Zhe Dong

Cicero Nogueira dos Santos, Zhe Dong, Daniel Cer et al. · berkeley, deepmind

Soft prompts have been recently proposed as a tool for adapting large frozen language models (LMs) to new tasks. In this work, we repurpose soft prompts to the task of injecting world knowledge into LMs. We introduce a method to train soft prompts via self-supervised learning on data from knowledge bases. The resulting soft knowledge prompts (KPs) are task independent and work as an external memory of the LMs. We perform qualitative and quantitative experiments and demonstrate that: (1) KPs can effectively model the structure of the training data; (2) KPs can be used to improve the performance of LMs in different knowledge intensive tasks.

Fedor Moiseev, Zhe Dong, Enrique Alfonseca et al.

Large language models (LLMs) have demonstrated human-level performance on a vast spectrum of natural language tasks. However, it is largely unexplored whether they can better internalize knowledge from a structured data, such as a knowledge graph, or from text. In this work, we propose a method to infuse structured knowledge into LLMs, by directly training T5 models on factual triples of knowledge graphs (KGs). We show that models pre-trained on Wikidata KG with our method outperform the T5 baselines on FreebaseQA and WikiHop, as well as the Wikidata-answerable subset of TriviaQA and NaturalQuestions. The models pre-trained on factual triples compare competitively with the ones on natural language sentences that contain the same knowledge. Trained on a smaller size KG, WikiMovies, we saw 3x improvement of exact match score on MetaQA task compared to T5 baseline. The proposed method has an advantage that no alignment between the knowledge graph and text corpus is required in curating training data. This makes our method particularly useful when working with industry-scale knowledge graphs.

Zhe Dong, Jianmo Ni, Daniel M. Bikel et al.

Dual encoders have been used for question-answering (QA) and information retrieval (IR) tasks with good results. Previous research focuses on two major types of dual encoders, Siamese Dual Encoder (SDE), with parameters shared across two encoders, and Asymmetric Dual Encoder (ADE), with two distinctly parameterized encoders. In this work, we explore different ways in which the dual encoder can be structured, and evaluate how these differences can affect their efficacy in terms of QA retrieval tasks. By evaluating on MS MARCO, open domain NQ and the MultiReQA benchmarks, we show that SDE performs significantly better than ADE. We further propose three different improved versions of ADEs by sharing or freezing parts of the architectures between two encoder towers. We find that sharing parameters in projection layers would enable ADEs to perform competitively with or outperform SDEs. We further explore and explain why parameter sharing in projection layer significantly improves the efficacy of the dual encoders, by directly probing the embedding spaces of the two encoder towers with t-SNE algorithm.

Fedor Moiseev, Gustavo Hernandez Abrego, Peter Dornbach et al.

Dual encoders have been used for retrieval tasks and representation learning with good results. A standard way to train dual encoders is using a contrastive loss with in-batch negatives. In this work, we propose an improved contrastive learning objective by adding queries or documents from the same encoder towers to the negatives, for which we name it as "contrastive loss with SAMe TOwer NEgatives" (SamToNe). By evaluating on question answering retrieval benchmarks from MS MARCO and MultiReQA, and heterogenous zero-shot information retrieval benchmarks (BEIR), we demonstrate that SamToNe can effectively improve the retrieval quality for both symmetric and asymmetric dual encoders. By directly probing the embedding spaces of the two encoding towers via the t-SNE algorithm (van der Maaten and Hinton, 2008), we observe that SamToNe ensures the alignment between the embedding spaces from the two encoder towers. Based on the analysis of the embedding distance distributions of the top-$1$ retrieved results, we further explain the efficacy of the method from the perspective of regularisation.

Haifeng Wang, Hua Wu, Tian Wu et al.

In this report, we introduce ERNIE 5.0, a natively autoregressive foundation model desinged for unified multimodal understanding and generation across text, image, video, and audio. All modalities are trained from scratch under a unified next-group-of-tokens prediction objective, based on an ultra-sparse mixture-of-experts (MoE) architecture with modality-agnostic expert routing. To address practical challenges in large-scale deployment under diverse resource constraints, ERNIE 5.0 adopts a novel elastic training paradigm. Within a single pre-training run, the model learns a family of sub-models with varying depths, expert capacities, and routing sparsity, enabling flexible trade-offs among performance, model size, and inference latency in memory- or time-constrained scenarios. Moreover, we systematically address the challenges of scaling reinforcement learning to unified foundation models, thereby guaranteeing efficient and stable post-training under ultra-sparse MoE architectures and diverse multimodal settings. Extensive experiments demonstrate that ERNIE 5.0 achieves strong and balanced performance across multiple modalities. To the best of our knowledge, among publicly disclosed models, ERNIE 5.0 represents the first production-scale realization of a trillion-parameter unified autoregressive model that supports both multimodal understanding and generation. To facilitate further research, we present detailed visualizations of modality-agnostic expert routing in the unified model, alongside comprehensive empirical analysis of elastic training, aiming to offer profound insights to the community.

Ziyue Huang, Mingming Zhang, Qingjie Liu et al.

The field of building detection from remote sensing images has made significant progress, but faces challenges in achieving high-accuracy detection due to the diversity in building appearances and the complexity of vast scenes. To address these challenges, we propose a novel approach called Context-Enhanced Detector (CEDet). Our approach utilizes a three-stage cascade structure to enhance the extraction of contextual information and improve building detection accuracy. Specifically, we introduce two modules: the Semantic Guided Contextual Mining (SGCM) module, which aggregates multi-scale contexts and incorporates an attention mechanism to capture long-range interactions, and the Instance Context Mining Module (ICMM), which captures instance-level relationship context by constructing a spatial relationship graph and aggregating instance features. Additionally, we introduce a semantic segmentation loss based on pseudo-masks to guide contextual information extraction. Our method achieves state-of-the-art performance on three building detection benchmarks, including CNBuilding-9P, CNBuilding-23P, and SpaceNet.

Zhe Dong, Yuzhe Sun, Tianzhu Liu et al.

Given a natural language expression and a remote sensing image, the goal of referring remote sensing image segmentation (RRSIS) is to generate a pixel-level mask of the target object identified by the referring expression. In contrast to natural scenarios, expressions in RRSIS often involve complex geospatial relationships, with target objects of interest that vary significantly in scale and lack visual saliency, thereby increasing the difficulty of achieving precise segmentation. To address the aforementioned challenges, a novel RRSIS framework is proposed, termed the cross-modal bidirectional interaction model (CroBIM). Specifically, a context-aware prompt modulation (CAPM) module is designed to integrate spatial positional relationships and task-specific knowledge into the linguistic features, thereby enhancing the ability to capture the target object. Additionally, a language-guided feature aggregation (LGFA) module is introduced to integrate linguistic information into multi-scale visual features, incorporating an attention deficit compensation mechanism to enhance feature aggregation. Finally, a mutual-interaction decoder (MID) is designed to enhance cross-modal feature alignment through cascaded bidirectional cross-attention, thereby enabling precise segmentation mask prediction. To further forster the research of RRSIS, we also construct RISBench, a new large-scale benchmark dataset comprising 52,472 image-language-label triplets. Extensive benchmarking on RISBench and two other prevalent datasets demonstrates the superior performance of the proposed CroBIM over existing state-of-the-art (SOTA) methods. The source code for CroBIM and the RISBench dataset will be publicly available at https://github.com/HIT-SIRS/CroBIM

Yuzhe Sun, Zhe Dong, Haochen Jiang et al.

Referring remote sensing image segmentation aims to localize specific targets described by natural language within complex overhead imagery. However, due to extreme scale variations, dense similar distractors, and intricate boundary structures, the reliability of cross-modal alignment exhibits significant \textbf{spatial non-uniformity}. Existing methods typically employ uniform fusion and refinement strategies across the entire image, which often introduces unnecessary linguistic perturbations in visually clear regions while failing to provide sufficient disambiguation in confused areas. To address this, we propose an \textbf{uncertainty-guided framework} that explicitly leverages a pixel-wise \textbf{referring uncertainty map} as a spatial prior to orchestrate adaptive inference. Specifically, we introduce a plug-and-play \textbf{Referring Uncertainty Scorer (RUS)}, which is trained via an online error-consistency supervision strategy to interpretably predict the spatial distribution of referential ambiguity. Building on this prior, we design two plug-and-play modules: 1) \textbf{Uncertainty-Gated Fusion (UGF)}, which dynamically modulates language injection strength to enhance constraints in high-uncertainty regions while suppressing noise in low-uncertainty ones; and 2) \textbf{Uncertainty-Driven Local Refinement (UDLR)}, which utilizes uncertainty-derived soft masks to focus refinement on error-prone boundaries and fine details. Extensive experiments demonstrate that our method functions as a unified, plug-and-play solution that significantly improves robustness and geometric fidelity in complex remote sensing scenes without altering the backbone architecture.

Zhe Dong, Yuzhe Sun, Haochen Jiang et al.

Remote sensing images inevitably suffer from various degradation factors during acquisition, including atmospheric interference, sensor limitations, and imaging conditions. These complex and heterogeneous degradations pose severe challenges to image quality and downstream interpretation tasks. Addressing limitations of existing all-in-one restoration methods that overly rely on implicit feature representations and lack explicit modeling of degradation physics, this paper proposes Physics-Guided Degradation-Adaptive Experts (PhyDAE). The method employs a two-stage cascaded architecture transforming degradation information from implicit features into explicit decision signals, enabling precise identification and differentiated processing of multiple heterogeneous degradations including haze, noise, blur, and low-light conditions. The model incorporates progressive degradation mining and exploitation mechanisms, where the Residual Manifold Projector (RMP) and Frequency-Aware Degradation Decomposer (FADD) comprehensively analyze degradation characteristics from manifold geometry and frequency perspectives. Physics-aware expert modules and temperature-controlled sparse activation strategies are introduced to enhance computational efficiency while ensuring imaging physics consistency. Extensive experiments on three benchmark datasets (MD-RSID, MD-RRSHID, and MDRS-Landsat) demonstrate that PhyDAE achieves superior performance across all four restoration tasks, comprehensively outperforming state-of-the-art methods. Notably, PhyDAE substantially improves restoration quality while achieving significant reductions in parameter count and computational complexity, resulting in remarkable efficiency gains compared to mainstream approaches and achieving optimal balance between performance and efficiency. Code is available at https://github.com/HIT-SIRS/PhyDAE.

Cristian-Ioan Blaga, Paul Suganthan, Sahil Dua et al.

Despite advances in multimodal learning, challenging benchmarks for mixed-modal image retrieval that combines visual and textual information are lacking. This paper introduces a novel benchmark to rigorously evaluate image retrieval that demands deep cross-modal contextual understanding. We present two new datasets: the Entity Image Dataset (EI), providing canonical images for Wikipedia entities, and the Mixed-Modal Image Retrieval Dataset (MMIR), derived from the WIT dataset. The MMIR benchmark features two challenging query types requiring models to ground textual descriptions in the context of provided visual entities: single entity-image queries (one entity image with descriptive text) and multi-entity-image queries (multiple entity images with relational text). We empirically validate the benchmark's utility as both a training corpus and an evaluation set for mixed-modal retrieval. The quality of both datasets is further affirmed through crowd-sourced human annotations. The datasets are accessible through the GitHub page: https://github.com/google-research-datasets/wit-retrieval.

Jinhyuk Lee, Feiyang Chen, Sahil Dua et al.

In this report, we introduce Gemini Embedding, a state-of-the-art embedding model leveraging the power of Gemini, Google's most capable large language model. Capitalizing on Gemini's inherent multilingual and code understanding capabilities, Gemini Embedding produces highly generalizable embeddings for text spanning numerous languages and textual modalities. The representations generated by Gemini Embedding can be precomputed and applied to a variety of downstream tasks including classification, similarity, clustering, ranking, and retrieval. Evaluated on the Massive Multilingual Text Embedding Benchmark (MMTEB), which includes over one hundred tasks across 250+ languages, Gemini Embedding substantially outperforms prior state-of-the-art models, demonstrating considerable improvements in embedding quality. Achieving state-of-the-art performance across MMTEB's multilingual, English, and code benchmarks, our unified model demonstrates strong capabilities across a broad selection of tasks and surpasses specialized domain-specific models.

Henrique Schechter Vera, Sahil Dua, Biao Zhang et al.

We introduce EmbeddingGemma, a new lightweight, open text embedding model based on the Gemma 3 language model family. Our innovative training recipe strategically captures knowledge from larger models via encoder-decoder initialization and geometric embedding distillation. We improve model robustness and expressiveness with a spread-out regularizer, and ensure generalizability by merging checkpoints from varied, optimized mixtures. Evaluated on the Massive Text Embedding Benchmark (MTEB) across multilingual, English, and code domains, EmbeddingGemma (300M) achieves state-of-the-art results. Notably, it outperforms prior top models, both proprietary and open, with fewer than 500M parameters, and provides performance comparable to models double its size, offering an exceptional performance-to-cost ratio. Remarkably, this lead persists when quantizing model weights or truncating embedding outputs. This makes EmbeddingGemma particularly well-suited for low-latency and high-throughput use cases such as on-device applications. We provide ablation studies exploring our key design choices. We release EmbeddingGemma to the community to promote further research.

Biao Zhang, Fedor Moiseev, Joshua Ainslie et al.

While decoder-only large language models (LLMs) have shown impressive results, encoder-decoder models are still widely adopted in real-world applications for their inference efficiency and richer encoder representation. In this paper, we study a novel problem: adapting pretrained decoder-only LLMs to encoder-decoder, with the goal of leveraging the strengths of both approaches to achieve a more favorable quality-efficiency trade-off. We argue that adaptation not only enables inheriting the capability of decoder-only LLMs but also reduces the demand for computation compared to pretraining from scratch. We rigorously explore different pretraining objectives and parameter initialization/optimization techniques. Through extensive experiments based on Gemma 2 (2B and 9B) and a suite of newly pretrained mT5-sized models (up to 1.6B), we demonstrate the effectiveness of adaptation and the advantage of encoder-decoder LLMs. Under similar inference budget, encoder-decoder LLMs achieve comparable (often better) pretraining performance but substantially better finetuning performance than their decoder-only counterpart. For example, Gemma 2B-2B outperforms Gemma 2B by $\sim$7\% after instruction tuning. Encoder-decoder adaptation also allows for flexible combination of different-sized models, where Gemma 9B-2B significantly surpasses Gemma 2B-2B by $>$3\%. The adapted encoder representation also yields better results on SuperGLUE. We will release our checkpoints to facilitate future research.

Paul Suganthan, Fedor Moiseev, Le Yan et al.

Decoder-based transformers, while revolutionizing language modeling and scaling to immense sizes, have not completely overtaken encoder-heavy architectures in natural language processing. Specifically, encoder-only models remain dominant in tasks like classification, regression, and ranking. This is primarily due to the inherent structure of decoder-based models, which limits their direct applicability to these tasks. In this paper, we introduce Gemma Encoder, adapting the powerful Gemma decoder model to an encoder architecture, thereby unlocking its potential for a wider range of non-generative applications. To optimize the adaptation from decoder to encoder, we systematically analyze various pooling strategies, attention mechanisms, and hyperparameters (e.g., dropout rate). Furthermore, we benchmark Gemma Encoder against established approaches on the GLUE benchmarks, and MS MARCO ranking benchmark, demonstrating its effectiveness and versatility.

Zhe Dong, Yuzhe Sun, Tianzhu Liu et al.

Referring remote sensing image segmentation (RRSIS) enables the precise delineation of regions within remote sensing imagery through natural language descriptions, serving critical applications in disaster response, urban development, and environmental monitoring. Despite recent advances, current approaches face significant challenges in processing aerial imagery due to complex object characteristics including scale variations, diverse orientations, and semantic ambiguities inherent to the overhead perspective. To address these limitations, we propose DiffRIS, a novel framework that harnesses the semantic understanding capabilities of pre-trained text-to-image diffusion models for enhanced cross-modal alignment in RRSIS tasks. Our framework introduces two key innovations: a context perception adapter (CP-adapter) that dynamically refines linguistic features through global context modeling and object-aware reasoning, and a progressive cross-modal reasoning decoder (PCMRD) that iteratively aligns textual descriptions with visual regions for precise segmentation. The CP-adapter bridges the domain gap between general vision-language understanding and remote sensing applications, while PCMRD enables fine-grained semantic alignment through multi-scale feature interaction. Comprehensive experiments on three benchmark datasets-RRSIS-D, RefSegRS, and RISBench-demonstrate that DiffRIS consistently outperforms existing methods across all standard metrics, establishing a new state-of-the-art for RRSIS tasks. The significant performance improvements validate the effectiveness of leveraging pre-trained diffusion models for remote sensing applications through our proposed adaptive framework.

Zhe Dong, Yuzhe Sun, Tianzhu Liu et al.

Satellite imagery and maps, as two fundamental data modalities in remote sensing, offer direct observations of the Earth's surface and human-interpretable geographic abstractions, respectively. The task of bidirectional translation between satellite images and maps (BSMT) holds significant potential for applications in urban planning and disaster response. However, this task presents two major challenges: first, the absence of precise pixel-wise alignment between the two modalities substantially complicates the translation process; second, it requires achieving both high-level abstraction of geographic features and high-quality visual synthesis, which further elevates the technical complexity. To address these limitations, we introduce EarthMapper, a novel autoregressive framework for controllable bidirectional satellite-map translation. EarthMapper employs geographic coordinate embeddings to anchor generation, ensuring region-specific adaptability, and leverages multi-scale feature alignment within a geo-conditioned joint scale autoregression (GJSA) process to unify bidirectional translation in a single training cycle. A semantic infusion (SI) mechanism is introduced to enhance feature-level consistency, while a key point adaptive guidance (KPAG) mechanism is proposed to dynamically balance diversity and precision during inference. We further contribute CNSatMap, a large-scale dataset comprising 302,132 precisely aligned satellite-map pairs across 38 Chinese cities, enabling robust benchmarking. Extensive experiments on CNSatMap and the New York dataset demonstrate EarthMapper's superior performance, achieving significant improvements in visual realism, semantic consistency, and structural fidelity over state-of-the-art methods. Additionally, EarthMapper excels in zero-shot tasks like in-painting, out-painting and coordinate-conditional generation, underscoring its versatility.

Gemini Team, Rohan Anil, Sebastian Borgeaud et al.

This report introduces a new family of multimodal models, Gemini, that exhibit remarkable capabilities across image, audio, video, and text understanding. The Gemini family consists of Ultra, Pro, and Nano sizes, suitable for applications ranging from complex reasoning tasks to on-device memory-constrained use-cases. Evaluation on a broad range of benchmarks shows that our most-capable Gemini Ultra model advances the state of the art in 30 of 32 of these benchmarks - notably being the first model to achieve human-expert performance on the well-studied exam benchmark MMLU, and improving the state of the art in every one of the 20 multimodal benchmarks we examined. We believe that the new capabilities of the Gemini family in cross-modal reasoning and language understanding will enable a wide variety of use cases. We discuss our approach toward post-training and deploying Gemini models responsibly to users through services including Gemini, Gemini Advanced, Google AI Studio, and Cloud Vertex AI.

Zhe Dong, Andriy Mnih, George Tucker

Training models with discrete latent variables is challenging due to the high variance of unbiased gradient estimators. While low-variance reparameterization gradients of a continuous relaxation can provide an effective solution, a continuous relaxation is not always available or tractable. Dong et al. (2020) and Yin et al. (2020) introduced a performant estimator that does not rely on continuous relaxations; however, it is limited to binary random variables. We introduce a novel derivation of their estimator based on importance sampling and statistical couplings, which we extend to the categorical setting. Motivated by the construction of a stick-breaking coupling, we introduce gradient estimators based on reparameterizing categorical variables as sequences of binary variables and Rao-Blackwellization. In systematic experiments, we show that our proposed categorical gradient estimators provide state-of-the-art performance, whereas even with additional Rao-Blackwellization, previous estimators (Yin et al., 2019) underperform a simpler REINFORCE with a leave-one-out-baseline estimator (Kool et al., 2019).

Zhe Dong, Andriy Mnih, George Tucker

Training models with discrete latent variables is challenging due to the difficulty of estimating the gradients accurately. Much of the recent progress has been achieved by taking advantage of continuous relaxations of the system, which are not always available or even possible. The Augment-REINFORCE-Merge (ARM) estimator provides an alternative that, instead of relaxation, uses continuous augmentation. Applying antithetic sampling over the augmenting variables yields a relatively low-variance and unbiased estimator applicable to any model with binary latent variables. However, while antithetic sampling reduces variance, the augmentation process increases variance. We show that ARM can be improved by analytically integrating out the randomness introduced by the augmentation process, guaranteeing substantial variance reduction. Our estimator, DisARM, is simple to implement and has the same computational cost as ARM. We evaluate DisARM on several generative modeling benchmarks and show that it consistently outperforms ARM and a strong independent sample baseline in terms of both variance and log-likelihood. Furthermore, we propose a local version of DisARM designed for optimizing the multi-sample variational bound, and show that it outperforms VIMCO, the current state-of-the-art method.

Zhe Dong, Bryan A. Seybold, Kevin P. Murphy et al.

We propose an efficient inference method for switching nonlinear dynamical systems. The key idea is to learn an inference network which can be used as a proposal distribution for the continuous latent variables, while performing exact marginalization of the discrete latent variables. This allows us to use the reparameterization trick, and apply end-to-end training with stochastic gradient descent. We show that the proposed method can successfully segment time series data, including videos and 3D human pose, into meaningful ``regimes'' by using the piece-wise nonlinear dynamics.

Zhe Dong, Deniz Oktay, Ben Poole et al.

Certain biological neurons demonstrate a remarkable capability to optimally compress the history of sensory inputs while being maximally informative about the future. In this work, we investigate if the same can be said of artificial neurons in recurrent neural networks (RNNs) trained with maximum likelihood. Empirically, we find that RNNs are suboptimal in the information plane. Instead of optimally compressing past information, they extract additional information that is not relevant for predicting the future. We show that constraining past information by injecting noise into the hidden state can improve RNNs in several ways: optimality in the predictive information plane, sample quality, heldout likelihood, and downstream classification performance.

Jieren Cheng, Chen Zhang, Xiangyan Tang et al.

Distributed denial of service (DDoS) attacks have caused huge economic losses to society. They have become one of the main threats to Internet security. Most of the current detection methods based on a single feature and fixed model parameters cannot effectively detect early DDoS attacks in cloud and big data environment. In this paper, an adaptive DDoS attack detection method (ADADM) based on multiple kernel learning (MKL) is proposed. Based on the burstiness of DDoS attack flow, the distribution of addresses and the interactivity of communication, we define five features to describe the network flow characteristic. Based on the ensemble learning framework, the weight of each dimension is adaptively adjusted by increasing the inter-class mean with a gradient ascent and reducing the intra-class variance with a gradient descent, and the classifier is established to identify an early DDoS attack by training simple multiple kernel learning (SMKL) models with two characteristics including inter-class mean squared difference growth (M-SMKL) and intra-class variance descent (S-SMKL). The sliding window mechanism is used to coordinate the S-SMKL and M-SMKL to detect the early DDoS attack. The experimental results indicate that this method can detect DDoS attacks early and accurately.