GENERATIVE · 2026-11-10
Generative — Paper Daily, 2026-11-10
Today's generative track keeps the already published conversational memory architecture and adds newly surfaced papers. One is Shopify's report on using Semantic IDs for both search and recommendation. The others cover latent reasoning without teacher-written chains of thought, a cascade that spends expensive language-model reranking only where it helps, and sparse-autoencoder steering of generative recommendations by play concepts.
4.90
LEAD STORY ·
Shopify
Search and recommendation at Shopify have to retrieve relevant products from a changing catalog of billions of items under tight latency limits. This report casts both tasks as generating semantic IDs: short sequences of content-derived tokens that represent products. The model is jointly pretrained on catalog and behavioral data so search and recommendation share a representation and a training objective.
4.25
Latent reasoning for generative recommendation usually needs a teacher model to write chain-of-thought traces, which is costly and hard to stabilize. This paper skips teacher generation. It builds semantic targets from the text meaning of items already in a user's history, and uses those targets to supervise the model's intermediate tokens that are reserved for reasoning steps, without looking at the next-item label.
4.10
Large language models are attractive rerankers, but running an expensive model on every request is costly and often hurts users with short histories. This production system, deployed for content recommendation at a large financial-planning software provider, reranks every request with a lightweight first tier.
3.85
Generative recommenders can predict item identifiers from a user's history, but they rarely offer readable, concept-level control over what gets recommended, especially if the model cannot be retrained. This game-discovery work defines play concepts with a language model, finds sparse-autoencoder features that are both associated with a concept and effective when edited, and then amplifies or suppresses those features at inference time. On two industrial generative recommenders, amplification raises the share of the target concept by 113.6% on average for an Atomic ID model and by 24.6% for a Semantic ID model. Selecting features by their steering effect works better than picking the features most strongly associated with the concept.
3.75
In conversational recommendation, retrieving the item a user wants does not mean the user sees it. The authors test this with Amazon review data and simulated users. Each simulated user has a target product in mind but may not name it, so it reveals requirements as the conversation goes. Each turn the system retrieves a batch of candidates and shows only the top 3. By the end of the dialogue the target has been retrieved in 74.75% of sessions, yet only 62.63% of sessions actually put it in front of the user, so the gap sits in choosing which candidates to show. The system itself splits work across several LLM agents: a supervisor asks clarifying questions when the request is vague and brings preferences from the user's earlier sessions into retrieval. Compared with one-shot recommendation, clarification plus memory raised the simulated users' acceptance rate by 28.9 percentage points. When evaluating conversational recommenders, the display step deserves its own measurement.