Vision-Language Pretraining

AI developers face a dual-use dilemma. A model capability that helps one user cure a disease can help another synthesize one. This dilemma could be resolved with access control, limiting dual-use AI capabilities to trusted deployments with a legitimate need. A gold standard for access control would be to serve separate models with different capabilities to different users. However, training and deploying multiple models is prohibitively expensive. To address this challenge, we propose gradient-routed auxiliary modules (GRAM), a pre-training method that adds modules to a neural network and selectively updates them to induce specialization. Ablating a module at inference time removes its capability from the network, approximating a model trained on filtered data. We evaluate GRAM on synthetic stories and realistic dual-use data spanning virology, cybersecurity, nuclear physics, and specialized code. These experiments show that GRAM preserves selected retain capabilities while disabling forgotten capabilities, and limits recovery better than post-hoc unlearning. Most importantly, a Chinchilla-optimal scaling analysis from 50M to 5B parameters shows that the forget capability gap between data-filtered and full-data models widens with scale while the retain gap stays constant, and that GRAM closely tracks data filtering. GRAM's training cost is independent of the number of supported capability profiles, yielding a 5x cost reduction in our 5-profile setting.

Effective data selection is essential for pretraining large language models (LLMs), improving efficiency and generalization to downstream tasks. However, existing approaches often rely on external pretrained models, making it difficult to separate the benefits of data selection from those introduced by external models. In addition, many methods estimate data importance from a fixed model state or short-horizon update, making it hard to capture how data preference changes as the model evolves during pretraining. In this paper, we introduce BLISS (BileveL Influence Scoring method for data Selection), a lightweight data selection method that operates entirely \emph{from scratch}, without external pretrained oracle models, while modeling dynamic data preference. BLISS uses a small proxy model as a surrogate for the LLM and trains a score model to estimate sample importance through multi-step proxy updates induced by score-weighted training data. We formulate data selection as a bilevel optimization problem: the upper-level objective optimizes the score model to assign sample weights, so minimizing the lower-level weighted training loss improves validation performance. Once optimized, the score model predicts influence scores, enabling efficient selection of high-quality samples for LLM pretraining. We validate BLISS by pretraining 410M/1B/2.8B Pythia and LLaMA-0.5B models on selected C4 subsets. Under the 1B setting, BLISS achieves a $1.7\times$ speedup in reaching the same performance as the state-of-the-art method, while delivering superior performance across multiple downstream tasks.

Image-based Joint-Embedding Predictive Architecture (I-JEPA) offers a promising approach to visual self-supervised learning through masked feature prediction. However with the inherent visual uncertainty at masked positions, feature prediction remains challenging and may fail to learn semantic representations. In this work, we propose Text-Conditional JEPA (TC-JEPA) that uses image captions to reduce the prediction uncertainty. Specifically, we modulate the predicted patch features using a fine-grained text conditioner that computes sparse cross-attention over input text tokens. With such conditioning, patch features become predictable as a function of text, thus are more semantically meaningful. We show TC-JEPA improves downstream performance and training stability, with promising scaling properties. TC-JEPA also offers a new vision-language pretraining paradigm based on feature prediction only, outperforming contrastive methods on diverse tasks, especially those requiring fine-grained visual understanding and reasoning.

Large Vision-Language Models (LVLMs) demonstrate impressive multimodal capabilities, yet suffer from hallucination—generating factually inaccurate content. Contrastive Decoding (CD) mitigates this by contrasting amateur and expert branches at the logit level. However, our investigation reveals that such logit-level interventions fundamentally compromise generation coherence, necessitating restrictive penalty constraints unrelated to hallucination suppression. We introduce Attention Contrastive Decoding (ACD), a training-free plug-in that complements logit-level CD by relocating part of the contrastive operations to the attention mechanism. Operating at an earlier stage of the forward pass, ACD performs smooth semantic-preserving interventions through an Adaptive Subtraction Strategy (ASS), which attenuates hallucination-associated attention patterns while amplifying critical visual information. Extensive experiments demonstrate that combining ACD with existing CD methods (e.g., VCD+ACD) produces substantially more coherent outputs with further reduced hallucinations, eliminating restrictive penalties while enabling trustworthy multimodal generation.

Visual Contrastive Decoding (VCD) mitigates hallucinations in Multimodal Large Language Models (MLLMs) by penalizing the output shift from noise-perturbed images, assuming this shift captures the hallucination direction. We prove this assumption flawed: noise-induced drift in Language-Image Pretrained (LIP) encoders is a \emph{coupled vector} entangling (i) structural degradation from corrupted visual information with (ii) hallucination induction from linguistic prior activation. VCD's indiscriminate penalty inevitably suppresses valid visual semantics. Our key insight is that Self-Supervised Learning (SSL) encoders exhibit \emph{only} structural degradation under noise—geometrically orthogonal to hallucination paths—enabling principled disentanglement via LIP--SSL differential response. We propose \textbf{Disentangled Visual Rectification (DVR)}, a training-free dual-stream framework performing visual-layer rectification and decoding-layer contrast on purified representations. DVR achieves approximately $5\times$ theoretical error reduction over VCD and establishes SOTA performance on POPE, MME, LLaVA-Bench and CHAIR benchmarks.

Inspired by the idea of Positive-incentive Noise (*Pi-Noise* or *$\pi$-Noise*) that aims at learning the reliable noise beneficial to tasks, we scientifically investigate the connection between contrastive learning and $\pi$-noise in this paper. By converting the contrastive loss to an auxiliary Gaussian distribution to quantitatively measure the difficulty of the specific contrastive model under the information theory framework, we properly define the task entropy, the core concept of $\pi$-noise, of contrastive learning. It is further proved that the predefined data augmentation in the standard contrastive learning paradigm can be regarded as a kind of point estimation of $\pi$-noise. Inspired by the theoretical study, a framework that develops a $\pi$-noise generator to learn the beneficial noise (instead of estimation) as data augmentations for contrast is proposed. The designed framework can be applied to diverse types of data and is also completely compatible with the existing contrastive models. From the visualization, we surprisingly find that the proposed method successfully learns effective augmentations.

The factors driving the performance of in-context learning (ICL) in large language models (LLMs) remain poorly understood despite ICL's surprising effectiveness, enabling models to adapt to new tasks from only a handful of examples. To clarify and improve these capabilities, we characterize how the statistical properties of the pretraining distribution (e.g., tail behavior, coverage) shape ICL. We develop a theoretical framework that encompasses generalization and task selection and show how distributional properties govern sample efficiency, task retrieval, and robustness. To this end, we generalize existing concentration results to heavy-tailed priors and dependent sequences, better reflecting the structure of LLM pretraining data. Our framework reveals a fundamental design trade-off: heavy-tailed pretraining distributions facilitate robust task selection under distribution shifts but are detrimental to generalization, especially in low-data regimes. We then empirically evaluate our predictions by studying how ICL performance varies with the pretraining distribution on challenging tasks such as stochastic differential equations and stochastic processes with memory. Together, these findings suggest that controlling key statistical properties of the pretraining distribution is essential for building ICL-capable and reliable LLMs.

Language models achieve impressive performance on a variety of knowledge, language, and reasoning tasks due to the scale and diversity of pretraining data available. The standard training recipe is a two-stage paradigm: pretraining first on the full corpus of data followed by specialization on a subset of high quality, specialized data from the full corpus. In the multi-domain setting, this involves continued pretraining of multiple models on each specialized domain, referred to as split model training. We propose a method for pretraining multiple models independently over a general pretraining corpus, and determining the optimal compute allocation between pretraining and continued pretraining using scaling laws. Our approach accurately predicts the loss of a model of size $N$ with $D$ pretraining and $D'$ specialization tokens, and extrapolates to larger model sizes and number of tokens. Applied to language model training, our approach improves performance consistently across common sense knowledge and reasoning benchmarks across different model sizes and compute budgets.

Language models are pretrained on sequences that blend statistical regularities, which make text fluent, with factual associations between specific tokens, which encode knowledge of facts. While recent works suggest that generalization depends critically on the interaction of these two streams, such as the diversity of the contexts in which facts appear, these effects remain difficult to study systematically. This paper introduces a flexible synthetic testbed that combines a statistical stream of generic tokens with an abstract factual stream of source-target token pairs, enabling fine-grained control over their interaction, such as their composition into a sequence (contextual structure) or the level of context diversity carrying the facts at training time. Through controlled experiments, we find that higher contextual diversity delays in-distribution factual learning, while low diversity can harm out-of-distribution generalization in ways that depend on the contextual structure. As a result, the optimal diversity level depends on the training budget. Beyond factual recall failures, we also identify failures in statistical generalization as we study how the interplay between contextual design and diversity level impacts different aspects of generalization. Furthermore, through a series of controlled interventions on the model components, we trace failure in different aspects of generalization to distinct optimization bottlenecks, highlighting the importance of the embedding and unembedding layers. Overall, our synthetic framework allows us to isolate effects that would be confounded in large-scale studies, offering a controlled testbed for future investigations.

Vision–language pre-training (VLP) holds great promise for general-purpose medical AI by leveraging radiology reports as rich textual supervision, yet existing methods struggle with 3D CT imaging due to inefficient visual backbones and coarse semantic alignment. To address these issues, we propose a tailored VLP framework featuring three key components: (1) a CNN–ViT hybrid encoder that replaces ViT’s patch embedding with a 3D CNN backbone to efficiently capture local anatomical details while preserving global attention and compatibility with pre-trained cross-modal priors; (2) a disease-level contrastive learning mechanism using learnable query tokens to dynamically extract disease-specific semantics from full reports and align them with corresponding visual features, thereby disentangling distinct diseases within the same anatomical region; and (3) a diagnosis-aware prompt strategy that employs real clinical phrases and aggregated disease prototypes to bridge the pre-training–inference gap and enhance zero-shot diagnostic reliability. Our model achieves state-of-the-art performance on CT-RATE (84.4\% AUC, +5.1%) and Rad-ChestCT (75.4\% AUC, +5.4%), with even larger gains (+9.8% AUC) on a challenging 60-disease benchmark, and demonstrates strong transferability to radiology report generation, underscoring the generality and clinical utility of our approach.

Vision-Language-Action (VLA) models are bottlenecked by the scarcity of expert demonstrations—expensive triplets of observations, language instructions, and actions. We propose that learning ''how to move'' can be decoupled from learning ''what to do,'' and that the former requires no task labels at all. Our two-stage framework, **Task-Agnostic Pretraining (TAP)** first pre-trains on abundant, cheap *task-agnostic* data (discarded off-task trajectories or autonomous robot play) using an Inverse Dynamics objective that predicts actions from consecutive observations. This self-supervised phase instills physical affordances—grasping, contact dynamics, end-effector control—without human annotation. A lightweight second stage then aligns these physical priors with language instructions using minimal expert data. On the SIMPLER benchmark, our approach matches models trained on 1M+ expert trajectories while using orders of magnitude less labeled data, achieving a 10\% absolute gain over standard behavior cloning. In real-world WidowX experiments, it surpasses internet-scale baselines under visual distribution shifts (e.g., 25\% vs. 0\% under camera perturbations), demonstrating that task-agnostic pretraining yields robust, transferable physical representations for Embodied AI.

Standard practice across domains from robotics to language is to first pretrain a policy on a large-scale demonstration dataset, and then finetune this policy, typically with reinforcement learning (RL), in order to improve performance on deployment domains. This finetuning step has proved critical in achieving human or super-human performance, yet while much attention has been given to developing more effective finetuning algorithms, little attention has been given to ensuring the pretrained policy is an effective initialization for RL finetuning. In this work we seek to understand how the pretrained policy affects finetuning performance, and how to pretrain policies in order to ensure they are effective initializations for finetuning. We first show theoretically that standard behavioral cloning (BC) can fail to ensure coverage over the demonstrator's actions, a minimal condition necessary for effective RL finetuning. We then show that if, instead of exactly fitting the observed demonstrations, we train a policy to model the posterior distribution of the demonstrator's behavior given the demonstration dataset, we do obtain a policy that ensures coverage over the demonstrator's actions, enabling more effective finetuning. Furthermore, this policy achieves this while ensuring pretrained performance is no worse than that of the BC policy. We then show this approach is practically implementable with modern generative models and leads to significantly improved RL finetuning performance on both realistic robotic control benchmarks and real-world robotic manipulation tasks, as compared to standard behavioral cloning.

Why does contrastive learning with simple images and augmentations yield useful representations for downstream tasks? We address this question by analytically computing the optimal representation in terms of the contrastive loss for a range of basic augmentations and any image dataset with stationary statistics. We show that for certain augmentations the optimum can be attained by a CNN whose first layer filters are sinusoids, followed by a pointwise nonlinearity, global average pooling, and a final linear layer that performs partial whitening. We also show that the optimal weights in such CNNs for more complicated augmentations are still sinusoids. The frequencies of the sinusoids and their weights can be computed using a simple “waterfilling” algorithm given the dataset’s expected power spectrum. Experiments with different image datasets and augmentations show that such CNNs trained with SGD empirically learn sinusoids in their first layer and to perform partial whitening.

Efficient transfer learning of vision–language models (VLMs) commonly suffers from a Base–New Trade-off (BNT): improving performance on unseen (new) classes often degrades accuracy on known (base) classes. Addressing how to boost recognition of unseen classes without sacrificing known-class performance remains a central challenge. Existing work often simplistically attributes the BNT to overfitting on known classes. We observe an interesting phenomenon: VLMs frequently exhibit asymmetric confusion on certain downstream data, i.e., samples of class A are systematically mispredicted as class B, while the reverse confusion (B → A) rarely occurs. For known classes, this kind of bias can be mitigated by tuning using a cross-entropy loss, but for unseen classes, such pretraining-induced bias persists and harms generalization. Motivated by this, we propose NeRP, a plug-and-play prompting correction strategy that improves discrimination on unseen classes without modifying model parameters. NeRP leverages neutral text prompts and reference images to measure class-wise prior preferences along the pre-trained inter-class geometry, and combines them with the sample likelihood to obtain the model’s surrogate score. If, for a given sample, the prior strongly favors the current prediction while the observed evidence is clearly insufficient, we perform a local flip between easily confusable class pairs, thereby correcting prior-dominated mispredictions. Extensive experiments across multiple backbones and 15 few-shot and cross-domain benchmarks show that NeRP substantially improves accuracy on unseen classes while preserving known-class prediction performance.

Large Vision–Language Models (LVLMs) have shown outstanding performance across various multimodal tasks, but they still suffer from hallucinations, where they generate incorrect information by relying on language priors without visual grounding. To alleviate this issue, prior work has explored contrastive decoding approaches that compare the output of an original LVLM with that of a contrast branch. However, existing methods typically generate contrast logits through preprocessing of the input image. Such input-level perturbations fail to sufficiently reflect the model’s internal degree of visual reliance during the decoding process. To address this limitation, we propose Dynamic Entropy-Guided Attention Perturbation (DEGAP) for contrastive decoding in LVLMs. DEGAP performs contrastive decoding by directly perturbing visual attention and leveraging the resulting logits, without requiring any additional image preprocessing. To this end, we analyze the layer-wise effects of visual attention perturbations and, based on these observations, dynamically select the layers at which attention perturbation is applied according to the model’s confidence. Experimental results on seven benchmarks demonstrate that DEGAP effectively mitigates various types of hallucinations and consistently outperforms state-of-the-art methods in general VQA performance.

Mislabeled samples in training datasets severely degrade the performance of deep networks, as overparameterized models tend to memorize erroneous labels. We address this challenge by proposing a novel approach for mislabeled data detection that leverages training dynamics. Our method is grounded in the key observation that correctly labeled samples exhibit consistent entropy decrease during training, while mislabeled samples maintain relatively high entropy throughout the training process. Building on this insight, we introduce a signed entropy integral (SEI) statistic that captures both the magnitude and temporal trend of prediction entropy across training epochs. SEI is broadly applicable to classification networks and demonstrates particular effectiveness when integrated with contrastive language-image pretraining (CLIP) architectures. Through extensive experiments on four medical imaging datasets---a domain particularly susceptible to labeling errors due to diagnostic complexity---spanning diverse modalities and pathologies, we demonstrate that SEI achieves state-of-the-art performance in mislabeled data identification, outperforming existing methods while maintaining computational efficiency and implementation simplicity. Our code is available at https://github.com/MedAITech/SEI.

Most existing CLIP-style medical vision--language pretraining methods rely on global or local alignment with substantial paired data. However, global alignment is easily dominated by non-diagnostic information, while local alignment fails to integrate key diagnostic evidence. As a result, learning reliable diagnostic representations becomes difficult, which limits their applicability in medical scenarios with limited paired data. To address this issue, we propose an LLM-Guided Diagnostic Evidence Alignment method (LGDEA), which shifts the pretraining objective toward evidence-level alignment that is more consistent with the medical diagnostic process. Specifically, we leverage LLMs to extract key diagnostic evidence from radiology reports and construct a shared diagnostic evidence space, enabling evidence-aware cross-modal alignment and allowing LGDEA to effectively exploit abundant unpaired medical images and reports, thereby substantially alleviating the reliance on paired data. Extensive experimental results demonstrate that our method achieves consistent and significant improvements on phrase grounding, image--text retrieval, and zero-shot classification, and even rivals pretraining methods that rely on substantial paired data.

Heterogeneous multi-modal remote sensing object detection aims to accurately detect objects from diverse sensors (e.g., RGB, SAR, Infrared). Existing approaches largely adopt a late alignment paradigm, in which modality alignment and task-specific optimization are entangled during downstream fine-tuning. This tight coupling complicates optimization and often results in unstable training and suboptimal generalization. To address these limitations, we propose BabelRS, a unified language-pivoted pretraining framework that explicitly decouples modality alignment from downstream task learning. BabelRS comprises two key components: Concept-Shared Instruction Aligning (CSIA) and Layerwise Visual-Semantic Annealing (LVSA). CSIA aligns each sensor modality to a shared set of linguistic concepts, using language as a semantic pivot to bridge heterogeneous visual representations. To further mitigate the granularity mismatch between high-level language representations and dense detection objectives, LVSA progressively aggregates multi-scale visual features to provide fine-grained semantic guidance. Extensive experiments demonstrate that BabelRS stabilizes training and consistently outperforms state-of-the-art methods without bells and whistles. Code: \href{https://github.com/zcablii/SM3Det}{github.com/zcablii/SM3Det}.

Pretraining transformers on long sequences, such as entire code repositories or collections of related documents, is bottlenecked by quadratic attention costs. We present Multipole Semantic Attention (MuSe), which accelerates 64k-context pretraining by 36% while matching baseline loss, requiring no architectural changes. MuSe is a training-time approximation that clusters queries and keys separately in representation space. This yields query-specific summaries that substantially outperform spatial blocking at matched sparsity, while also enabling drop-in compatibility with existing pretrained models—we validate on Llama 3.1-8B and 3.2-1B without retraining. We pretrain language models up to 1B parameters at 64k context on code and scientific documents, confirming that MuSe preserves quality and long-context utilization during training.