An Empirical Study on Zero-Data Bootstrapping for Conversational Recommender Systems
Rohan Surana, Junda Wu, Zhouhang Xie, Yu Xia, Nathan Kallus, Julian McAuley
University of California, San Diego · Netflix Inc. · Cornell University
Published August 28, 2026 · on the feed September 3, 2026 · ▲ 1 · arXiv 2504.15476
The abstract
Conversational Recommender Systems (CRS) typically require domain-specific dialogue data, which is costly, scarce, and often unavailable in new domains. We conduct a systematic empirical study of zero-data CRS bootstrapping: generating synthetic conversational supervision from non-conversational signals---item reviews, metadata, and user-item interactions---without any in-domain dialogue corpus. We compare two information-theoretic selection strategies, Jensen-Shannon diversity and Fisher information, across domain signals, model architectures, datasets, and fine-tuning paradigms. Our results show that domain-grounded synthetic data consistently outperforms zero-shot prompting and naive synthetic baselines; active selection improves data efficiency over random sampling; metadata and collaborative filtering signals each improve selection quality; and, in low-resource settings, synthetic data can outperform scarce real dialogues while further complementing them. These findings establish non-conversational domain signals as a viable path toward building CRS without conversational training data. The code is available at https://anonymous.4open.science/r/zero_data_crs/ .
The Brief
2:08A short podcast summary of the paper. Written and voiced by AI.
Transcript
An Empirical Study on Zero-Data Bootstrapping for Conversational Recommender Systems. By Rohan Surana and five others, from the University of California, San Diego, Netflix, and one other institution. Published on the twenty-eighth of August. No code repository is linked to it.
Conversational recommender systems usually need dialogue data from the domain they serve. That data is costly and scarce, and in a new domain it often does not exist at all. The claim this study tests is that you can start without it. Zero-data bootstrapping means generating synthetic conversational supervision from signals that are not conversations: item reviews, metadata, and records of which users interacted with which items. No in-domain dialogue corpus is used.
The study is systematic. It compares two information-theoretic strategies for choosing which synthetic examples to keep, Jensen-Shannon diversity and Fisher information, and varies the domain signals, the model architectures, the datasets and the fine-tuning paradigms around them. Four results come back. Domain-grounded synthetic data consistently beats both zero-shot prompting and naive synthetic baselines. Active selection is more data efficient than random sampling. Metadata and collaborative filtering signals each improve the quality of selection. And in low-resource settings, synthetic data can outperform scarce real dialogues, while also complementing them when the two are used together.
The boundaries matter here. The comparison against real dialogue is made in low-resource settings, where real conversations are few, and the claim is that synthetic data can win there, not that it always does. The selection results are relative to random sampling, and the baselines are zero-shot prompting and naive synthetic data. This is an empirical study rather than a new architecture: its contribution is a map of which signals and which selection strategies help.
The authors read the results as establishing non-conversational domain signals as a viable route to building these systems with no conversational training data. Reviews, metadata and interaction logs are things a new domain tends to have already. Dialogue is the thing it lacks, and the study's point is that the first can stand in for the second.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
A new domain, a chat recommender to build, and not a single conversation to learn from.
You want to launch a recommender that chats with users in a new domain, such as movies. Fine-tuning a small in-house model needs lots of domain dialogues, and you have none: they are costly to annotate and raise privacy worries.
The road ahead
- 1 · The problem ~7 min You must build a chatting recommender for a new domain, but the dialogue data it needs to learn from does not exist.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Turn reviews, metadata and interaction records into synthetic chats, choose carefully which samples to convert, and fine-tune a small model on the result.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Gather the domain signals, have a teacher LLM turn selected samples into synthetic chats, then fine-tune the recommender on them.
- 6 · The experiment ~6 min You’ll guess: In low-resource settings, how did synthetic training data compare with a small set of real dialogues for the recommender?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.