Yuanhang Su

Amala Sanjay Deshmukh, Kateryna Chumachenko, Tuomas Rintamaki et al. · amazon-science, nvidia

We introduce Nemotron 3 Nano Omni, the latest model in the Nemotron multimodal series and the first to natively support audio inputs alongside text, images, and video. Nemotron 3 Nano Omni delivers consistent accuracy improvements over its predecessor, Nemotron Nano V2 VL, across all modalities, enabled by advances in architecture, training data and recipes. In particular, Nemotron 3 delivers leading results in real-world document understanding, long audio-video comprehension, and agentic computer use. Built on the highly efficient Nemotron 3 Nano 30B-A3B backbone, Nemotron 3 Nano Omni further incorporates innovative multimodal token-reduction techniques to deliver substantially lower inference latency and higher throughput than other models of similar size. We are releasing model checkpoints in BF16, FP8, and FP4 formats, along with portions of the training data and codebase to facilitate further research and development.

Hanrong Ye, Chao-Han Huck Yang, Arushi Goel et al.

Advancing machine intelligence requires developing the ability to perceive across multiple modalities, much as humans sense the world. We introduce OmniVinci, an initiative to build a strong, open-source, omni-modal LLM. We carefully study the design choices across model architecture and data curation. For model architecture, we present three key innovations: (i) OmniAlignNet for strengthening alignment between vision and audio embeddings in a shared omni-modal latent space; (ii) Temporal Embedding Grouping for capturing relative temporal alignment between vision and audio signals; and (iii) Constrained Rotary Time Embedding for encoding absolute temporal information in omni-modal embeddings. We introduce a curation and synthesis pipeline that generates 24M single-modal and omni-modal conversations. We find that modalities reinforce one another in both perception and reasoning. Our model, OmniVinci, outperforms Qwen2.5-Omni with +19.05 on DailyOmni (cross-modal understanding), +1.7 on MMAR (audio), and +3.9 on Video-MME (vision), while using just 0.2T training tokens - a 6 times reduction compared to Qwen2.5-Omni's 1.2T. We finally demonstrate omni-modal advantages in downstream applications spanning robotics, medical AI, and smart factory.

Yuanhang Su, C. -C. Jay Kuo

In this work, we first analyze the memory behavior in three recurrent neural networks (RNN) cells; namely, the simple RNN (SRN), the long short-term memory (LSTM) and the gated recurrent unit (GRU), where the memory is defined as a function that maps previous elements in a sequence to the current output. Our study shows that all three of them suffer rapid memory decay. Then, to alleviate this effect, we introduce trainable scaling factors that act like an attention mechanism to adjust memory decay adaptively. The new design is called the extended LSTM (ELSTM). Finally, to design a system that is robust to previous erroneous predictions, we propose a dependent bidirectional recurrent neural network (DBRNN). Extensive experiments are conducted on different language tasks to demonstrate the superiority of the proposed ELSTM and DBRNN solutions. The ELTSM has achieved up to 30% increase in the labeled attachment score (LAS) as compared to LSTM and GRU in the dependency parsing (DP) task. Our models also outperform other state-of-the-art models such as bi-attention and convolutional sequence to sequence (convseq2seq) by close to 10% in the LAS. The code is released as an open source (https://github.com/yuanhangsu/ELSTM-DBRNN)

Yuanhang Su, Kai Fan, Nguyen Bach et al.

Unsupervised neural machine translation (UNMT) has recently achieved remarkable results with only large monolingual corpora in each language. However, the uncertainty of associating target with source sentences makes UNMT theoretically an ill-posed problem. This work investigates the possibility of utilizing images for disambiguation to improve the performance of UNMT. Our assumption is intuitively based on the invariant property of image, i.e., the description of the same visual content by different languages should be approximately similar. We propose an unsupervised multi-modal machine translation (UMNMT) framework based on the language translation cycle consistency loss conditional on the image, targeting to learn the bidirectional multi-modal translation simultaneously. Through an alternate training between multi-modal and uni-modal, our inference model can translate with or without the image. On the widely used Multi30K dataset, the experimental results of our approach are significantly better than those of the text-only UNMT on the 2016 test dataset.

Yuanhang Su, Ruiyuan Lin, C. -C. Jay Kuo

A PCA based sequence-to-vector (seq2vec) dimension reduction method for the text classification problem, called the tree-structured multi-stage principal component analysis (TMPCA) is presented in this paper. Theoretical analysis and applicability of TMPCA are demonstrated as an extension to our previous work (Su, Huang & Kuo). Unlike conventional word-to-vector embedding methods, the TMPCA method conducts dimension reduction at the sequence level without labeled training data. Furthermore, it can preserve the sequential structure of input sequences. We show that TMPCA is computationally efficient and able to facilitate sequence-based text classification tasks by preserving strong mutual information between its input and output mathematically. It is also demonstrated by experimental results that a dense (fully connected) network trained on the TMPCA preprocessed data achieves better performance than state-of-the-art fastText and other neural-network-based solutions.

Yuanhang Su, Yuzhong Huang, C. -C. Jay Kuo

A novel text data dimension reduction technique, called the tree-structured multi-linear principal component anal- ysis (TMPCA), is proposed in this work. Being different from traditional text dimension reduction methods that deal with the word-level representation, the TMPCA technique reduces the dimension of input sequences and sentences to simplify the following text classification tasks. It is shown mathematically and experimentally that the TMPCA tool demands much lower complexity (and, hence, less computing power) than the ordinary principal component analysis (PCA). Furthermore, it is demon- strated by experimental results that the support vector machine (SVM) method applied to the TMPCA-processed data achieves commensurable or better performance than the state-of-the-art recurrent neural network (RNN) approach.