The DeepSeek models have shown exceptional performance in general scene understanding, question-answering (QA), and text generation tasks, owing to their efficient training paradigm and strong reasoning capabilities. In this study, we investigate the dialogue capabilities of the DeepSeek model in robotic surgery scenarios, focusing on tasks such as Single Phrase QA, Visual QA, and Detailed Description. The Single Phrase QA tasks further include sub-tasks such as surgical instrument recognition, action understanding, and spatial position analysis. We conduct extensive evaluations using publicly available datasets, including EndoVis18 and CholecT50, along with their corresponding dialogue data. Our empirical study shows that, compared to existing general-purpose multimodal large language models, DeepSeek-VL2 performs better on complex understanding tasks in surgical scenes. Additionally, although DeepSeek-V3 is purely a language model, we find that when image tokens are directly inputted, the model demonstrates better performance on single-sentence QA tasks. However, overall, the DeepSeek models still fall short of meeting the clinical requirements for understanding surgical scenes. Under general prompts, DeepSeek models lack the ability to effectively analyze global surgical concepts and fail to provide detailed insights into surgical scenarios. Based on our observations, we argue that the DeepSeek models are not ready for vision-language tasks in surgical contexts without fine-tuning on surgery-specific datasets.
Multilingual Natural Language Generation (NLG) is challenging due to the lack of training data for low-resource languages. However, some low-resource languages have up to tens of millions of speakers globally, making it important to improve NLG tools for them. Table-to-Text NLG is an excellent measure of models' reasoning abilities but is very challenging in the multilingual setting. System outputs are often not attributable, or faithful, to the data in the source table. Intermediate planning techniques like Question-Answer (QA) blueprints have been shown to improve attributability on summarisation tasks. This work explores whether QA blueprints make multilingual Table-to-Text outputs more attributable to the input tables. This paper extends the challenging multilingual Table-to-Text dataset, TaTA, which includes African languages, with QA blueprints. Sequence-to-sequence language models are then finetuned on this dataset, with and without blueprints. Results show that QA blueprints improve performance for models finetuned and evaluated only on English examples, but do not demonstrate gains in the multilingual setting. This is due to inaccuracies in machine translating the blueprints from English into target languages when generating the training data, and models failing to rely closely on the blueprints they generate. An in-depth analysis is conducted on why this is challenging.
Large language models (LLMs) excel on a variety of reasoning benchmarks, but previous studies suggest they sometimes struggle to generalize to unseen questions, potentially due to over-reliance on memorized training examples. However, the precise conditions under which LLMs switch between reasoning and memorization during text generation remain unclear. In this work, we provide a mechanistic understanding of LLMs' reasoning-memorization dynamics by identifying a set of linear features in the model's residual stream that govern the balance between genuine reasoning and memory recall. These features not only distinguish reasoning tasks from memory-intensive ones but can also be manipulated to causally influence model performance on reasoning tasks. Additionally, we show that intervening in these reasoning features helps the model more accurately activate the most relevant problem-solving capabilities during answer generation. Our findings offer new insights into the underlying mechanisms of reasoning and memory in LLMs and pave the way for the development of more robust and interpretable generative AI systems.
Various layer-skipping methods have been proposed to accelerate token generation in large language models (LLMs). However, they have overlooked a fundamental question: How do computational demands vary across the generation of different tokens? In this work, we introduce FlexiDepth, a method that dynamically adjusts the number of Transformer layers used in text generation. By incorporating a plug-in router and adapter, FlexiDepth enables adaptive layer-skipping in LLMs without modifying their original parameters. Introducing FlexiDepth to Llama-3-8B model achieves layer skipping of 8 layers out of 32, and meanwhile maintains the full 100\% benchmark performance. Experimental results with FlexiDepth demonstrate that computational demands in LLMs significantly vary based on token type. Specifically, generating repetitive tokens or fixed phrases requires fewer layers, whereas producing tokens involving computation or high uncertainty requires more layers. Interestingly, this adaptive allocation pattern aligns with human intuition. To advance research in this area, we open sourced FlexiDepth and a dataset documenting FlexiDepth's layer allocation patterns for future exploration.
Data plays a pivotal role in Text-Based Person Retrieval (TBPR) research. Mainstream research paradigm necessitates real-world person images with manual textual annotations for training models, posing privacy-sensitive and labor-intensive issues. Several pioneering efforts explore synthetic data for TBPR but still rely on real data, keeping the aforementioned issues and also resulting in diversity-deficient issue in synthetic datasets, thus impacting TBPR performance. Moreover, these works tend to explore synthetic data for TBPR through limited perspectives, leading to exploration-restricted issue. In this paper, we conduct an empirical study to explore the potential of synthetic data for TBPR, highlighting three key aspects. (1) We propose an inter-class image generation pipeline, in which an automatic prompt construction strategy is introduced to guide generative Artificial Intelligence (AI) models in generating various inter-class images without reliance on original data. (2) We develop an intra-class image augmentation pipeline, in which the generative AI models are applied to further edit the images for obtaining various intra-class images. (3) Building upon the proposed pipelines and an automatic text generation pipeline, we explore the effectiveness of synthetic data in diverse scenarios through extensive experiments. Additionally, we experimentally investigate various noise-robust learning strategies to mitigate the inherent noise in synthetic data. We will release the code, along with the synthetic large-scale dataset generated by our pipelines, which are expected to advance practical TBPR research.
Advances in hardware and language model architecture have spurred a revolution in natural language generation. However, autoregressive models compute probability distributions over next-token choices, and sampling from these distributions, known as decoding, has received significantly less attention than other design choices. Existing decoding strategies are largely based on heuristics, resulting in methods that are hard to apply or improve in a principled manner. We develop the theory of decoding strategies for language models by expressing popular decoding algorithms as equilibrium states in the language of ergodic theory and stating the functions they optimize. Using this, we analyze the effect of the local normalization step of top-k, nucleus, and temperature sampling, used to make probabilities sum to one. We argue that local normalization distortion is a fundamental defect of decoding strategies and quantify the size of this distortion and its effect on mathematical proxies for the quality and diversity of generated text. Contrary to the prevailing explanation, we argue that the major cause of the under-performance of top-k sampling relative to nucleus sampling is local normalization distortion. This yields conclusions for the future design of decoding algorithms and the detection of machine-generated text.
Aligning large language models (LLMs) with human values and safety constraints is challenging, especially when objectives like helpfulness, truthfulness, and avoidance of harm conflict. Reinforcement Learning from Human Feedback (RLHF) has achieved notable success in steering models, but is complex and can be unstable. Recent approaches such as Direct Preference Optimization (DPO) simplify preference-based fine-tuning but may introduce bias or trade-off certain objectives~\cite{dpo}. In this work, we propose a Group Relative Policy Optimization (GRPO) framework with a multi-label reward regression model to achieve safe and aligned language generation. The GRPO algorithm optimizes a policy by comparing groups of sampled responses, eliminating the need for a separate value critic and improving training efficiency~\cite{grpo}. We train a reward model to predict multiple alignment scores (e.g., safety, helpfulness, etc.), which are combined into a single reward signal. We provide a theoretical derivation for using this learned multi-aspect reward within GRPO and discuss its advantages and limitations. Empirically, our approach improves all the safety and quality metrics evaluated in language generation tasks on model scales (0.5B, 7B, and 14B parameters), demonstrating a robust balance of objectives. We compare GRPO to PPO-based RLHF and DPO, highlighting that GRPO achieves alignment with significantly lower computational cost and explicit multi-objective handling. \textbf{We will open-source all trained models at https://huggingface.co/hydroxai.
Large language models (LLMs) demonstrate strong language generation capabilities but often struggle with structured reasoning, leading to inconsistent or suboptimal problem-solving. To mitigate this limitation, Guilford's Structure of Intellect (SOI) model - a foundational framework from intelligence theory - is leveraged as the basis for cognitive prompt engineering. The SOI model categorizes cognitive operations such as pattern recognition, memory retrieval, and evaluation, offering a systematic approach to enhancing LLM reasoning and decision-making. This position paper presents a novel cognitive prompting approach for enforcing SOI-inspired reasoning for improving clarity, coherence, and adaptability in model responses.
In this work, we undertake the challenge of augmenting the existing generative capabilities of pre-trained text-only large language models (LLMs) with multi-modal generation capability while satisfying two core constraints: C1 preserving the preservation of original language generative capabilities with negligible performance degradation, and C2 adhering to a small parameter budget to learn the new modality, ensuring scalability and efficiency. In contrast to current approaches that add dedicated modules, thereby significantly increasing the parameter count, we propose a method that leverages the underutilized capacity inherent in deep models. Specifically, we exploit the parameter redundancy within Mixture-of-Experts (MoEs) as a source of additional capacity for learning a new modality, enabling better parameter efficiency (C1). Moreover, we preserve the original language generation capabilities by applying low-rank adaptation exclusively to the tokens of the new modality (C2). Furthermore, we introduce a novel parameter initialization scheme based on the Gromov-Wasserstein distance to improve convergence and training stability. Through an extensive analysis of the routing mechanism, we uncover the emergence of modality-specific pathways and decreased redundancy within the experts that can efficiently unlock multi-modal generative capabilities. Overall, our method can be seamlessly applied to a wide range of contemporary LLMs, providing a new pathway for transitioning from uni-modal to multi-modal architectures.
Recent years have witnessed the rapid progression of deep learning, pushing us closer to the realization of AGI (Artificial General Intelligence). Probabilistic modeling is critical to many of these advancements, which provides a foundational framework for capturing data distributions. However, as the scale and complexity of AI applications grow, traditional probabilistic modeling faces escalating challenges, such as high-dimensional parameter spaces, heterogeneous data sources, and evolving real-world requirements often render classical approaches insufficiently flexible. This paper proposes a novel concept, Probability Engineering, which treats the already-learned probability distributions within deep learning as engineering artifacts. Rather than merely fitting or inferring distributions, we actively modify and reinforce them to better address the diverse and evolving demands of modern AI. Specifically, Probability Engineering introduces novel techniques and constraints to refine existing probability distributions, improving their robustness, efficiency, adaptability, or trustworthiness. We showcase this paradigm through a series of applications spanning Bayesian deep learning, Edge AI (including federated learning and knowledge distillation), and Generative AI (such as text-to-image generation with diffusion models and high-quality text generation with large language models). These case studies demonstrate how probability distributions once treated as static objects can be engineered to meet the diverse and evolving requirements of large-scale, data-intensive, and trustworthy AI systems. By systematically expanding and strengthening the role of probabilistic modeling, Probability Engineering paves the way for more robust, adaptive, efficient, and trustworthy deep learning solutions in today's fast-growing AI era.
The emergence of large language models (LLMs) has significantly advanced the development of natural language processing (NLP), especially in text generation tasks like question answering. However, model hallucinations remain a major challenge in natural language generation (NLG) tasks due to their complex causes. We systematically expand on the causes of factual hallucinations from the perspective of knowledge shortcuts, analyzing hallucinations arising from correct and defect-free data and demonstrating that knowledge-shortcut hallucinations are prevalent in generative models. To mitigate this issue, we propose a high similarity pruning algorithm at the data preprocessing level to reduce spurious correlations in the data. Additionally, we design a specific detection method for knowledge-shortcut hallucinations to evaluate the effectiveness of our mitigation strategy. Experimental results show that our approach effectively reduces knowledge-shortcut hallucinations, particularly in fine-tuning tasks, without negatively impacting model performance in question answering. This work introduces a new paradigm for mitigating specific hallucination issues in generative models, enhancing their robustness and reliability in real-world applications.
This article provides a brief overview of the field of Natural Language Generation. The term Natural Language Generation (NLG), in its broadest definition, refers to the study of systems that verbalize some form of information through natural language. That information could be stored in a large database or knowledge graph (in data-to-text applications), but NLG researchers may also study summarisation (text-to-text) or image captioning (image-to-text), for example. As a subfield of Natural Language Processing, NLG is closely related to other sub-disciplines such as Machine Translation (MT) and Dialog Systems. Some NLG researchers exclude MT from their definition of the field, since there is no content selection involved where the system has to determine what to say. Conversely, dialog systems do not typically fall under the header of Natural Language Generation since NLG is just one component of dialog systems (the others being Natural Language Understanding and Dialog Management). However, with the rise of Large Language Models (LLMs), different subfields of Natural Language Processing have converged on similar methodologies for the production of natural language and the evaluation of automatically generated text.
Computational efficiency has remained a critical consideration in scaling high-capacity language models, with inference latency and resource consumption presenting significant constraints on real-time applications. The study has introduced a structured latency perturbation mechanism that modifies computational pathways through recursive state induction, enabling dynamic suppression of redundant activations while preserving generative fidelity. A formal mathematical framework has been established to describe recursive perturbations, ensuring that modifications remain adaptive rather than statically imposed. Experiments have demonstrated that applying recursive state adjustments reduces inference latency across varying sequence lengths, with longer text generations benefiting from cumulative efficiency improvements. Comparative evaluations against structured pruning and quantization have indicated that latency gains can be achieved without compromising token retention or memory utilization. The analysis of computational overhead has suggested that selectively suppressing redundant activations contributes to improved power efficiency, particularly in scenarios requiring extended text generation. An assessment of linguistic stability has shown that token-level consistency remains largely intact under controlled perturbation thresholds, reinforcing the viability of structural latency modifications as an alternative to weight-centric optimization techniques. The results have supported the hypothesis that recursive state induction offers an effective method for reducing computational complexity without requiring architectural modifications or external augmentation.
Adapting Large Language Models (LLMs) to downstream tasks using Reinforcement Learning (RL) has proven to be an effective approach. However, LLMs do not inherently define the structure of an agent for RL training, particularly in terms of defining the action space. This paper studies learning a compact latent action space to enhance the controllability and exploration of RL for LLMs. We propose Controlling Large Language Models with Latent Actions (CoLA), a framework that integrates a latent action space into pre-trained LLMs. We apply CoLA to the Llama-3.1-8B model. Our experiments demonstrate that, compared to RL with token-level actions, CoLA's latent action enables greater semantic diversity in text generation. For enhancing downstream tasks, we show that CoLA with RL achieves a score of 42.4 on the math500 benchmark, surpassing the baseline score of 38.2, and reaches 68.2 when augmented with a Monte Carlo Tree Search variant. Furthermore, CoLA with RL consistently improves performance on agent-based tasks without degrading the pre-trained LLM's capabilities, unlike the baseline. Finally, CoLA reduces computation time by half in tasks involving enhanced thinking prompts for LLMs by RL. These results highlight CoLA's potential to advance RL-based adaptation of LLMs for downstream applications.
We introduce LeX-Art, a comprehensive suite for high-quality text-image synthesis that systematically bridges the gap between prompt expressiveness and text rendering fidelity. Our approach follows a data-centric paradigm, constructing a high-quality data synthesis pipeline based on Deepseek-R1 to curate LeX-10K, a dataset of 10K high-resolution, aesthetically refined 1024$\times$1024 images. Beyond dataset construction, we develop LeX-Enhancer, a robust prompt enrichment model, and train two text-to-image models, LeX-FLUX and LeX-Lumina, achieving state-of-the-art text rendering performance. To systematically evaluate visual text generation, we introduce LeX-Bench, a benchmark that assesses fidelity, aesthetics, and alignment, complemented by Pairwise Normalized Edit Distance (PNED), a novel metric for robust text accuracy evaluation. Experiments demonstrate significant improvements, with LeX-Lumina achieving a 79.81% PNED gain on CreateBench, and LeX-FLUX outperforming baselines in color (+3.18%), positional (+4.45%), and font accuracy (+3.81%). Our codes, models, datasets, and demo are publicly available.
Multimodal generative models that can understand and generate across multiple modalities are dominated by autoregressive (AR) approaches, which process tokens sequentially from left to right, or top to bottom. These models jointly handle images, text, video, and audio for various tasks such as image captioning, question answering, and image generation. In this work, we explore discrete diffusion models as a unified generative formulation in the joint text and image domain, building upon their recent success in text generation. Discrete diffusion models offer several advantages over AR models, including improved control over quality versus diversity of generated samples, the ability to perform joint multimodal inpainting (across both text and image domains), and greater controllability in generation through guidance. Leveraging these benefits, we present the first Unified Multimodal Discrete Diffusion (UniDisc) model which is capable of jointly understanding and generating text and images for a variety of downstream tasks. We compare UniDisc to multimodal AR models, performing a scaling analysis and demonstrating that UniDisc outperforms them in terms of both performance and inference-time compute, enhanced controllability, editability, inpainting, and flexible trade-off between inference time and generation quality. Code and additional visualizations are available at https://unidisc.github.io.
Long-context video modeling is essential for enabling generative models to function as world simulators, as they must maintain temporal coherence over extended time spans. However, most existing models are trained on short clips, limiting their ability to capture long-range dependencies, even with test-time extrapolation. While training directly on long videos is a natural solution, the rapid growth of vision tokens makes it computationally prohibitive. To support exploring efficient long-context video modeling, we first establish a strong autoregressive baseline called Frame AutoRegressive (FAR). FAR models temporal dependencies between continuous frames, converges faster than video diffusion transformers, and outperforms token-level autoregressive models. Based on this baseline, we observe context redundancy in video autoregression. Nearby frames are critical for maintaining temporal consistency, whereas distant frames primarily serve as context memory. To eliminate this redundancy, we propose the long short-term context modeling using asymmetric patchify kernels, which apply large kernels to distant frames to reduce redundant tokens, and standard kernels to local frames to preserve fine-grained detail. This significantly reduces the training cost of long videos. Our method achieves state-of-the-art results on both short and long video generation, providing an effective baseline for long-context autoregressive video modeling.
Token prediction stability remains a challenge in autoregressive generative models, where minor variations in early inference steps often lead to significant semantic drift over extended sequences. A structured modulation mechanism was introduced to regulate hidden state transitions, ensuring that latent representation trajectories remain aligned with prior contextual dependencies while preserving generative flexibility. The modulation framework was designed to function within transformer-based architectures, dynamically constraining representation evolution without imposing external memory dependencies or extensive architectural modifications. Empirical evaluations demonstrated that structured latent adjustments contributed to reductions in perplexity fluctuations, entropy variance, and lexical instability, improving coherence in long-form text generation. Gradient propagation stability was further analyzed, revealing that the modulation process led to smoother optimization pathways, mitigating erratic fluctuations in weight updates across successive inference steps. The computational efficiency of the modulation process was assessed, showing that its integration within transformer-based architectures introduced only marginal overhead while maintaining compatibility with existing optimization frameworks. The structured modulation constraints also influenced syntactic variation, preventing excessive repetition while maintaining balanced sentence length distributions. Comparative evaluations against baseline models reinforced the role of controlled latent state evolution in improving pronoun resolution, logical consistency, and contextual alignment across autoregressive text generation tasks.
In this report, we present Qwen2.5-Omni, an end-to-end multimodal model designed to perceive diverse modalities, including text, images, audio, and video, while simultaneously generating text and natural speech responses in a streaming manner. To enable the streaming of multimodal information inputs, both audio and visual encoders utilize a block-wise processing approach. To synchronize the timestamps of video inputs with audio, we organize the audio and video sequentially in an interleaved manner and propose a novel position embedding approach, named TMRoPE(Time-aligned Multimodal RoPE). To concurrently generate text and speech while avoiding interference between the two modalities, we propose \textbf{Thinker-Talker} architecture. In this framework, Thinker functions as a large language model tasked with text generation, while Talker is a dual-track autoregressive model that directly utilizes the hidden representations from the Thinker to produce audio tokens as output. Both the Thinker and Talker models are designed to be trained and inferred in an end-to-end manner. For decoding audio tokens in a streaming manner, we introduce a sliding-window DiT that restricts the receptive field, aiming to reduce the initial package delay. Qwen2.5-Omni is comparable with the similarly sized Qwen2.5-VL and outperforms Qwen2-Audio. Furthermore, Qwen2.5-Omni achieves state-of-the-art performance on multimodal benchmarks like Omni-Bench. Notably, Qwen2.5-Omni's performance in end-to-end speech instruction following is comparable to its capabilities with text inputs, as evidenced by benchmarks such as MMLU and GSM8K. As for speech generation, Qwen2.5-Omni's streaming Talker outperforms most existing streaming and non-streaming alternatives in robustness and naturalness.
Vision-language models (VLMs) have achieved remarkable advancements, capitalizing on the impressive capabilities of large language models (LLMs) across diverse tasks. Despite this, a critical challenge known as hallucination occurs when models overconfidently describe objects or attributes absent from the image, a problem exacerbated by the tendency of VLMs to rely on linguistic priors. This limitation reduces model reliability in high-stakes applications. In this work, we have observed the characteristic of logits' continuity consistency enhancement and introduced a straightforward and efficient method, Cross-Temporal Prediction Connection (TPC), designed to enhance the semantic consistency of logits by connecting them temporally across timesteps. TPC amplifies information flow and improves coherence, effectively reducing hallucination. Extensive experiments show that TPC surpasses existing representatives, delivering superior performance in both accuracy and efficiency while maintaining robustness in open-ended text generation tasks.