We introduce Hibiki, a decoder-only model for simultaneous speech translation. Hibiki leverages a multistream language model to synchronously process source and target speech, and jointly produces text and audio tokens to perform speech-to-text and speech-to-speech translation. We furthermore address the fundamental challenge of simultaneous interpretation, which unlike its consecutive counterpart, where one waits for the end of the source utterance to start translating, adapts its flow to accumulate just enough context to produce a correct translation in real-time, chunk by chunk. To do so, we introduce a weakly-supervised method that leverages the perplexity of an off-the-shelf text translation system to identify optimal delays on a per-word basis and create aligned synthetic data. After supervised training, Hibiki performs adaptive, simultaneous speech translation with vanilla temperature sampling. On a French-English simultaneous speech translation task, Hibiki demonstrates state-of-the-art performance in translation quality, speaker fidelity and naturalness. Moreover, the simplicity of its inference process makes it compatible with batched translation and even real-time on-device deployment. We provide examples as well as models and inference code.
Recent advancements in foundation models have transformed computer vision, driving significant performance improvements across diverse domains, including digital histopathology. However, the advantages of domain-specific histopathology foundation models over general-purpose models for specialized tasks such as cell analysis remain underexplored. This study investigates the representation learning gap between these two categories by analyzing multi-level patch embeddings applied to cell instance segmentation and classification. We implement an encoder-decoder architecture with a consistent decoder and various encoders. These include convolutional, vision transformer (ViT), and hybrid encoders pre-trained on ImageNet-22K or LVD-142M, representing general-purpose foundation models. These are compared against ViT encoders from the recently released UNI, Virchow2, and Prov-GigaPath foundation models, trained on patches extracted from hundreds of thousands of histopathology whole-slide images. The decoder integrates patch embeddings from different encoder depths via skip connections to generate semantic and distance maps. These maps are then post-processed to create instance segmentation masks where each label corresponds to an individual cell and to perform cell-type classification. All encoders remain frozen during training to assess their pre-trained feature extraction capabilities. Using the PanNuke and CoNIC histopathology datasets, and the newly introduced Nissl-stained CytoDArk0 dataset for brain cytoarchitecture studies, we evaluate instance-level detection, segmentation accuracy, and cell-type classification. This study provides insights into the comparative strengths and limitations of general-purpose vs. histopathology foundation models, offering guidance for model selection in cell-focused histopathology and brain cytoarchitecture analysis workflows.
Audio-Visual Speech Recognition (AVSR) combines lip-based video with audio and can improve performance in noise, but most methods are trained only on English data. One limitation is the lack of large-scale multilingual video data, which makes it hard to train models from scratch. In this work, we propose mWhisper-Flamingo for multilingual AVSR which combines the strengths of a pre-trained audio model (Whisper) and video model (AV-HuBERT). To enable better multi-modal integration and improve the noisy multilingual performance, we introduce decoder modality dropout where the model is trained both on paired audio-visual inputs and separate audio/visual inputs. mWhisper-Flamingo achieves state-of-the-art WER on MuAViC, an AVSR dataset of 9 languages. Audio-visual mWhisper-Flamingo consistently outperforms audio-only Whisper on all languages in noisy conditions.
Chain-of-Thought (CoT) reasoning has become a powerful framework for improving complex problem-solving capabilities in Multimodal Large Language Models (MLLMs). However, the verbose nature of textual reasoning introduces significant inefficiencies. In this work, we propose $\textbf{Heima}$ (as hidden llama), an efficient reasoning framework that leverages reasoning CoTs at hidden latent space. We design the Heima Encoder to condense each intermediate CoT into a compact, higher-level hidden representation using a single thinking token, effectively minimizing verbosity and reducing the overall number of tokens required during the reasoning process. Meanwhile, we design corresponding Heima Decoder with traditional Large Language Models (LLMs) to adaptively interpret the hidden representations into variable-length textual sequence, reconstructing reasoning processes that closely resemble the original CoTs. Experimental results across diverse reasoning MLLM benchmarks demonstrate that Heima model achieves higher generation efficiency while maintaining or even better zero-shot task accuracy. Moreover, the effective reconstruction of multimodal reasoning processes with Heima Decoder validates both the robustness and interpretability of our approach.
Designing the physical encoder is crucial for accurate image reconstruction in computational imaging (CI) systems. Currently, these systems are designed via end-to-end (E2E) optimization, where the encoder is modeled as a neural network layer and is jointly optimized with the decoder. However, the performance of E2E optimization is significantly reduced by the physical constraints imposed on the encoder. Also, since the E2E learns the parameters of the encoder by backpropagating the reconstruction error, it does not promote optimal intermediate outputs and suffers from gradient vanishing. To address these limitations, we reinterpret the concept of knowledge distillation (KD) for designing a physically constrained CI system by transferring the knowledge of a pretrained, less-constrained CI system. Our approach involves three steps: (1) Given the original CI system (student), a teacher system is created by relaxing the constraints on the student's encoder. (2) The teacher is optimized to solve a less-constrained version of the student's problem. (3) The teacher guides the training of the student through two proposed knowledge transfer functions, targeting both the encoder and the decoder feature space. The proposed method can be employed to any imaging modality since the relaxation scheme and the loss functions can be adapted according to the physical acquisition and the employed decoder. This approach was validated on three representative CI modalities: magnetic resonance, single-pixel, and compressive spectral imaging. Simulations show that a teacher system with an encoder that has a structure similar to that of the student encoder provides effective guidance. Our approach achieves significantly improved reconstruction performance and encoder design, outperforming both E2E optimization and traditional non-data-driven encoder designs.
Variational Autoencoders (VAEs) are essential tools in generative modeling and image reconstruction, with their performance heavily influenced by the encoder-decoder architecture. This study aims to improve the quality of reconstructed images by enhancing their resolution and preserving finer details, particularly when working with low-resolution inputs (16x16 pixels), where traditional VAEs often yield blurred or in-accurate results. To address this, we propose a hybrid model that combines quantum computing techniques in the VAE encoder with convolutional neural networks (CNNs) in the decoder. By upscaling the resolution from 16x16 to 32x32 during the encoding process, our approach evaluates how the model reconstructs images with enhanced resolution while maintaining key features and structures. This method tests the model's robustness in handling image reconstruction and its ability to preserve essential details despite training on lower-resolution data. We evaluate our proposed down sampling filter for Quantum VAE (Q-VAE) on the MNIST and USPS datasets and compare it with classical VAEs and a variant called Classical Direct Passing VAE (CDP-VAE), which uses windowing pooling filters in the encoding process. Performance is assessed using metrics such as the Frechet Inception Distance (FID) and Mean Squared Error (MSE), which measure the fidelity of reconstructed images. Our results demonstrate that the Q-VAE consistently outperforms both the Classical VAE and CDP-VAE, achieving significantly lower FID and MSE scores. Additionally, CDP-VAE yields better performance than C-VAE. These findings highlight the potential of quantum-enhanced VAEs to improve image reconstruction quality by enhancing resolution and preserving essential features, offering a promising direction for future applications in computer vision and synthetic data generation.
We present FireRedASR, a family of large-scale automatic speech recognition (ASR) models for Mandarin, designed to meet diverse requirements in superior performance and optimal efficiency across various applications. FireRedASR comprises two variants: FireRedASR-LLM: Designed to achieve state-of-the-art (SOTA) performance and to enable seamless end-to-end speech interaction. It adopts an Encoder-Adapter-LLM framework leveraging large language model (LLM) capabilities. On public Mandarin benchmarks, FireRedASR-LLM (8.3B parameters) achieves an average Character Error Rate (CER) of 3.05%, surpassing the latest SOTA of 3.33% with an 8.4% relative CER reduction (CERR). It demonstrates superior generalization capability over industrial-grade baselines, achieving 24%-40% CERR in multi-source Mandarin ASR scenarios such as video, live, and intelligent assistant. FireRedASR-AED: Designed to balance high performance and computational efficiency and to serve as an effective speech representation module in LLM-based speech models. It utilizes an Attention-based Encoder-Decoder (AED) architecture. On public Mandarin benchmarks, FireRedASR-AED (1.1B parameters) achieves an average CER of 3.18%, slightly worse than FireRedASR-LLM but still outperforming the latest SOTA model with over 12B parameters. It offers a more compact size, making it suitable for resource-constrained applications. Moreover, both models exhibit competitive results on Chinese dialects and English speech benchmarks and excel in singing lyrics recognition. To advance research in speech processing, we release our models and inference code at https://github.com/FireRedTeam/FireRedASR.
The reasoning abilities of Large Language Models (LLMs) can be improved by structurally denoising their weights, yet existing techniques primarily focus on denoising the feed-forward network (FFN) of the transformer block, and can not efficiently utilise the Multi-head Attention (MHA) block, which is the core of transformer architectures. To address this issue, we propose a novel intuitive framework that, at its very core, performs MHA compression through a multi-head tensorisation process and the Tucker decomposition. This enables both higher-dimensional structured denoising and compression of the MHA weights, by enforcing a shared higher-dimensional subspace across the weights of the multiple attention heads. We demonstrate that this approach consistently enhances the reasoning capabilities of LLMs across multiple benchmark datasets, and for both encoder-only and decoder-only architectures, while achieving compression rates of up to $\sim 250$ times in the MHA weights, all without requiring any additional data, training, or fine-tuning. Furthermore, we show that the proposed method can be seamlessly combined with existing FFN-only-based denoising techniques to achieve further improvements in LLM reasoning performance.
The dominance of large decoder-only language models has overshadowed encoder-decoder architectures, despite their fundamental efficiency advantages in sequence processing. For small language models (SLMs) - those with 1 billion parameters or fewer - our systematic analysis across GPU, CPU, and NPU platforms reveals that encoder-decoder architectures achieve 47% lower first-token latency and 4.7x higher throughput compared to decoder-only models on edge devices. These gains may be attributed to encoder-decoder's one-time input processing and efficient separation of understanding and generation phases. We introduce a novel knowledge distillation framework that enables encoder-decoder models to leverage capabilities from large scalable decoder-only teachers while preserving their architectural advantages, achieving up to 6 average performance points improvement across diverse tasks, with significant gains in asymmetric sequence tasks where input and output distributions can benefit from different processing approaches. When combined with modern advances like Rotary Positional Embeddings (RoPE) and Vision encoders, our systematic investigation demonstrates that encoder-decoder architectures provide a more practical path toward deploying capable language models in resource-constrained environments. Our findings challenge the prevailing trend toward decoder-only scaling, showing that architectural choices become increasingly crucial as parameter budgets decrease, particularly for on-device and edge deployments where computational efficiency is paramount.
This work introduces a novel, fully differentiable linear-time complexity transformer decoder and a transformer decoder to correct 5G New Radio (NR) LDPC. We propose a scalable approach to decode linear block codes with $O(n)$ complexity rather than $O(n^2)$ for regular transformers. The architectures' performances are compared to Belief Propagation (BP), the production-level decoding algorithm used for 5G New Radio (NR) LDPC codes. We achieve bit error rate performance that matches a regular Transformer decoder and surpases one iteration BP, also achieving competitive time performance against BP, even for larger block codes. We utilize Sionna, Nvidia's 5G & 6G physical layer research software, for reproducible results.
Recent approaches in hierarchical text classification (HTC) rely on the capabilities of a pre-trained transformer model and exploit the label semantics and a graph encoder for the label hierarchy. In this paper, we introduce an effective hierarchical text classifier RADAr (Transformer-based Autoregressive Decoder Architecture) that is based only on an off-the-shelf RoBERTa transformer to process the input and a custom autoregressive decoder with two decoder layers for generating the classification output. Thus, unlike existing approaches for HTC, the encoder of RADAr has no explicit encoding of the label hierarchy and the decoder solely relies on the label sequences of the samples observed during training. We demonstrate on three benchmark datasets that RADAr achieves results competitive to the state of the art with less training and inference time. Our model consistently performs better when organizing the label sequences from children to parents versus the inverse, as done in existing HTC approaches. Our experiments show that neither the label semantics nor an explicit graph encoder for the hierarchy is needed. This has strong practical implications for HTC as the architecture has fewer requirements and provides a speed-up by a factor of 2 at inference time. Moreover, training a separate decoder from scratch in conjunction with fine-tuning the encoder allows future researchers and practitioners to exchange the encoder part as new models arise. The source code is available at https://github.com/yousef-younes/RADAr.
Modeling real-world spatio-temporal data is exceptionally difficult due to inherent high dimensionality, measurement noise, partial observations, and often expensive data collection procedures. In this paper, we present Sparse Identification of Nonlinear Dynamics with SHallow REcurrent Decoder networks (SINDy-SHRED), a method to jointly solve the sensing and model identification problems with simple implementation, efficient computation, and robust performance. SINDy-SHRED uses Gated Recurrent Units to model the temporal sequence of sparse sensor measurements along with a shallow decoder network to reconstruct the full spatio-temporal field from the latent state space. Our algorithm introduces a SINDy-based regularization for which the latent space progressively converges to a SINDy-class functional, provided the projection remains within the set. In restricting SINDy to a linear model, a Koopman-SHRED model is generated. SINDy-SHRED (i) learns a symbolic and interpretable generative model of a parsimonious and low-dimensional latent space for the complex spatio-temporal dynamics, (ii) discovers new physics models even for well-known physical systems, (iii) achieves provably robust convergence with an observed globally convex loss landscape, and (iv) achieves superior accuracy, data efficiency, and training time, all with fewer model parameters. We conduct systematic experimental studies on PDE data such as turbulent flows, real-world sensor measurements for sea surface temperature, and direct video data. The interpretable SINDy and Koopman models of latent state dynamics enable stable and accurate long-term video predictions, outperforming all current baseline deep learning models in accuracy, training time, and data requirements, including Convolutional LSTM, PredRNN, ResNet, and SimVP.
Multi-modal retrieval-augmented Question Answering (MRAQA), integrating text and images, has gained significant attention in information retrieval (IR) and natural language processing (NLP). Traditional ranking methods rely on small encoder-based language models, which are incompatible with modern decoder-based generative large language models (LLMs) that have advanced various NLP tasks. To bridge this gap, we propose RAMQA, a unified framework combining learning-to-rank methods with generative permutation-enhanced ranking techniques. We first train a pointwise multi-modal ranker using LLaVA as the backbone. Then, we apply instruction tuning to train a LLaMA model for re-ranking the top-k documents using an innovative autoregressive multi-task learning approach. Our generative ranking model generates re-ranked document IDs and specific answers from document candidates in various permutations. Experiments on two MRAQA benchmarks, WebQA and MultiModalQA, show significant improvements over strong baselines, highlighting the effectiveness of our approach. Code and data are available at: https://github.com/TonyBY/RAMQA
Monocular depth estimation (MDE) is a challenging task in computer vision, often hindered by the cost and scarcity of high-quality labeled datasets. We tackle this challenge using auxiliary datasets from related vision tasks for an alternating training scheme with a shared decoder built on top of a pre-trained vision foundation model, while giving a higher weight to MDE. Through extensive experiments we demonstrate the benefits of incorporating various in-domain auxiliary datasets and tasks to improve MDE quality on average by ~11%. Our experimental analysis shows that auxiliary tasks have different impacts, confirming the importance of task selection, highlighting that quality gains are not achieved by merely adding data. Remarkably, our study reveals that using semantic segmentation datasets as Multi-Label Dense Classification (MLDC) often results in additional quality gains. Lastly, our method significantly improves the data efficiency for the considered MDE datasets, enhancing their quality while reducing their size by at least 80%. This paves the way for using auxiliary data from related tasks to improve MDE quality despite limited availability of high-quality labeled data. Code is available at https://jugit.fz-juelich.de/ias-8/mdeaux.
In this paper, we present an encoder-decoder model leveraging Flan-T5 for post-Automatic Speech Recognition (ASR) Generative Speech Error Correction (GenSEC), and we refer to it as FlanEC. We explore its application within the GenSEC framework to enhance ASR outputs by mapping n-best hypotheses into a single output sentence. By utilizing n-best lists from ASR models, we aim to improve the linguistic correctness, accuracy, and grammaticality of final ASR transcriptions. Specifically, we investigate whether scaling the training data and incorporating diverse datasets can lead to significant improvements in post-ASR error correction. We evaluate FlanEC using the HyPoradise dataset, providing a comprehensive analysis of the model's effectiveness in this domain. Furthermore, we assess the proposed approach under different settings to evaluate model scalability and efficiency, offering valuable insights into the potential of instruction-tuned encoder-decoder models for this task.
Video Moment Retrieval and Highlight Detection aim to find corresponding content in the video based on a text query. Existing models usually first use contrastive learning methods to align video and text features, then fuse and extract multimodal information, and finally use a Transformer Decoder to decode multimodal information. However, existing methods face several issues: (1) Overlapping semantic information between different samples in the dataset hinders the model's multimodal aligning performance; (2) Existing models are not able to efficiently extract local features of the video; (3) The Transformer Decoder used by the existing model cannot adequately decode multimodal features. To address the above issues, we proposed the LD-DETR model for Video Moment Retrieval and Highlight Detection tasks. Specifically, we first distilled the similarity matrix into the identity matrix to mitigate the impact of overlapping semantic information. Then, we designed a method that enables convolutional layers to extract multimodal local features more efficiently. Finally, we fed the output of the Transformer Decoder back into itself to adequately decode multimodal information. We evaluated LD-DETR on four public benchmarks and conducted extensive experiments to demonstrate the superiority and effectiveness of our approach. Our model outperforms the State-Of-The-Art models on QVHighlight, Charades-STA and TACoS datasets. Our code is available at https://github.com/qingchen239/ld-detr.
Video deblurring is essential task for autonomous driving, facial recognition, and security surveillance. Traditional methods directly estimate motion blur kernels, often introducing artifacts and leading to poor results. Recent approaches utilize the detection of sharp frames within video sequences to enhance deblurring. However, existing datasets rely on fixed number of sharp frames, which may be too restrictive for some applications and may introduce a bias during model training. To address these limitations and enhance domain adaptability, this work first introduces GoPro Random Sharp (GoProRS), a new dataset where the the frequency of sharp frames within the sequence is customizable, allowing more diverse training and testing scenarios. Furthermore, it presents a novel video deblurring model, called SPEINet, that integrates sharp frame features into blurry frame reconstruction through an attention-based encoder-decoder architecture, a lightweight yet robust sharp frame detection and an edge extraction phase. Extensive experimental results demonstrate that SPEINet outperforms state-of-the-art methods across multiple datasets, achieving an average of +3.2% PSNR improvement over recent techniques. Given such promising results, we believe that both the proposed model and dataset pave the way for future advancements in video deblurring based on the detection of sharp frames.
The standard language modeling (LM) loss by itself has been shown to be inadequate for effective dialogue modeling. As a result, various training approaches, such as auxiliary loss functions and leveraging human feedback, are being adopted to enrich open-domain dialogue systems. One such auxiliary loss function is Bag-of-Words (BoW) loss, defined as the cross-entropy loss for predicting all the words/tokens of the next utterance. In this work, we propose a novel auxiliary loss named Bag-of-Keywords (BoK) loss to capture the central thought of the response through keyword prediction and leverage it to enhance the generation of meaningful and interpretable responses in open-domain dialogue systems. BoK loss upgrades the BoW loss by predicting only the keywords or critical words/tokens of the next utterance, intending to estimate the core idea rather than the entire response. We incorporate BoK loss in both encoder-decoder (T5) and decoder-only (DialoGPT) architecture and train the models to minimize the weighted sum of BoK and LM (BoK-LM) loss. We perform our experiments on two popular open-domain dialogue datasets, DailyDialog and Persona-Chat. We show that the inclusion of BoK loss improves the dialogue generation of backbone models while also enabling post-hoc interpretability. We also study the effectiveness of BoK-LM loss as a reference-free metric and observe comparable performance to the state-of-the-art metrics on various dialogue evaluation datasets.
Text-to-SQL enables users to interact with databases through natural language, simplifying the retrieval and synthesis of information. Despite the success of large language models (LLMs) in converting natural language questions into SQL queries, their broader adoption is limited by two main challenges: achieving robust generalization across diverse queries and ensuring interpretative confidence in their predictions. To tackle these issues, our research investigates the integration of selective classifiers into Text-to-SQL systems. We analyse the trade-off between coverage and risk using entropy based confidence estimation with selective classifiers and assess its impact on the overall performance of Text-to-SQL models. Additionally, we explore the models' initial calibration and improve it with calibration techniques for better model alignment between confidence and accuracy. Our experimental results show that encoder-decoder T5 is better calibrated than in-context-learning GPT 4 and decoder-only Llama 3, thus the designated external entropy-based selective classifier has better performance. The study also reveal that, in terms of error detection, selective classifier with a higher probability detects errors associated with irrelevant questions rather than incorrect query generations.
Code-switching, the alternation of languages within a single discourse, presents a significant challenge for Automatic Speech Recognition. Despite the unique nature of the task, performance is commonly measured with established metrics such as Word-Error-Rate (WER). However, in this paper, we question whether these general metrics accurately assess performance on code-switching. Specifically, using both Connectionist-Temporal-Classification and Encoder-Decoder models, we show fine-tuning on non-code-switched data from both matrix and embedded language improves classical metrics on code-switching test sets, although actual code-switched words worsen (as expected). Therefore, we propose Point-of-Interest Error Rate (PIER), a variant of WER that focuses only on specific words of interest. We instantiate PIER on code-switched utterances and show that this more accurately describes the code-switching performance, showing huge room for improvement in future work. This focused evaluation allows for a more precise assessment of model performance, particularly in challenging aspects such as inter-word and intra-word code-switching.