ICML 2025 Papers

Modern methods for fine-tuning a Vision Transformer (ViT) like Low-Rank Adaptation (LoRA) and its variants demonstrate impressive performance. However, these methods ignore the high-dimensional nature of Multi-Head Attention (MHA) weight tensors. To address this limitation, we propose Canonical Rank Adaptation (CaRA). CaRA leverages tensor mathematics, first by tensorising the transformer into two different tensors; one for projection layers in MHA and the other for feed-forward layers. Second, the tensorised formulation is fine-tuned using the low-rank adaptation in Canonical-Polyadic Decomposition (CPD) form. Employing CaRA efficiently minimizes the number of trainable parameters. Experimentally, CaRA outperforms existing Parameter-Efficient Fine-Tuning (PEFT) methods in visual classification benchmarks such as Visual Task Adaptation Benchmark (VTAB)-1k and Fine-Grained Visual Categorization (FGVC).

3D multiple object tracking (MOT) plays a crucial role in autonomous driving perception. Recent end-to-end query-based trackers simultaneously detect and track objects, which have shown promising potential for the 3D MOT task. However, existing methods are still in the early stages of development and lack systematic improvements, failing to track objects in certain complex scenarios, like occlusions and the small size of target object’s situations. In this paper, we first summarize the current end-to-end 3D MOT framework by decomposing it into three constituent parts: query initialization, query propagation, and query matching. Then we propose corresponding improvements, which lead to a strong yet simple tracker: S2-Track. Specifically, for query initialization, we present 2D-Prompted Query Initialization, which leverages predicted 2D object and depth information to prompt an initial estimate of the object’s 3D location. For query propagation, we introduce an Uncertainty-aware Probabilistic Decoder to capture the uncertainty of complex environment in object prediction with probabilistic attention. For query matching, we propose a Hierarchical Query Denoising strategy to enhance training robustness and convergence. As a result, our S2-Track achieves state-of-the-art performance on nuScenes benchmark, i.e., 66.3% AMOTA on test split, surpassing the previous best end-to-end solution by a significant margin of 8.9% AMOTA. We achieve 1st place on the nuScenes tracking task leaderboard.

In this paper, we present a novel approach, termed Double-Filter,to “slim down” the fine-tuning process of vision-language pre-trained (VLP) models via filtering redundancies in feature inputs and architectural components. We enhance the fine-tuning process using two approaches. First, we develop a new patch selection method incorporating image patch filtering through background and foreground separation, followed by a refined patch selection process. Second, we design a genetic algorithm to eliminate redundant fine-grained architecture layers, improving the efficiency and effectiveness of the model. The former makes patch selection semantics more comprehensive, improving inference efficiency while ensuring semantic representation. The latter’s fine-grained layer filter removes architectural redundancy to the extent possible and mitigates the impact on performance. Experimental results demonstrate that the proposed Double-Filter achieves superior efficiency of model fine-tuning and maintains competitive performance compared with the advanced efficient fine-tuning methods on three downstream tasks, VQA, NLVR and Retrieval. In addition, it has been proven to be effective under METER and ViLT VLP models.

Recent RL research has utilized reward shaping--particularly complex shaping rewards such as intrinsic motivation (IM)--to encourage agent exploration in sparse-reward environments. While often effective, ``reward hacking'' can lead to the shaping reward being optimized at the expense of the extrinsic reward, resulting in a suboptimal policy. Potential-Based Reward Shaping (PBRS) techniques such as Generalized Reward Matching (GRM) and Policy-Invariant Explicit Shaping (PIES) have mitigated this. These methods allow for implementing IM without altering optimal policies. In this work we show that they are effectively unsuitable for complex, exploration-heavy environments with long-duration episodes. To remedy this, we introduce Action-Dependent Optimality Preserving Shaping (ADOPS), a method of converting intrinsic rewards to an optimality-preserving form that allows agents to utilize IM more effectively in the extremely sparse environment of Montezuma's Revenge. We also prove ADOPS accommodates reward shaping functions that cannot be written in a potential-based form: while PBRS-based methods require the cumulative discounted intrinsic return be independent of actions, ADOPS allows for intrinsic cumulative returns to be dependent on agents' actions while still preserving the optimal policy set. We show how action-dependence enables ADOPS's to preserve optimality while learning in complex, sparse-reward environments where other methods struggle.

Vision GNN (ViG) demonstrates superior performance by representing images as graph structures, providing a more natural way to capture irregular semantic patterns beyond traditional grid or sequence-based representations. To efficiently adapt ViG to downstream tasks, parameter-efficient fine-tuning techniques like visual prompting become increasingly essential. However, existing prompting methods are primarily designed for Transformer-based models, neglecting the rich topological relationships among nodes and edges in graph-based representations, limiting their capacity to model complex semantics. In this paper, we propose Vision Graph Prompting (VGP), a novel framework tailored for vision graph structures. Our core insight reveals that semantically connected components in the graph exhibit low-rank properties. Building on this observation, we introduce a semantic low-rank prompting method that decomposes low-rank semantic features and integrates them with prompts on vision graph topologies, capturing both global structural patterns and fine-grained semantic dependencies. Extensive experiments demonstrate our method significantly improves ViG’s transfer performance on diverse downstream tasks, achieving results comparable to full fine-tuning while maintaining parameter efficiency.

Motion transfer is to transfer pose in driving video to object of source image, so that object of source image moves. Although great progress has been made recently in unsupervised motion transfer, many unsupervised methods still struggle to accurately model large displacement motions when large motion differences occur between source and driving images. To solve the problem, we propose an unsupervised anytime interpolation based large displacement motion transfer method, which can generate a series of anytime interpolated images between source and driving images. By decomposing large displacement motion into many small displacement motions, difficulty of large displacement motion estimation is reduced. In the process, we design a selector that can select optimal interpolated image from generated interpolated images for downstream tasks. Since there are no real images as labels in the interpolation process, we propose a bidirectional training strategy. Some constraints are added to optimal interpolated image to generate a reasonable interpolated image. To encourage network to generate high-quality images, a pre-trained Vision Transformer model is used to design constraint losses. Finally, experiments show that compared with the large displacement motion between source and driving images, small displacement motion between interpolated and driving images is easier to realize motion transfer. Compared with existing state-of-art methods, our method has significant improvements in motion-related metrics.

Vision-language instruction-tuning models have recently achieved significant performance improvements. In this work, we discover that large-scale 3D parallel training on those models leads to an imbalanced computation load across different devices. The vision and language parts are inherently heterogeneous: their data distribution and model architecture differ significantly, which affects distributed training efficiency. To address this issue, we rebalance the computational load from data, model, and memory perspectives, achieving more balanced computation across devices. Specifically, for the data, instances are grouped into new balanced mini-batches within and across devices. A search-based method is employed for the model to achieve a more balanced partitioning. For memory optimization, we adaptively adjust the re-computation strategy for each partition to utilize the available memory fully. These three perspectives are not independent but are closely connected, forming an omniverse balanced training framework. Extensive experiments are conducted to validate the effectiveness of our method. Compared with the open-source training code of InternVL-Chat, training time is reduced greatly, achieving about 1.8$\times$ speed-up. Our method's efficacy and generalizability are further validated across various models and datasets. Codes will be released at https://github.com/ModelTC/OmniBal.

Actor-critic algorithms have been instrumental in boosting the performance of numerous challenging applications involving continuous control, such as highly robust and agile robot motion control. However, their theoretical understanding remains largely underdeveloped. Existing analyses mostly focus on finite state-action spaces and on simplified variants of actor-critic, such as double-loop updates with i.i.d. sampling, which are often impractical for real-world applications. We consider the canonical and widely adopted single-timescale updates with Markovian sampling in continuous state-action space. Specifically, we establish finite-time convergence by introducing a novel Lyapunov analysis framework, which provides a unified convergence characterization of both the actor and the critic. Our approach is less conservative than previous methods and offers new insights into the coupled dynamics of actor-critic updates.

Generalist robots that can perform a range of different tasks in open-world settings must be able to not only reason about the steps needed to accomplish their goals, but also process complex instructions, prompts, and even feedback during task execution. Intricate instructions (e.g., "Could you make me a vegetarian sandwich?" or "I don't like that one") require not just the ability to physically perform the individual steps, but the ability to situate complex commands and feedback in the physical world. In this work, we describe a system that uses vision-language models in a hierarchical structure, first reasoning over complex prompts and user feedback to deduce the most appropriate next step to fulfill the task, and then performing that step with low-level actions. In contrast to direct instruction following methods that can fulfill simple commands ("pick up the cup"), our system can reason through complex prompts and incorporate situated feedback during task execution ("that's not trash"). We evaluate our system across three robotic platforms, including single-arm, dual-arm, and dual-arm mobile robots, demonstrating its ability to handle tasks such as cleaning messy tables, making sandwiches, and grocery shopping. Videos are available at https://www.pi.website/research/hirobot

Generalizing vision-based reinforcement learning (RL) agents to novel environments remains a difficult and open challenge. Current trends are to collect large-scale datasets or use data augmentation techniques to prevent overfitting and improve downstream generalization. However, the computational and data collection costs increase exponentially with the number of task variations and can destabilize the already difficult task of training RL agents. In this work, we take inspiration from recent advances in computational neuroscience and propose a model, Associative Latent DisentAnglement (ALDA), that builds on standard off-policy RL towards zero-shot generalization. Specifically, we revisit the role of latent disentanglement in RL and show how combining it with a model of associative memory achieves zero-shot generalization on difficult task variations *without* relying on data augmentation. Finally, we formally show that data augmentation techniques are a form of weak disentanglement and discuss the implications of this insight.

Designing effective reward functions remains a fundamental challenge in reinforcement learning (RL), as it often requires extensive human effort and domain expertise. While RL from human feedback has been successful in aligning agents with human intent, acquiring high-quality feedback is costly and labor-intensive, limiting its scalability. Recent advancements in foundation models present a promising alternative--leveraging AI-generated feedback to reduce reliance on human supervision in reward learning. Building on this paradigm, we introduce ERL-VLM, an enhanced rating-based RL method that effectively learns reward functions from AI feedback. Unlike prior methods that rely on pairwise comparisons, ERL-VLM queries large vision-language models (VLMs) for absolute ratings of individual trajectories, enabling more expressive feedback and improved sample efficiency. Additionally, we propose key enhancements to rating-based RL, addressing instability issues caused by data imbalance and noisy labels. Through extensive experiments across both low-level and high-level control tasks, we demonstrate that ERL-VLM significantly outperforms existing VLM-based reward generation methods. Our results demonstrate the potential of AI feedback for scaling RL with minimal human intervention, paving the way for more autonomous and efficient reward learning.

We reveal that feedforward network (FFN) layers, rather than attention layers, are the primary contributors to Vision Transformer (ViT) inference latency, with their impact signifying as model size increases. This finding highlights a critical opportunity for optimizing the efficiency of large-scale ViTs by focusing on FFN layers. In this work, we propose a novel channel idle mechanism that facilitates post-training structural reparameterization for efficient FFN layers during testing. Specifically, a set of feature channels remains idle and bypasses the nonlinear activation function in each FFN layer, thereby forming a linear pathway that enables structural reparameterization during inference. This mechanism results in a family of **RePa**rameterizable **Vi**sion **T**ransformers (RePaViTs), which achieve remarkable latency reductions with acceptable sacrifices (sometimes gains) in accuracy across various ViTs. The benefits of our method scale consistently with model sizes, demonstrating greater speed improvements and progressively narrowing accuracy gaps or even higher accuracies on larger models. In particular, RePa-ViT-Large and RePa-ViT-Huge enjoy **66.8%** and **68.7%** speed-ups with **+1.7%** and **+1.1%** higher top-1 accuracies under the same training strategy, respectively. RePaViT is the first to employ structural reparameterization on FFN layers to expedite ViTs to our best knowledge, and we believe that it represents an auspicious direction for efficient ViTs. Source code is available at https://github.com/Ackesnal/RePaViT.

Generative recommendation (GR) is an emerging paradigm where user actions are tokenized into discrete token patterns and autoregressively generated as predictions. However, existing GR models tokenize each action independently, assigning the same fixed tokens to identical actions across all sequences without considering contextual relationships. This lack of context-awareness can lead to suboptimal performance, as the same action may hold different meanings depending on its surrounding context. To address this issue, we propose ActionPiece to explicitly incorporate context when tokenizing action sequences. In ActionPiece, each action is represented as a *set* of item features. Given the action sequence corpora, we construct the vocabulary by merging feature patterns as new tokens, based on their co-occurrence frequency both within individual sets and across adjacent sets. Considering the unordered nature of feature sets, we further introduce set permutation regularization, which produces multiple segmentations of action sequences with the same semantics. Our code is available at: https://github.com/google-deepmind/action_piece.

In-Context Reinforcement Learning (ICRL) represents a promising paradigm for developing generalist agents that learn at inference time through trial-and-error interactions, analogous to how large language models adapt contextually, but with a focus on reward maximization. However, the scalability of ICRL beyond toy tasks and single-domain settings remains an open challenge. In this work, we present the first steps toward scaling ICRL by introducing a fixed, cross-domain model capable of learning behaviors through in-context reinforcement learning. Our results demonstrate that Algorithm Distillation, a framework designed to facilitate ICRL, offers a compelling and competitive alternative to expert distillation to construct versatile action models. These findings highlight the potential of ICRL as a scalable approach for generalist decision-making systems.

Typographic attacks are often attributed to the ability of multimodal pre-trained models to fuse textual semantics into visual representations, yet the mechanisms and locus of such interference remain unclear. We examine whether such models genuinely encode textual semantics or primarily rely on texture-based visual features. To disentangle orthographic form from meaning, we introduce the ToT dataset, which includes controlled word pairs that either share semantics with distinct appearances (synonyms) or share appearance with differing semantics (paronyms). A layer-wise analysis of Intrinsic Dimension (ID) reveals that early layers exhibit competing dynamics between orthographic and semantic representations. In later layers, semantic accuracy increases as ID decreases, but this improvement largely stems from orthographic disambiguation. Notably, clear semantic differentiation emerges only in the final block, challenging the common assumption that semantic understanding is progressively constructed across depth. These findings reveal how current vision-language models construct text representations through texture-dependent processes, prompting a reconsideration of the gap between visual perception and semantic understanding. The code is available at: https://github.com/Ovsia/Textural-or-Textual

Visual language is a system of communication that conveys information through symbols, shapes, and spatial arrangements. Diagrams are a typical example of a visual language depicting complex concepts and their relationships in the form of an image. The symbolic nature of diagrams presents significant challenges for building models capable of understanding them. Yet, recent studies seem to suggest that Large Vision-Language Models (LVLMs) can even tackle complex reasoning tasks involving diagrams. In this paper, we investigate this phenomenon by developing a comprehensive test suite to evaluate the diagram comprehension capability of LVLMs. Our test suite uses a variety of questions focused on concept entities and their relationships over a set of synthetic as well as real diagrams across several domains to evaluate the recognition and reasoning abilities of models. Our evaluation of six LVLMs shows that while these models can accurately identify and reason about entities, their ability to understand relationships is notably limited. Further testing reveals that the decent performance on diagram understanding largely stems from leveraging their background knowledge as shortcuts to identify and reason about the relational information. Thus, we conclude that LVLMs have a limited capability for genuine diagram understanding, and their impressive performance in diagram reasoning is an illusion emanating from other confounding factors, such as the background knowledge in the models.

Pre-trained vision-language models (VLMs) have revolutionized the field of machine learning, demonstrating exceptional performance across a wide range of tasks. However, their robustness remains vulnerable to the spurious-correlation problem. Existing works often involve fine-tuning the model with labeled data or relying on large language models (LLMs) to generate more complex prompts. Although effective to some extent, these methods introduce new challenges, including additional computational costs and dependence on the quality of prompts without fully utilizing the vision modality. To address these limitations, we propose a novel method named ERICT to Enhance model Robustness by Identifying Concept Tokens. ERICT mitigates spurious correlation directly in the inference stage and comprises two key steps: (1) Identify concept tokens capturing invariant features through auxiliary prompts to generate a token-level mask. (2) Apply the mask to the attention weights of the CLS token in the vision encoder to help the model focus on the relevant image region. Extensive experiments show that ERICT significantly improves the overall performance including that of the worst group, and achieves new state-of-the-art results.

Offline reinforcement learning is widely applied in multiple fields due to its advantages in efficiency and risk control. However, a major problem it faces is the distribution shift between offline datasets and online environments. This mismatch leads to out-of-distribution (OOD) state-action pairs that fall outside the scope of the training data. Therefore, existing conservative training policies may not provide reliable decisions when the test environment deviates greatly from the offline dataset. In this paper, we propose Test-time Adapted Reinforcement Learning (TARL) to address this problem. TARL constructs unsupervised test-time optimization objectives for discrete and continuous control tasks, using test data without depending on environmental rewards. In discrete control tasks, it minimizes the entropy of predicted action probabilities to decrease uncertainty and avoid OOD state-action pairs. For continuous control tasks, it represents and minimizes action uncertainty based on the normal distribution of policy network outputs. Moreover, to prevent model bias caused by overfitting and error accumulation during the test-time update process, TARL enforces a KL divergence constraint between the fine-tuned policy and the original policy. For efficiency, TARL only updates the layer normalization layer parameters during testing. Extensive experiments on popular Atari game benchmarks and the D4RL dataset demonstrate the superiority of our method. Our method achieved a significant improvement over CQL, with a 13.6% episode return relative increase on the hopper-expert-v2 task.

Off-policy deep reinforcement learning (RL) agents typically leverage replay buffers for reusing past experiences during learning. This can help sample efficiency when the collected data is informative and aligned with the learning objectives; when that is not the case, it has the effect of ``polluting'' the replay buffer with data that can exacerbate optimization challenges in addition to wasting environment interactions due to redundant sampling. We argue that sampling these uninformative and wasteful transitions can be avoided by addressing the **sunk cost fallacy** which, in the context of deep RL, is the tendency towards continuing an episode until termination. To address this, we propose the *learn to stop* (**LEAST**) mechanism which uses statistics based on $Q$-values and gradient to guide early episode termination which helps agents recognize when to terminate unproductive episodes early. We demonstrate that our method improves learning efficiency on a variety of RL algorithms, evaluated on both the MuJoCo and DeepMind Control Suite benchmarks.