SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2606.04591 · 2026

Fine-grained Fragment Retrieval in Multi-modal Long-form Dialogues

Jie Zhou, Jinchao Zhang, Xiaoyue Duan, Hanbo Bi, Yingchao Feng, Zexi Jia, Zhiqiang Yuan, Jiapei Zhang, Chongyang Li, Qiwei Yan

arXiv · PDF · Open in the Atlas

Code that ran

We have not lifted any functions out of this paper's repositories yet, so there is nothing we have run. We have no repository linked to it either. That is a statement about our graph, not about the paper.

Abstract

With the widespread adoption of multi-modal communication platforms, long-form dialogues that interleave text and images have become increasingly common. In realworld applications such as customer support, knowledge management, and collaborative tools, users often seek to revisit coherent dialogue fragments related to specific topics or events, rather than isolated utterances or images. To address this need, we propose the task of Fine-grained Fragment Retrieval (FFR), which requires a model to locate semantically relevant fragments composed of multiple utterances and images within multi-modal long-form dialogues. We explore two core FFR settings: (1) FFR within Single-Dialogue, where the model retrieves fragments from a given long-form dialogue; and (2) FFR within Dialogue Corpus, which involves retrieving relevant fragments from a large-scale dialogue corpus, reflecting practical open-domain retrieval scenarios. To support the former, we introduce F 2 RVLM, a generation-based retrieval model trained with reinforcement learning, which incorporates multi-objective rewards and difficulty-aware curriculum sampling to enhance the consistency and coherence of retrieved fragments. For the latter, we develop FFRS, a two-stage retrieval system that combines offline fragment-level indexing with online retrieval, enabling efficient and accurate retrieval at scale. Specifically, each long-form dialogue is decomposed into minimal semantic units (i.e., fragments), which are encoded into a vector database by a Fragment Embedding Model (FEM). During inference, FEM rapidly recalls Top-K candidate fragments

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2606.04591")
get_code_for_paper("2606.04591")
have("2606.04591")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2606.04591.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2606.04591.svg)](https://syntology.ai/paper/2606.04591)

Connect an agent — have() is free.