar induction algorithm that derives a powerful context-free grammar from action sequence data. We also develop an efficient parser to generate plausible next-action candidates beyond the ground truth. For noisy input, we enhance model robustness by randomly deleting or replacing actions in observed sequences. Our experiments on the 50Salads, EGTEA Gaze+, and Epic-Kitchens-100 datasets demonstrate significant performance improvements over existing state-of-the-art methods.

Large Vision Language Models (LVLMs) have demonstrated remarkable abilities in understanding and reasoning about both visual and textual information. However, existing evaluation methods for LVLMs, primarily based on benchmarks like Visual Question Answering and image captioning, often fail to capture the full scope of LVLMs' capabilities. These benchmarks are limited by issues such as inadequate assessment of detailed visual perception, data contamination, and a lack of focus on multi-turn reasoning. To address these challenges, we propose LVLM-Playground, a game-based evaluation framework designed to provide a comprehensive assessment of LVLMs' cognitive and reasoning skills in structured environments. LVLM-Playground uses a set of games to evaluate LVLMs on four core tasks: Perceiving, Question Answering, Rule Following, and End-to-End Playing, with each target task designed to assess specific abilities, including visual perception, reasoning, decision-making, etc. Based on this framework, we conduct extensive experiments that explore the limitations of current LVLMs, such as handling long structured outputs and perceiving detailed and dense elements. Code and data are publicly available at https://github.com/xinke-wang/LVLM-Playground.

Although large vision-language-action (VLA) models pretrained on extensive robot datasets offer promising generalist policies for robotic learning, they still struggle with spatial-temporal dynamics in interactive robotics, making them less effective in handling complex tasks, such as manipulation. In this work, we introduce visual trace prompting, a simple yet effective approach to facilitate VLA models’ spatial-temporal awareness for action prediction by encoding state-action trajectories visually. We develop a new TraceVLA model by finetuning OpenVLA on our own collected dataset of 150K robot manipulation trajectories using visual trace prompting. Evaluations of TraceVLA across 137 configurations in SimplerEnv and 4 tasks on a physical WidowX robot demonstrate state-of-the-art performance, outperforming OpenVLA by 10% on SimplerEnv and 3.5x on real-robot tasks and exhibiting robust generalization across diverse embodiments and scenarios. To further validate the effectiveness and generality of our method, we present a compact VLA model based on 4B Phi-3-Vision, pretrained on the Open-X-Embodiment and finetuned on our dataset, rivals the 7B OpenVLA baseline while significantly improving inference efficiency.

Although reward models have been successful in improving multimodal large language models, the reward models themselves remain brutal and contain minimal information. Notably, existing reward models only mimic human annotations by assigning only one feedback to any text, no matter how long the text is. In the realm of multimodal language models, where models are required to process both images and texts, a naive reward model may learn implicit biases toward texts and become less grounded in images. In this paper, we propose a **T**oken-**L**evel **D**etective **R**eward Model (**TLDR**) to provide fine-grained annotations to each text token. We first introduce a perturbation-based method to generate synthetic hard negatives and their token-level labels to train TLDR models. Then we show the rich usefulness of TLDR models both in assisting off-the-shelf models to self-correct their generations, and in serving as a hallucination evaluation tool. We show that TLDR automatically trains a token-level likelihood optimization, and can improve the base model's performance significantly. Finally, we show that TLDR models can significantly speed up human annotation by 3 times to acquire a broader range of high-quality vision language data.

We present Polynomial Attention Drop-in Replacement (PADRe), a novel and unifying framework designed to replace the conventional self-attention mechanism in transformer models. Notably, several recent alternative attention mechanisms, including Hyena, Mamba, SimA, Conv2Former, and Castling-ViT, can be viewed as specific instances of our PADRe framework. PADRe leverages polynomial functions and draws upon established results from approximation theory, enhancing computational efficiency without compromising accuracy. PADRe's key components include multiplicative nonlinearities, which we implement using straightforward, hardware-friendly operations such as Hadamard products, incurring only linear computational and memory costs. PADRe further avoids the need for using complex functions such as Softmax, yet it maintains comparable or superior accuracy compared to traditional self-attention. We assess the effectiveness of PADRe as a drop-in replacement for self-attention across diverse computer vision tasks. These tasks include image classification, image-based 2D object detection, and 3D point cloud object detection. Empirical results demonstrate that PADRe runs significantly faster than the conventional self-attention (11x~43x faster on server GPU and mobile NPU) while maintaining similar accuracy when substituting self-attention in the transformer models.

Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities across multimodal tasks such as visual perception and reasoning, leading to good performance on various multimodal evaluation benchmarks. However, these benchmarks keep a static nature and overlap with the pre-training data, resulting in fixed complexity constraints and data contamination issues. This raises the concern regarding the validity of the evaluation. To address these two challenges, we introduce a dynamic multimodal evaluation protocol called Vision-Language Bootstrapping (VLB). VLB provides a robust and comprehensive assessment for LVLMs with reduced data contamination and flexible complexity. To this end, VLB dynamically generates new visual question-answering samples through a multimodal bootstrapping module that modifies both images and language, while ensuring that newly generated samples remain consistent with the original ones by a judge module. By composing various bootstrapping strategies, VLB offers dynamic variants of existing benchmarks with diverse complexities, enabling the evaluation to co-evolve with the ever-evolving capabilities of LVLMs. Extensive experimental results across multiple benchmarks, including SEEDBench, MMBench, and MME, show that VLB significantly reduces data contamination and exposes performance limitations of LVLMs.

The Mixture-of-Experts (MoE) has gained increasing attention in studying Large Vision-Language Models (LVLMs). It uses a sparse model to replace the dense model, achieving comparable performance while activating fewer parameters during inference, thus significantly reducing the inference cost. Existing MoE methods in LVLM encourage different experts to specialize in different tokens, and they usually employ a router to predict the routing of each token. However, the router is not optimized concerning distinct parameter optimization directions generated from tokens within an expert. This may lead to severe interference between tokens within an expert. To address this problem, we propose to use the token-level gradient analysis to Solving Token Gradient Conflict (STGC) in this paper. Specifically, we first use token-level gradients to identify conflicting tokens in experts. After that, we add a regularization loss tailored to encourage conflicting tokens routing from their current experts to other experts, for reducing interference between tokens within an expert. Our method can serve as a plug-in for diverse LVLM methods, and extensive experimental results demonstrate its effectiveness. demonstrate its effectiveness. The code will be publicly available at https://github.com/longrongyang/STGC.

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.

Modeling human-like action-to-reaction generation has significant real-world applications, like human-robot interaction and games. Despite recent advancements in single-person motion generation, it is still challenging to well handle action-to-reaction generation, due to the difficulty of directly predicting reaction from action sequence without prompts, and the absence of a unified representation that effectively encodes multi-person motion. To address these challenges, we introduce Think-Then-React (TTR), a large language-model-based framework designed to generate human-like reactions. First, with our fine-grained multimodal training strategy, TTR is capable to unify two processes during inference: a thinking process that explicitly infers action intentions and reasons corresponding reaction description, which serve as semantic prompts, and a reacting process that predicts reactions based on input action and the inferred semantic prompts. Second, to effectively represent multi-person motion in language models, we propose a unified motion tokenizer by decoupling egocentric pose and absolute space features, which effectively represents action and reaction motion with same encoding. Extensive experiments demonstrate that TTR outperforms existing baselines, achieving significant improvements in evaluation metrics, such as reducing FID from 3.988 to 1.942.

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.

We revisit the ``dataset classification'' experiment suggested by Torralba & Efros (2011) a decade ago, in the new era with large-scale, diverse, and hopefully less biased datasets as well as more capable neural network architectures. Surprisingly, we observe that modern neural networks can achieve excellent accuracy in classifying which dataset an image is from: e.g., we report 84.7% accuracy on held-out validation data for the three-way classification problem consisting of the YFCC, CC, and DataComp datasets. Our further experiments show that such a dataset classifier could learn semantic features that are generalizable and transferable, which cannot be explained by memorization. We hope our discovery will inspire the community to rethink issues involving dataset bias.

Contrastive language-image pretraining (CLIP) has significantly advanced image-based vision learning. A pressing topic subsequently arises: how can we effectively adapt CLIP to the video domain? Recent studies have focused on adjusting either the textual or visual branch of CLIP for action recognition. However, we argue that adaptations of both branches are crucial. In this paper, we propose a **C**ontrastive **L**anguage-**A**ction **V**ideo Learn**er** (**CLAVER**), designed to shift CLIP's focus from the alignment of static visual objects and concrete nouns to the alignment of dynamic action behaviors and abstract verbs. Specifically, we introduce a novel Kronecker mask attention for temporal modeling. Our tailored Kronecker mask offers three benefits 1) it expands the temporal receptive field for each token, 2) it serves as an effective spatiotemporal heterogeneity inductive bias, mitigating the issue of spatiotemporal homogenization, and 3) it can be seamlessly plugged into transformer-based models. Regarding the textual branch, we leverage large language models to generate diverse, sentence-level and semantically rich interpretive prompts of actions, which shift the model's focus towards the verb comprehension. Extensive experiments on various benchmarks and learning scenarios demonstrate the superiority and generality of our approach. The code will be available soon.

Vision Language Models (VLMs) have become essential backbones for multi-modal intelligence, yet significant safety challenges limit their real-world application. While textual inputs can often be effectively safeguarded, adversarial visual inputs can often easily bypass VLM defense mechanisms. Existing defense methods are either resource-intensive, requiring substantial data and compute, or fail to simultaneously ensure safety and usefulness in responses. To address these limitations, we propose a novel two-phase inference-time alignment framework, **E**valuating **T**hen **A**ligning (ETA): i) Evaluating input visual contents and output responses to establish a robust safety awareness in multimodal settings, and ii) Aligning unsafe behaviors at both shallow and deep levels by conditioning the VLMs' generative distribution with an interference prefix and performing sentence-level best-of-$N$ to search the most harmless and helpful generation paths. Extensive experiments show that ETA outperforms baseline methods in terms of harmlessness, helpfulness, and efficiency, reducing the unsafe rate by 87.5\% in cross-modality attacks and achieving 96.6\% win-ties in GPT-4 helpfulness evaluation.