A single LLM jointly fine-tuned for conversational dense retrieval and retrieval-augmented generation beats separate retriever-plus-generator pipelines on most test collections, though its headline benchmark was contaminated by inclusion in the training set.
Ask Optimal Questions: Aligning Large Language Models with Retriever's Preference in Conversation
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
abstract
Conversational search, unlike single-turn retrieval tasks, requires understanding the current question within a dialogue context. The common approach of rewrite-then-retrieve aims to decontextualize questions to be self-sufficient for off-the-shelf retrievers, but most existing methods produce sub-optimal query rewrites due to the limited ability to incorporate signals from the retrieval results. To overcome this limitation, we present a novel framework RetPO (Retriever's Preference Optimization), which is designed to optimize a language model (LM) for reformulating search queries in line with the preferences of the target retrieval systems. The process begins by prompting a large LM to produce various potential rewrites and then collects retrieval performance for these rewrites as the retrievers' preferences. Through the process, we construct a large-scale dataset called RF collection, containing Retrievers' Feedback on over 410K query rewrites across 12K conversations. Furthermore, we fine-tune a smaller LM on this dataset to align it with the retrievers' feedback. Our resulting model demonstrates superiority on two benchmarks, surpassing the previous state-of-the-art performance of rewrite-then-retrieve approaches.
fields
cs.CL 1years
2025 1verdicts
REJECT 1representative citing papers
citing papers explorer
-
UniConv: Unifying Retrieval and Response Generation for Large Language Models in Conversations
A single LLM jointly fine-tuned for conversational dense retrieval and retrieval-augmented generation beats separate retriever-plus-generator pipelines on most test collections, though its headline benchmark was contaminated by inclusion in the training set.