attribute engineering

In multi-attribute controllable text generation, prompt design is challenging, and attribute overlap/conflict renders linear combination ineffective. Method: This paper proposes a novel language model composition paradigm grounded in the law of total probability and conditional mutual information minimization. We introduce the first “palette”-inspired compositional theory, analogizing attribute intensities to color channels and formalizing their mapping to generative style; we rigorously prove two theoretical properties—positive correlation and attribute enhancement—to enable interpretable, principled composition. Our approach integrates probabilistic graphical modeling, conditional mutual information optimization, single/multi-attribute fine-tuning, and collaborative prompt-space modeling. Results: Experiments demonstrate significant improvements over linear baselines across both single- and multi-attribute control tasks, achieving consistent gains in attribute consistency, text quality, and control precision.

This work proposes a controllable generation framework based on a structured semantic state space to address the challenges of flexibility, stability, and controllability in complex interactive scenarios, where existing generative models often suffer from semantic drift, inconsistent tone, or structural fragmentation. The approach encodes user input, environmental conditions, and historical context into manipulable latent representations and employs directional control vectors to guide content generation. A multi-level control mechanism is introduced, integrating constraints for semantic consistency and structural stability alongside a penalty for semantic drift. Experimental results on public dialogue datasets demonstrate that the proposed method significantly enhances semantic clarity, contextual coherence, and controllable expressiveness of generated responses.

This work proposes an unsupervised synthetic dialogue generation framework tailored for industrial settings where human-annotated data are scarce, relying solely on intent definitions. To enhance diversity, the approach explicitly incorporates topic and stylistic attributes and introduces two novel post-processing stylization models—Univ and Exam—combined with a large language model–based discriminative filtering mechanism to improve data quality. The study reveals that stylistic diversity has a significantly greater impact on the utility of synthetic data than topic diversity, and that integrating stylistic attributes during generation outperforms post-hoc style transfer. Experimental results demonstrate that the proposed method achieves 93.3% of the performance of models trained on human-annotated data across both industrial and public benchmarks, substantially enhancing the practicality of unlabeled synthetic dialogues.

Existing diffusion models struggle to perform fine-grained, continuous, and intensity-controllable editing of image aesthetic attributes (e.g., “brightness”, “refinement”) due to reliance on ambiguous text prompts or costly human preference annotations, limiting scalability. Method: We propose a plug-and-play aesthetic control framework that leverages a pretrained vision-language model to quantify semantic similarity of abstract aesthetics, and introduces a lightweight value encoder that maps [0,1] intensity scalars into differentiable embeddings—seamlessly integrated into text-conditioned diffusion sampling. Contribution/Results: Our method requires no human preference labels, enables independent or joint control over multiple attributes, supports continuous cross-intensity editing, and is compatible with mainstream open-source generators (e.g., Stable Diffusion). Experiments demonstrate significant improvements over baselines in both single-attribute fidelity and multi-attribute coordination, achieving high practicality, flexibility, and scalability.

To address weak generalization in skill discovery caused by insufficient semantic diversity, this paper proposes an LLM-guided prompt-constrained skill learning framework. Methodologically, it is the first to explicitly maximize semantic dissimilarity among skills by directly leveraging large language models’ semantic priors; skills are learned within user-defined natural-language prompts’ semantic subspaces via joint optimization of reward-free reinforcement learning and a skill discriminator, enabling semantic-aware state-space exploration. Contributions include: (1) establishing an interpretable mapping from natural language to the skill semantic space; and (2) supporting plug-and-play, prompt-driven downstream task adaptation. Evaluated on legged robot navigation and robotic arm manipulation tasks, the method significantly outperforms five baselines. Crucially, altering only the input prompt activates distinct semantic behavior modes—demonstrating strong generalization and precise controllability.

This work addresses the lack of systematic research and unified frameworks in artificial intelligence for high-level semantic tasks such as humor, irony, metaphor, and empathy. It introduces, for the first time, the concept of High-Level Semantic Intelligence (HLSI), formally delineating its scope and proposing a comprehensive framework encompassing task taxonomies, data construction methodologies, modeling strategies, and evaluation protocols. By integrating advances in natural language processing, multimodal learning, and cognitive modeling, the study synthesizes existing datasets and algorithms to provide a thorough review of current research on high-level semantic understanding and generation across textual, spoken, visual, and multimodal contexts. This synthesis establishes a foundational theoretical basis and outlines key directions toward achieving human-like semantic intelligence.

Current text-to-image models often sacrifice diversity to strictly adhere to input prompts, yielding outputs confined to a single visual interpretation. This work proposes a novel paradigm termed “semantic browsing,” which decouples semantic decision-making from pixel generation to introduce controlled diversity while preserving prompt fidelity. By leveraging rich semantic text representations, vision-language models (VLMs), and a customized agent-based workflow, the method explicitly guides structured and interpretable semantic variations. The resulting navigable design space ensures that each generated image variant corresponds to a distinct, meaningful semantic choice, substantially enhancing both output diversity and user controllability.

This work addresses the challenges of unpredictability and lack of personalization in large language models for dialogue generation, which hinder controllable output. The authors propose a lightweight, model-agnostic, and reusable ontology-driven control framework that defines dialogue attributes through an ontology to establish modular and interpretable constraints. By integrating a hybrid fine-tuning strategy, the framework effectively steers models to generate responses aligned with specified dimensions. Evaluated across seven open-source dialogue large language models, the approach consistently outperforms baseline methods on tasks involving English proficiency and content polarity, demonstrating strong performance even with smaller-scale models. Moreover, it significantly enhances controllability, cross-domain adaptability, and alignment with desired conversational strategies.

Existing multimodal dialogue generation approaches struggle to achieve fine-grained alignment and controllable expression across speech, vision, and text. This work proposes a conditionally controllable multimodal dialogue generation framework grounded in the natural interaction patterns of human communication. By constructing a high-quality annotation pipeline leveraging cinematic data, we introduce MM-Dia—the first multimodal dialogue dataset supporting style-controllable spoken dialogue synthesis—and establish MM-Dia-Bench, a benchmark for evaluating cross-modal style consistency. Experimental results demonstrate that the proposed method significantly enhances fine-grained controllability in generated dialogues. Furthermore, evaluations using MM-Dia-Bench reveal substantial gaps in current models’ ability to replicate the expressive richness characteristic of human multimodal interaction.

Traditional meaning representations (MRs) exhibit limited expressiveness in conversational natural language generation, often failing to produce semantically accurate and fluent responses consistently across diverse tasks and data conditions. This work proposes enhancing both training and inference by incorporating semantically rich MR–utterance pairs as input augmentation. For the first time, the approach is systematically evaluated within a unified framework spanning multiple domains, corpora, and evaluation metrics. Experimental results demonstrate substantial improvements in generation quality, particularly in complex tasks and challenging scenarios such as low-resource, high-variation, and zero-shot settings. Human-evaluated semantic metrics prove more sensitive to subtle semantic errors, confirming the model’s robustness and rapid adaptability at both semantic and communicative-intent levels.