GENERATIVE · 2026-11-11
Generative — Paper Daily, 2026-11-11
Today's main thread is adding business constraints to generative retrieval: a generated ID must exist and must also be deliverable and match the request. TS-GEAR encodes the delivery range in the SID prefix, RoutingTrie blocks dead-end branches before each beam selection, and LEGIT checks the rules at every token while writing a whole page and finds that the main cost is waiting between CPU and GPU. AST rebuilds the IDs themselves as paths in a tree. A second thread asks how assistants understand users: IntentLens calls tools to fill in needs the user left unclear, and SituRecBench tests models that ask questions and give directions in 3D stores. CloseSort gives an exact way to merge pairwise preferences, which may contain cycles, into one list.
4.60
LEAD STORY ·
Alibaba
In food delivery search, a merchant that cannot reach the user's current location is not merely ranked low, it must not appear at all. Common SIDs are built only from content and collaborative signals and at best treat location as a soft signal, so unreachable merchants are filtered after generation, waste beam slots, and stay outside end-to-end training.
4.40
Generative recommendation numbers products with SIDs, and the dominant approach is residual quantization: each level has one global code table and looks up whatever the previous level left unexplained, which amounts to assuming that each token carries its own independent meaning. The authors checked RQ-VAE IDs on Amazon Beauty.
4.20
On devices such as smart glasses, users often give only a picture and one vague sentence, leaving style, budget and venue requirements unsaid. Existing recommenders either rely on past behavior or assume the request is already explicit, so direct similarity between query and candidates easily ranks the wrong places first. For example, a user points at a cafe dish and wants a similar vibe but a more affordable lunch.
4.15
Generative retrieval often has to honor a condition, such as returning only products from one shop, while a trie only guarantees that a generated SID exists and says nothing about shops or countries. The usual approach filters after generation, but by then the beam is full of other shops' candidates and few results may survive.
4.15
UberUber Eats generative search writes a whole page of stores and menu items in one decode, and every token must keep each store real, deliverable and unrepeated, with each item belonging to the store above it. These rules keep changing with the request and with what is already on the page, a trie over the catalog only keeps IDs real, and the grammar-constrained decoder the authors tried could not be both exact and fast. The authors compared dozens of implementations: the direct GPU version makes each token about 5 times as slow as unconstrained decoding, while the GPU compute units are active at most 5.42% of the time. Hardware counters and ablations both point to kernel launches and CPU waits on the GPU as the overhead. The arithmetic of the constraint is light, and the expense is the round trip to the GPU for each group of sequences. Walking the trie on the CPU while the model runs its forward pass brings the cost down to about 1.7 times, with no custom kernel. Teams generating whole pages can keep per-token constraints on the CPU, as the authors suggest.
3.40
Existing benchmarks look either at dialogue and preferences or at navigation and manipulation in 3D, so they cannot tell whether a shopping assistant can ask about needs, read the scene and give directions at once. This benchmark puts LLMs in 3D supermarkets and furniture stores, sharing the user's first-person view, and judges them on whether the user buys. Using simulated users, the authors tested 7 VLMs: the best, Gemini 3.1 Pro, reaches a purchase rate of 80% in the supermarket and 66% in the furniture store, while the two small models stay at about three in ten. The guidance of GPT-5.6 Sol is followed most often, yet its purchase rate trails Gemini. In a supermarket ablation with GPT-5.6 Sol, removing the first-person image barely changes the purchase rate and only makes the guidance followed less closely, and only removing the text annotations of the scene as well lowers purchases and dialogue ratings. First-person information works mainly through the text annotations. For teams building shopping agents in this kind of setting, a text description of the products and places currently in view is the input most worth keeping.
3.15
Merging pairwise win probabilities into one list is, at maximum likelihood, a weighted Kemeny problem that is NP-hard in general. Known exact rank-two algorithms use only who beat whom and ignore by how much, and LLM judges' preferences can also form cycles. This paper proves that when the matrix of win log-odds has rank two, the problem can be solved exactly even with cycles. Each entry gets a planar vector, one closing vector makes all vectors sum to zero, and the likelihood of a list then corresponds to the area of the polygon those vectors trace end to end. Sorting by angle maximizes the area, and cutting at the closing vector gives the best list. On LLM judge data, the rank-two component captures 89.2% of the energy, and on samples with cycles its gap to the optimum is 43.9% smaller than that of Borda, while on samples without cycles Copeland does better. For people working on rank fusion and model evaluation, its advantage is concentrated in data where pairwise preferences form cycles.