Retrieval-Augmented Generation (RAG) enhances large language models with external knowledge, but long retrieved contexts incur high computational cost and often degrade performance due to context–knowledge mismatch. Parametric RAG (PRAG) addresses this by integrating external knowledge into model parameters via lightweight adapters. However, existing PRAG methods are fundamentally document-centric, treating each document as an independent unit of knowledge, which leads to poor scalability and weak supervision. We argue that knowledge is inherently distributed across semantically related documents, and document-level parameterization is a suboptimal granularity. To this end, we propose ParaShard, a knowledge-centric framework that organizes knowledge into semantically coherent knowledge shards and parameterizes each shard with a shared adapter. This design enables richer supervision across related documents while significantly reducing the number of required adapters. At inference time, ParaShard dynamically selects and composes multiple shards using a diversity-aware strategy to ensure comprehensive knowledge coverage. Experiments on four knowledge-intensive QA benchmarks show that ParaShard achieves superior average performance across diverse knowledge-intensive tasks compared with strong RAG and PRAG baselines, with particularly strong gains on multi-hop reasoning, and substantially reduces storage cost compared to PRAG, demonstrating the knowledge injection effectiveness and storage efficiency of knowledge-centric parameterization.

Generative retrieval (GR) retrieves documents by directly generating document identifiers (DocIDs) from queries using a sequence-to-sequence model. Recent reasoning-enhanced GR methods improve retrieval effectiveness by generating intermediate reasoning before DocID decoding. However, they typically apply reasoning uniformly to all queries, even though many queries can already be handled effectively through direct DocID decoding. This introduces unnecessary token generation and retrieval latency when reasoning provides little additional benefit. In this work, we propose Generative Retrieval with Adaptive Reasoning (GEAR), a retrieval framework that dynamically allocates reasoning computation based on query-specific needs. Instead of treating reasoning as a fixed pre-retrieval procedure, GEAR learns when to reason, what retrieval-oriented context to generate, and when to transition from free-form generation to constrained DocID decoding. This adaptive behavior is implemented through a learnable retrieval transition policy optimized jointly for retrieval quality and reasoning cost. Experiments on NQ, MS MARCO, and TREC DL 2022 show that GEAR maintains retrieval effectiveness comparable to strong reasoning-enhanced GR methods while substantially reducing reasoning overhead, triggering non-trivial reasoning for only 37.1% of queries and reducing both generated tokens and retrieval latency.

Large language models (LLMs) are emerging as versatile foundations for generative retrieval (GR), a paradigm that directly generates relevant document identifiers (docids) given a query. Existing ways of using LLMs for GR typically rely on heavy post-hoc training to fit a relevance data prior, often leading to overfitting and poor cross-domain generalization. Can retrieval, like other language tasks, be performed without retriever-specific training using LLMs? We conduct a preliminary study and find that constrained decoding, required to ensure valid docids, often misaligns with the model’s natural free-generation patterns. We propose FreeGR, a plug-and-play framework that adapts retrieval data to a LLM’s habitual generation behavior. FreeGR consists of three components: an LLM self-annotated docid generator, an LLM-preferred input context designer, and two decoding strategies that trade off retrieval effectiveness and efficiency. It requires no retriever-specific parameter optimization, is model-agnostic, and scales across LLM families and sizes while preserving their generative capabilities. Experiments on the NQ, MS MARCO, and BEIR datasets show that FreeGR achieves competitive retrieval effectiveness compared with several trained generative and dense retrievers, without retriever-specific parameter optimization.

Dense retrieval is a fundamental component of e-commerce search, projecting user queries and items into a shared semantic representation space for large-scale, real-time retrieval. While LLM-based embedding models are increasingly replacing BERT-style encoders, prevalent training pipelines still largely rely on BERT-like supervised fine-tuning and hard negative mining. Such paradigms require intricate offline negative construction, thereby constraining model iteration efficiency and limiting the evolution of learned semantic representations. We propose Retrieval-GRPO, a dense retrieval framework based on reinforcement learning. It dispenses with offline hard negative mining by performing dynamic Top-K candidates retrieval during training. Moreover, we introduce a relevance-oriented LLM as a reward model to provide real-time supervisory signals. Guided by the resulting reinforcement learning objective, the retriever updates embeddings to better capture semantic correspondence and achieve improved generalization on complex long-tail queries. Extensive offline and online experiments substantiate the effectiveness of Retrieval-GRPO in Taobao Search, which serves China’s largest e-commerce platform.