Multimodal Retrieval-Augmented Generation (MM-RAG) has emerged as a promising approach for enhancing the reliability and factuality of large vision-language models (LVLMs). While end-to-end loss backpropagation is infeasible due to non-differentiable operations during the forward process, current methods primarily focus on component-level optimizations, necessitate extensive component-specific training datasets and suffer from a gap between local and global optimization objectives. In this paper, we propose a new paradigm that backpropagates global rewards from the system output to each component and then transforms these rewards into specific local losses, enabling each component to perform gradient descent and thus ensuring end-to-end optimization. Specifically, we first insert two lightweight multimodal components, a query translator and an adaptive reranker, to address the heterogeneity of multimodal knowledge and the varying knowledge demands for different questions, and then tune only these inserted components using our proposed paradigm to integrate the entire system. Our method achieves SOTA performance on multiple knowledge-intensive multimodal benchmarks with high training efficiency, relying exclusively on supervised signals from an external reward model. Experimental results and our detailed analysis of the evolution of components during training collectively reveal the advantages and considerable potential of this paradigm as a promising direction for MM-RAG research.

Fine-grained control is essential for precise and customizable text generation, yet existing benchmarks evaluate models on only a few attributes, typically fewer than five. We introduce UltraBench, a new benchmark for extremely fine-grained controllable generation (EFCG), which evaluates large language models (LLMs) under dense, multi-attribute constraints. Each sample includes approximately 70 attributes, combining LLM-extracted soft constraints (e.g., style and tone) with programmatically enforced hard constraints (e.g., word count). Using UltraBench, we conduct a comprehensive evaluation of state-of-the-art LLMs and prompting strategies. Models such as GPT-4.1 and Qwen3-8B perform well on individual constraints, achieving instruction-level accuracy above 70%, but consistently fail to satisfy all constraints simultaneously. To understand this limitation, we analyze model behavior across different dimensions. First, we observe a clear position bias: models tend to prioritize constraints presented later in the prompt while neglecting those that appear earlier. Second, we find that structural and formatting-related constraints are significantly more difficult to satisfy than content- or style-based ones, suggesting that current models struggle to coordinate global structure with token-level control. Finally, our error analysis reveals distinct failure modes: GPT-4.1 often attempts to follow constraints but falls short in precision, whereas LLaMA frequently omits constraints, particularly in multi-turn settings. These findings highlight fundamental limitations in EFCG and underscore the need for new methods that support dense, instruction-sensitive generation.

Instruction-tuned large language models (LLMs) employ structured templates, such as role markers and special tokens, to enforce format consistency during inference. However, we identify a critical limitation of such formatting: it induces a phenomenon we term diversity collapse, where the model generates semantically similar outputs for open-ended inputs, undermining creativity and variability. We systematically evaluate this effect across tasks like story completion and free-form generation, finding that (1) diversity collapse persists even under high-temperature sampling, and (2) structural tokens in templates significantly constrain the model’s output space. To contextualize these findings, we fine-tune using a range of structured prompts and then evaluate them across three axes: downstream task performance, alignment behavior, and output diversity. Our analysis shows that format consistency between fine-tuning and inference is crucial for structure-sensitive tasks (e.g., GSM8K, IFEval), but has marginal influence on knowledge-heavy tasks (e.g., MMLU, WebQuestions). In contrast, output diversity is primarily governed by the presence or absence of structural tokens, with minimal formatting yielding the most diverse outputs. These findings reveal that current prompting conventions, while beneficial for alignment, may inadvertently suppress output diversity, underscoring the need for diversity-aware prompt design and instruction tuning.