GENERATIVE · 2026-11-18
Generative — Paper Daily, 2026-11-18
Today's generative track turns on two questions: whether generated identifiers and reasoning actually improve ranking, and whether a large model acting as a shopping guide can be pulled off course by surface cues. One paper adds only a multiple-choice task to the learning objective so SID generation also learns relative comparison; another aligns the tokenizer to retrieval first, then deploys it in industrial video search. Reasoning optimization starts rewarding each sentence for how it changes the likelihood of the target identifier, while a reinforcement-learning study shows that ratio-based objectives can collapse training and that the update itself needs to be bounded. An audit finds that shopping recommenders still favor longer product descriptions after relevance is controlled, and that product advice cannot be judged only by whether a reply names the community-listed item, because naming it is not the same as selecting it.
4.78
Generative recommenders often predict the target item's SID directly and never explicitly learn negatives or relative ranking, which sits poorly with the idea behind BPR, where the rank of a positive against negatives matters. Adding a simple multiple-choice auxiliary task to the learning objective, and synthesizing its training data with task identifiers and candidate construction, supplies that comparison signal with only a small change.
4.48
Generative retrieval produces a document's SID, so where the tokenizer puts quantization error directly limits what can still be recalled. Common SID methods quantize residual codes on document embeddings alone and do not check whether a query can still find its relevant documents after quantization.
4.28
When a large model picks a product from retrieved e-commerce candidates, the choice should depend on relevance to the user query, but the length of a description can also quietly add weight. Relevance labels are rarely available, so this kind of structural bias is hard to measure directly. This work rewrites each product description with randomly assigned paraphrasers and presents the candidates in random order.
4.20
When generative recommenders add reasoning, existing methods usually reward a whole trajectory from the final recommendation, so they never measure how much each sentence helps the target SID, and different trajectories in a group often get nearly the same reward. This work measures predictive gain: how much each reasoning sentence changes the log-likelihood of the target SID.
4.09
Applying reinforcement learning directly to generative recommendation often makes ratio-based objectives unstable: clipping can zero the gradient on some samples, and negative-advantage updates can grow with the importance ratio, so a few negatives dominate training. GBPO uses stop-gradient, confidence-aware normalization so that both positive and negative updates stay bounded, and the update shrinks when the policy is already confident. On three Amazon datasets, HR@10 rises 0.71 to 2.21 percentage points over supervised TIGER. On Steam, it was the only method that avoided both performance collapse and a non-finite loss in every repeated run. With a continuous reward or a different teacher model, all fifteen competing runs tripped the collapse rule. For reinforcement post-training of an SID generator, stabilizing the update geometry matters more than clipping the ratio more aggressively.
3.91
Matching a product name in a reply to a community-listed item cannot tell whether the advice is good: several products may fit, and the recorded answer may itself be poor advice. When a community-listed alternative is shown beside a model's own sampled car recommendation, with sources and order hidden, six models prefer that alternative in 1.9% to 6.5% of eligible disagreements, consistently in both presentation orders. This work splits advice into the user's stated needs, assumptions the model added, candidate generation, filtering, and final choice, and lists 19 failure modes with the evidence needed to assess them. It looks only at the request and the written reply, without access to the model's internals. Naming a reference product is not the same as selecting it, and how the answer is extracted or the candidate list is built also changes the score. To improve product advice, first separate a missing candidate from a candidate that was generated and then chosen wrongly.