ATAS: Any-to-Any Self-Distillation for Enhanced Open-Vocabulary Dense Prediction

Vision-language models such as CLIP have recently propelled open-vocabulary dense prediction tasks by enabling recognition of a broad range of visual concepts. However, CLIP still struggles with fine-grained, region-level understanding, hindering its effectiveness on these dense prediction tasks. We identify two pivotal factors required to address this limitation: semantic coherence and fine-grained vision-language alignment. Current adaptation methods often improve fine-grained alignment at the expense of semantic coherence, and often rely on extra modules or supervised fine-tuning. To overcome these issues, we propose Any-to-Any Self-Distillation (ATAS), a novel approach that simultaneously enhances semantic coherence and fine-grained alignment by leveraging own knowledge of a model across all representation levels. Unlike prior methods, ATAS uses only unlabeled images and an internal self-distillation process to refine representations of CLIP vision encoders, preserving local semantic consistency while sharpening local detail recognition. On open-vocabulary object detection and semantic segmentation benchmarks, ATAS achieves substantial performance gains, outperforming baseline CLIP models. These results validate the effectiveness of our approach and underscore the importance of jointly maintaining semantic coherence and fine-grained alignment for advanced open-vocabulary dense prediction.

From Threat to Tool: Leveraging Refusal-Aware Injection Attacks for Safety Alignment

Safely aligning large language models (LLMs) often demands extensive human-labeled preference data, a process that's both costly and time-consuming. While synthetic data offers a promising alternative, current methods frequently rely on complex iterative prompting or auxiliary models. To address this, we introduce Refusal-Aware Adaptive Injection (RAAI), a straightforward, training-free, and model-agnostic framework that repurposes LLM attack techniques. RAAI works by detecting internal refusal signals and adaptively injecting predefined phrases to elicit harmful, yet fluent, completions. Our experiments show RAAI effectively jailbreaks LLMs, increasing the harmful response rate from a baseline of 2.15% to up to 61.04% on average across four benchmarks. Crucially, fine-tuning LLMs with the synthetic data generated by RAAI improves model robustness against harmful prompts while preserving general capabilities on standard tasks like MMLU and ARC. This work highlights how LLM attack methodologies can be reframed as practical tools for scalable and controllable safety alignment.

Parameter-Efficient Fine-Tuning with Column Space Projection

Fine-tuning large language models (LLMs) with minimal computational overhead is essential for efficiently adapting them to downstream tasks under resource constraints. Parameter-efficient fine-tuning (PEFT) methods, such as Low-Rank Adaptation (LoRA), facilitate this by updating only a small subset of parameters. However, recent studies show that LoRA diverges from full fine-tuning (Full FT) in its learning behavior, particularly in terms of spectral properties. Motivated by these findings, we propose PiCa, the first theoretically grounded PEFT method based on the spectral properties of fine-tuned weights. PiCa projects gradients onto the low-rank column subspace of pre-trained weights and exhibits learning patterns more closely aligned with Full FT. Furthermore, we show that combining PiCa with weight sharing drastically reduces the number of trainable parameters without compromising performance, enabling to achieve superior performance than LoRA using 13x fewer trainable parameters. Extensive experiments demonstrate PiCa achieves the state-of-the-art performance compared to existing PEFT methods.

Generalized and Personalized Federated Learning with Foundation Models via Orthogonal Transformations

Federated Learning (FL) aims to train models across decentralized clients or devices holding local data without the need for centralized data collection, thus enhancing data privacy and security. However, achieving both generalization and personalization in heterogeneous settings remains a significant challenge. To address this, we introduce FedOT, a novel approach that leverages black-box foundation models. FedOT shares only a global task-dependent classifier across clients while locally adapting features through orthogonal transformations. By enforcing orthogonality, FedOT mitigates gradient conflicts across diverse clients, preserves semantic integrity, and achieves robust performance even in the presence of substantial data heterogeneity. The strategy of combining global and local parameters enables a more balanced approach for both generalization and personalization, outperforming baseline FL methods across multiple benchmarks. Furthermore, our extensive analysis confirms that joint optimization of global classifiers and local orthogonal transformations yields superior performance and suggests broader applicability.

Regularized Personalization of Text-to-Image Diffusion Models without Distributional Drift

Personalization using text-to-image diffusion models involves adapting a pretrained model to novel subjects with only a few image examples. This task presents a fundamental challenge, as the model must not only learn the new subject effectively but also preserve its ability to generate diverse and coherent outputs across a wide range of prompts. In other words, successful personalization requires integrating new concepts without forgetting previously learned generative capabilities. Forgetting denotes unintended distributional drift, where the model's output distribution deviates from that of the original pretrained model. In this paper, we provide an analysis of this issue and identify a mismatch between standard training objectives and the goals of personalization. To address this, we propose a new training objective based on a Lipschitz-bounded formulation that explicitly constrains deviation from the pretrained distribution. Our method provides improved control over distributional drift and performs well even in data-scarce scenarios. Experimental results demonstrate that our approach consistently outperforms existing personalization methods, achieving higher CLIP-T, CLIP-I, and DINO scores.

Energy-based Preference Optimization for Test-time Adaptation

Test-Time Adaptation (TTA) enhances model robustness by enabling adaptation to target distributions that differ from training distributions, improving real-world generalizability. Existing TTA approaches focus on adjusting the conditional distribution; however these methods often depend on uncertain predictions in the absence of label information, leading to unreliable performance. Energy-based frameworks suggest a promising alternative to address distribution shifts without relying on uncertain predictions, instead computing the marginal distribution of target data. However, they involve the critical challenge of requiring extensive SGLD sampling, which is impractical for test-time scenarios requiring immediate adaptation. In this work, we propose Energy-based Preference Optimization for Test-time Adaptation (EPOTTA), which is based on a sampling free strategy. We first parameterize the target model using a pretrained model and residual energy function, enabling marginal likelihood maximization of target data without sampling. Building on the observation that the parameterization is mathematically equivalent to DPO objective, we then directly adapt the model to a target distribution without explicitly training the residual. Our experiments verify that EPOTTA is well-calibrated and performant while achieving computational efficiency.

"Well, Keep Thinking": Enhancing LLM Reasoning with Adaptive Injection Decoding

Large language models (LLMs) exhibit strong reasoning abilities, often attributed to few-shot or zero-shot chain-of-thought (CoT) prompting. While effective, these methods require labor-intensive prompt engineering, raising the question of whether reasoning can be induced without reliance on explicit prompts. In this work, we unlock the reasoning capabilities of LLMs without explicit prompting. Inspired by zero-shot CoT and CoT-decoding, we propose a novel decoding strategy that systematically nudges LLMs to continue reasoning, thereby preventing immature reasoning processes. Specifically, we monitor the model's generation and inject a designated phrase whenever it is likely to conclude its response prematurely, before completing the reasoning process. Our experimental evaluations on diverse reasoning benchmarks demonstrate that our proposed strategy substantially improves LLM reasoning capabilities, highlighting the potential of decoding-based interventions as an alternative to traditional prompting techniques.

Object-Centric World Model for Language-Guided Manipulation

A world model is essential for an agent to predict the future and plan in domains such as autonomous driving and robotics. To achieve this, recent advancements have focused on video generation, which has gained significant attention due to the impressive success of diffusion models. However, these models require substantial computational resources. To address these challenges, we propose a world model leveraging object-centric representation space using slot attention, guided by language instructions. Our model perceives the current state as an object-centric representation and predicts future states in this representation space conditioned on natural language instructions. This approach results in a more compact and computationally efficient model compared to diffusion-based generative alternatives. Furthermore, it flexibly predicts future states based on language instructions, and offers a significant advantage in manipulation tasks where object recognition is crucial. In this paper, we demonstrate that our latent predictive world model surpasses generative world models in visuo-linguo-motor control tasks, achieving superior sample and computation efficiency. We also investigate the generalization performance of the proposed method and explore various strategies for predicting actions using object-centric representations.

Retaining and Enhancing Pre-trained Knowledge in Vision-Language Models with Prompt Ensembling

The advancement of vision-language models, particularly the Contrastive Language-Image Pre-training (CLIP) model, has revolutionized the field of machine learning by enabling robust zero-shot learning capabilities. These capabilities allow models to understand and respond to previously unseen data without task-specific training. However, adapting CLIP to integrate specialized knowledge from various domains while retaining its zero-shot capabilities remains a significant challenge. To address this, we introduce a novel prompt ensemble learning approach called Group-wise Prompt Ensemble (GPE). This method aims to enhance CLIP's zero-shot capabilities by incorporating new domain knowledge while improving its adaptability and robustness against data distribution shifts. Our approach hinges on three main strategies: prompt grouping with masked attention to optimize CLIP's adaptability while safeguarding its zero-shot capabilities; the incorporation of auxiliary prompts for the seamless integration of new domain insights without disrupting the original model's representation; and an ensemble learning strategy that effectively merges original and new knowledge. Through rigorous experimentation, including more challenging cross-dataset transfer evaluations, our GPE method redefines the benchmarks for the adaptability and efficiency of vision-language models, surpassing existing models across various scenarios.

When Vision Models Meet Parameter Efficient Look-Aside Adapters Without Large-Scale Audio Pretraining

Recent studies show that pretrained vision models can boost performance in audio downstream tasks. To enhance the performance further, an additional pretraining stage with large scale audio data is typically required to infuse audio specific knowledge into the vision model. However, such approaches require extensive audio data and a carefully designed objective function. In this work, we propose bypassing the pretraining stage by directly fine-tuning the vision model with our Look Aside Adapter (LoAA) designed for efficient audio understanding. Audio spectrum data is represented across two heterogeneous dimensions time and frequency and we refine adapters to facilitate interactions between tokens across these dimensions. Our experiments demonstrate that our adapters allow vision models to reach or surpass the performance of pretrained audio models in various audio and speech tasks, offering a resource efficient and effective solution for leveraging vision models in audio applications.