Researchers are exploring new approaches to generative retrieval (GR) and dense retrieval (DR) for tasks like recommendation and question answering. One study found that while GR excels at identifying the best relevant item, DR performs better on exact-target recall. The research suggests that incorporating static item value into DR can reverse GR's lead, and even a smaller dense encoder can outperform a large SID-LLM in e-commerce relevance judgments. Another paper proposes a bit-factorized identifier design for GR that improves adaptation to unseen documents without retraining, significantly boosting retrieval performance and reducing inference latency. Further work disentangles the effects of paradigm, identifier, and decoding in GR, showing that decoding strategies can dramatically impact diffusion models and that random identifiers can retain a high percentage of performance, indicating a tendency for memorization in current GR models. AI
IMPACT These papers explore methods to improve retrieval efficiency and accuracy, potentially impacting recommendation systems and search engines by enabling better handling of unseen data and optimizing model decoding.
RANK_REASON The cluster contains multiple academic papers published on arXiv discussing novel techniques and evaluations in generative retrieval.
Read on arXiv cs.IR (Information Retrieval) →
- Hugging Face
- MS300K
- NQ320K
- arXiv
- MS MARCO
- TriviaQA
- alphaXiv
- Amazon Toys
- CatalyzeX Code Finder for Papers
- CORE Recommender
- DagsHub
- Dense Passage Retrieval for Open-Domain Question Answering
- e-commerce
- Generative Retrieval
- Language Models
- SID-LLM
AI-generated summary · Google Gemini · from 6 sources. How we write summaries →
