arXiv subjects

Chenze Shao

Publications and source records attributed to Chenze Shao.

At least 19 recordsLinked to original sources

arXiv2026

DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation

Speculative decoding accelerates Large Language Model (LLM) inference by decoupling draft generation from target verification. While recent parallel drafters efficiently propose long token sequences in a single forward pass, they suffer from rapid acceptance decay due to a lack of inter-token dependencies. Furthermore, indiscriminately verifying these extended blocks wastes critical batch capacity on tokens with high rejection risks, severely degrading throughput in high-concurrency serving systems. We introduce DSpark, a speculative decoding framework that unifies high-throughput parallel generation with adaptive, load-aware verification. To maintain draft quality, DSpark utilizes a semi-autoregressive architecture, coupling a parallel backbone with a lightweight sequential module, to introduce intra-block dependency modeling and mitigate suffix decay. To optimize system efficiency, DSpark employs confidence-scheduled verification, dynamically tailoring the verification length for each request based on estimated prefix survival probabilities and engine-specific throughput profiles. On offline benchmarks across diverse domains, DSpark substantially improves the accepted length over state-of-the-art autoregressive and parallel drafters. When deployed within the DeepSeek-V4 serving system under live user traffic, DSpark successfully mitigates verification waste. Compared to the established production baseline (MTP-1), DSpark accelerates per-user generation speeds by 60 to 85 percent at matched throughput levels. More importantly, by preventing severe throughput degradation under strict interactivity constraints, it enables performance tiers that were previously unattainable, shifting the Pareto frontier of our serving system.

cs.AI↗

arXiv2026

DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence

We present a preview version of DeepSeek-V4 series, including two strong Mixture-of-Experts (MoE) language models -- DeepSeek-V4-Pro with 1.6T parameters (49B activated) and DeepSeek-V4-Flash with 284B parameters (13B activated) -- both supporting a context length of one million tokens. DeepSeek-V4 series incorporate several key upgrades in architecture and optimization: (1) a hybrid attention architecture that combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to improve long-context efficiency; (2) Manifold-Constrained Hyper-Connections (mHC) that enhance conventional residual connections; (3) and the Muon optimizer for faster convergence and greater training stability. We pre-train both models on more than 32T diverse and high-quality tokens, followed by a comprehensive post-training pipeline that unlocks and further enhances their capabilities. DeepSeek-V4-Pro-Max, the maximum reasoning effort mode of DeepSeek-V4-Pro, redefines the state-of-the-art for open models, outperforming its predecessors in core tasks. Meanwhile, DeepSeek-V4 series are highly efficient in long-context scenarios. In the one-million-token context setting, DeepSeek-V4-Pro requires only 27% of single-token inference FLOPs and 10% of KV cache compared with DeepSeek-V3.2. This enables us to routinely support one-million-token contexts, thereby making long-horizon tasks and further test-time scaling more feasible. The model checkpoints are available at https://huggingface.co/collections/deepseek-ai/deepseek-v4.

cs.CL↗

arXiv2026

General learned delegation by clones

Frontier language models improve with additional test-time computation, but serial reasoning or uncoordinated parallel sampling can be compute-inefficient under fixed inference budgets. We propose SELFCEST, which equips a base model with the ability to spawn same-weight clones in separate parallel contexts by agentic reinforcement learning. Training is end-to-end under a global task reward with shared-parameter rollouts, yielding a learned controller that allocates both generation and context budget across branches. Across challenging math reasoning benchmarks and long-context multi-hop QA, SELFCEST improves the accuracy-cost Pareto frontier relative to monolithic baselines at matched inference budget, and exhibits out-of-distribution generalization in both domains.

cs.AI↗

arXiv2025

Continuous Autoregressive Language Models

The efficiency of large language models (LLMs) is fundamentally limited by their sequential, token-by-token generation process. We argue that overcoming this bottleneck requires a new design axis for LLM scaling: increasing the semantic bandwidth of each generative step. To this end, we introduce Continuous Autoregressive Language Models (CALM), a paradigm shift from discrete next-token prediction to continuous next-vector prediction. CALM uses a high-fidelity autoencoder to compress a chunk of K tokens into a single continuous vector, from which the original tokens can be reconstructed with over 99.9\% accuracy. This allows us to model language as a sequence of continuous vectors instead of discrete tokens, which reduces the number of generative steps by a factor of K. The paradigm shift necessitates a new modeling toolkit; therefore, we develop a comprehensive likelihood-free framework that enables robust training, evaluation, and controllable sampling in the continuous domain. Experiments show that CALM significantly improves the performance-compute trade-off, achieving the performance of strong discrete baselines at a significantly lower computational cost. More importantly, these findings establish next-vector prediction as a powerful and scalable pathway towards ultra-efficient language models. Code: https://github.com/shaochenze/calm. Project: https://shaochenze.github.io/blog/2025/CALM.

cs.CL↗

arXiv2025

Efficient Speech Language Modeling via Energy Distance in Continuous Latent Space

We introduce SLED, an alternative approach to speech language modeling by encoding speech waveforms into sequences of continuous latent representations and modeling them autoregressively using an energy distance objective. The energy distance offers an analytical measure of the distributional gap by contrasting simulated and target samples, enabling efficient training to capture the underlying continuous autoregressive distribution. By bypassing reliance on residual vector quantization, SLED avoids discretization errors and eliminates the need for the complicated hierarchical architectures common in existing speech language models. It simplifies the overall modeling pipeline while preserving the richness of speech information and maintaining inference efficiency. Empirical results demonstrate that SLED achieves strong performance in both zero-shot and streaming speech synthesis, showing its potential for broader applications in general-purpose speech language models.

cs.CL↗

arXiv2025

Beyond Next Token Prediction: Patch-Level Training for Large Language Models

The prohibitive training costs of Large Language Models (LLMs) have emerged as a significant bottleneck in the development of next-generation LLMs. In this paper, we show that it is possible to significantly reduce the training costs of LLMs without sacrificing their performance. Specifically, we introduce patch-level training for LLMs, in which multiple tokens are aggregated into a unit of higher information density, referred to as a `patch', to serve as the fundamental text unit for training LLMs. During patch-level training, we feed the language model shorter sequences of patches and train it to predict the next patch, thereby processing the majority of the training data at a significantly reduced cost. Following this, the model continues token-level training on the remaining training data to align with the inference mode. Experiments on a diverse range of models (370M-2.7B parameters) demonstrate that patch-level training can reduce the overall training costs to 0.5$\times$, without compromising the model performance compared to token-level training. Source code: https://github.com/shaochenze/PatchTrain.

cs.CL↗

arXiv2025

Continuous Visual Autoregressive Generation via Score Maximization

Conventional wisdom suggests that autoregressive models are used to process discrete data. When applied to continuous modalities such as visual data, Visual AutoRegressive modeling (VAR) typically resorts to quantization-based approaches to cast the data into a discrete space, which can introduce significant information loss. To tackle this issue, we introduce a Continuous VAR framework that enables direct visual autoregressive generation without vector quantization. The underlying theoretical foundation is strictly proper scoring rules, which provide powerful statistical tools capable of evaluating how well a generative model approximates the true distribution. Within this framework, all we need is to select a strictly proper score and set it as the training objective to optimize. We primarily explore a class of training objectives based on the energy score, which is likelihood-free and thus overcomes the difficulty of making probabilistic predictions in the continuous space. Previous efforts on continuous autoregressive generation, such as GIVT and diffusion loss, can also be derived from our framework using other strictly proper scores. Source code: https://github.com/shaochenze/EAR.

cs.CV↗

arXiv2024

Language Generation with Strictly Proper Scoring Rules

Language generation based on maximum likelihood estimation (MLE) has become the fundamental approach for text generation. Maximum likelihood estimation is typically performed by minimizing the log-likelihood loss, also known as the logarithmic score in statistical decision theory. The logarithmic score is strictly proper in the sense that it encourages honest forecasts, where the expected score is maximized only when the model reports true probabilities. Although many strictly proper scoring rules exist, the logarithmic score is the only local scoring rule among them that depends exclusively on the probability of the observed sample, making it capable of handling the exponentially large sample space of natural text. In this work, we propose a straightforward strategy for adapting scoring rules to language generation, allowing for language modeling with any non-local scoring rules. Leveraging this strategy, we train language generation models using two classic strictly proper scoring rules, the Brier score and the Spherical score, as alternatives to the logarithmic score. Experimental results indicate that simply substituting the loss function, without adjusting other hyperparameters, can yield substantial improvements in model's generation capabilities. Moreover, these improvements can scale up to large language models (LLMs) such as LLaMA-7B and LLaMA-13B. Source code: \url{https://github.com/shaochenze/ScoringRulesLM}.

cs.CL↗

arXiv2024

Understanding and Addressing the Under-Translation Problem from the Perspective of Decoding Objective

Neural Machine Translation (NMT) has made remarkable progress over the past years. However, under-translation and over-translation remain two challenging problems in state-of-the-art NMT systems. In this work, we conduct an in-depth analysis on the underlying cause of under-translation in NMT, providing an explanation from the perspective of decoding objective. To optimize the beam search objective, the model tends to overlook words it is less confident about, leading to the under-translation phenomenon. Correspondingly, the model's confidence in predicting the End Of Sentence (EOS) diminishes when under-translation occurs, serving as a mild penalty for under-translated candidates. Building upon this analysis, we propose employing the confidence of predicting EOS as a detector for under-translation, and strengthening the confidence-based penalty to penalize candidates with a high risk of under-translation. Experiments on both synthetic and real-world data show that our method can accurately detect and rectify under-translated outputs, with minor impact on other correct translations.

cs.CL↗

arXiv2023

Non-autoregressive Machine Translation with Probabilistic Context-free Grammar

Non-autoregressive Transformer(NAT) significantly accelerates the inference of neural machine translation. However, conventional NAT models suffer from limited expression power and performance degradation compared to autoregressive (AT) models due to the assumption of conditional independence among target tokens. To address these limitations, we propose a novel approach called PCFG-NAT, which leverages a specially designed Probabilistic Context-Free Grammar (PCFG) to enhance the ability of NAT models to capture complex dependencies among output tokens. Experimental results on major machine translation benchmarks demonstrate that PCFG-NAT further narrows the gap in translation quality between NAT and AT models. Moreover, PCFG-NAT facilitates a deeper understanding of the generated sentences, addressing the lack of satisfactory explainability in neural machine translation.Code is publicly available at https://github.com/ictnlp/PCFG-NAT.

cs.CL↗

arXiv2023

Beyond MLE: Convex Learning for Text Generation

Maximum likelihood estimation (MLE) is a statistical method used to estimate the parameters of a probability distribution that best explain the observed data. In the context of text generation, MLE is often used to train generative language models, which can then be used to generate new text. However, we argue that MLE is not always necessary and optimal, especially for closed-ended text generation tasks like machine translation. In these tasks, the goal of model is to generate the most appropriate response, which does not necessarily require it to estimate the entire data distribution with MLE. To this end, we propose a novel class of training objectives based on convex functions, which enables text generation models to focus on highly probable outputs without having to estimate the entire data distribution. We investigate the theoretical properties of the optimal predicted distribution when applying convex functions to the loss, demonstrating that convex functions can sharpen the optimal distribution, thereby enabling the model to better capture outputs with high probabilities. Experiments on various text generation tasks and models show the effectiveness of our approach. It enables autoregressive models to bridge the gap between greedy and beam search, and facilitates the learning of non-autoregressive models with a maximum improvement of 9+ BLEU points. Moreover, our approach also exhibits significant impact on large language models (LLMs), substantially enhancing their generative capability on various tasks. Source code is available at \url{https://github.com/ictnlp/Convex-Learning}.

cs.CL↗

arXiv2023

Non-autoregressive Streaming Transformer for Simultaneous Translation

Simultaneous machine translation (SiMT) models are trained to strike a balance between latency and translation quality. However, training these models to achieve high quality while maintaining low latency often leads to a tendency for aggressive anticipation. We argue that such issue stems from the autoregressive architecture upon which most existing SiMT models are built. To address those issues, we propose non-autoregressive streaming Transformer (NAST) which comprises a unidirectional encoder and a non-autoregressive decoder with intra-chunk parallelism. We enable NAST to generate the blank token or repetitive tokens to adjust its READ/WRITE strategy flexibly, and train it to maximize the non-monotonic latent alignment with an alignment-based latency loss. Experiments on various SiMT benchmarks demonstrate that NAST outperforms previous strong autoregressive SiMT baselines.

cs.CL↗

arXiv2023

Fuzzy Alignments in Directed Acyclic Graph for Non-Autoregressive Machine Translation

Non-autoregressive translation (NAT) reduces the decoding latency but suffers from performance degradation due to the multi-modality problem. Recently, the structure of directed acyclic graph has achieved great success in NAT, which tackles the multi-modality problem by introducing dependency between vertices. However, training it with negative log-likelihood loss implicitly requires a strict alignment between reference tokens and vertices, weakening its ability to handle multiple translation modalities. In this paper, we hold the view that all paths in the graph are fuzzily aligned with the reference sentence. We do not require the exact alignment but train the model to maximize a fuzzy alignment score between the graph and reference, which takes captured translations in all modalities into account. Extensive experiments on major WMT benchmarks show that our method substantially improves translation performance and increases prediction confidence, setting a new state of the art for NAT on the raw training data.

cs.CL↗

arXiv2023

Viterbi Decoding of Directed Acyclic Transformer for Non-Autoregressive Machine Translation

Non-autoregressive models achieve significant decoding speedup in neural machine translation but lack the ability to capture sequential dependency. Directed Acyclic Transformer (DA-Transformer) was recently proposed to model sequential dependency with a directed acyclic graph. Consequently, it has to apply a sequential decision process at inference time, which harms the global translation accuracy. In this paper, we present a Viterbi decoding framework for DA-Transformer, which guarantees to find the joint optimal solution for the translation and decoding path under any length constraint. Experimental results demonstrate that our approach consistently improves the performance of DA-Transformer while maintaining a similar decoding speedup.

cs.CL↗

arXiv2022

Rephrasing the Reference for Non-Autoregressive Machine Translation

Non-autoregressive neural machine translation (NAT) models suffer from the multi-modality problem that there may exist multiple possible translations of a source sentence, so the reference sentence may be inappropriate for the training when the NAT output is closer to other translations. In response to this problem, we introduce a rephraser to provide a better training target for NAT by rephrasing the reference sentence according to the NAT output. As we train NAT based on the rephraser output rather than the reference sentence, the rephraser output should fit well with the NAT output and not deviate too far from the reference, which can be quantified as reward functions and optimized by reinforcement learning. Experiments on major WMT benchmarks and NAT baselines show that our approach consistently improves the translation quality of NAT. Specifically, our best variant achieves comparable performance to the autoregressive Transformer, while being 14.7 times more efficient in inference.

cs.CL↗

arXiv2022

Non-Monotonic Latent Alignments for CTC-Based Non-Autoregressive Machine Translation

Non-autoregressive translation (NAT) models are typically trained with the cross-entropy loss, which forces the model outputs to be aligned verbatim with the target sentence and will highly penalize small shifts in word positions. Latent alignment models relax the explicit alignment by marginalizing out all monotonic latent alignments with the CTC loss. However, they cannot handle non-monotonic alignments, which is non-negligible as there is typically global word reordering in machine translation. In this work, we explore non-monotonic latent alignments for NAT. We extend the alignment space to non-monotonic alignments to allow for the global word reordering and further consider all alignments that overlap with the target sentence. We non-monotonically match the alignments to the target sentence and train the latent alignment model to maximize the F1 score of non-monotonic matching. Extensive experiments on major WMT benchmarks show that our method substantially improves the translation performance of CTC-based models. Our best model achieves 30.06 BLEU on WMT14 En-De with only one-iteration decoding, closing the gap between non-autoregressive and autoregressive models.

cs.CL↗

arXiv2022

One Reference Is Not Enough: Diverse Distillation with Reference Selection for Non-Autoregressive Translation

Non-autoregressive neural machine translation (NAT) suffers from the multi-modality problem: the source sentence may have multiple correct translations, but the loss function is calculated only according to the reference sentence. Sequence-level knowledge distillation makes the target more deterministic by replacing the target with the output from an autoregressive model. However, the multi-modality problem in the distilled dataset is still nonnegligible. Furthermore, learning from a specific teacher limits the upper bound of the model capability, restricting the potential of NAT models. In this paper, we argue that one reference is not enough and propose diverse distillation with reference selection (DDRS) for NAT. Specifically, we first propose a method called SeedDiv for diverse machine translation, which enables us to generate a dataset containing multiple high-quality reference translations for each source sentence. During the training, we compare the NAT output with all references and select the one that best fits the NAT output to train the model. Experiments on widely-used machine translation benchmarks demonstrate the effectiveness of DDRS, which achieves 29.82 BLEU with only one decoding pass on WMT14 En-De, improving the state-of-the-art performance for NAT by over 1 BLEU. Source code: https://github.com/ictnlp/DDRS-NAT

cs.CL↗

arXiv2022

Overcoming Catastrophic Forgetting beyond Continual Learning: Balanced Training for Neural Machine Translation

Neural networks tend to gradually forget the previously learned knowledge when learning multiple tasks sequentially from dynamic data distributions. This problem is called \textit{catastrophic forgetting}, which is a fundamental challenge in the continual learning of neural networks. In this work, we observe that catastrophic forgetting not only occurs in continual learning but also affects the traditional static training. Neural networks, especially neural machine translation models, suffer from catastrophic forgetting even if they learn from a static training set. To be specific, the final model pays imbalanced attention to training samples, where recently exposed samples attract more attention than earlier samples. The underlying cause is that training samples do not get balanced training in each model update, so we name this problem \textit{imbalanced training}. To alleviate this problem, we propose Complementary Online Knowledge Distillation (COKD), which uses dynamically updated teacher models trained on specific data orders to iteratively provide complementary knowledge to the student model. Experimental results on multiple machine translation tasks show that our method successfully alleviates the problem of imbalanced training and achieves substantial improvements over strong baseline systems.

cs.CL↗