New framework optimizes token-level rewards for generative document retrieval

Researchers have developed a new reinforcement learning framework to improve generative document retrieval. This method addresses the issue of coarse-grained feedback in existing models by assigning token-level relevance rewards. By measuring how each token decision impacts retrieval quality, the framework enables more precise credit assignment, guiding the model to favor decisions that directly contribute to document relevance. Experiments show this fine-grained supervision consistently outperforms sequence-level reward baselines. AI

IMPACT This research could lead to more accurate and efficient document retrieval systems by improving how generative models learn from feedback.

RANK_REASON The item is an academic paper detailing a new framework for generative document retrieval. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.IR (Information Retrieval) →

• alphaXiv
• arXiv
• arXivLabs
• Bibliographic Explorer
• CatalyzeX Code Finder for Papers
• Connected Papers
• CORE Recommender
• cs.IR
• DagsHub
• DocIDs
• Gotit.pub
• Hugging Face
• Influence Flower
• Litmaps
• ScienceCast
• scite Smart Citations
• Token-Level Credit Assignment Optimization for Generative Document Retrieval

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →