Every morning we screen several hundred new arXiv, conference and industry papers across four tracks —
generative recommendation, personalization, AI Search and ML infrastructure — score each
on business relevance, novelty, practicality, evidence and timeliness, and write long-form deep dives with
figures for the best ones.
English issues start from 2026-10-17. Earlier issues are Chinese-only and are not listed
here — browse them in the Chinese archive.
4.65
The Shopee authors measured three-level semantic IDs (SIDs) built by standard RQ-KMeans: the top 10% most-exposed full codes took 89.62% of exposure, so a generative recommender's probability mass concentrates on a few popular codes. UniRec has the model first generate the item's level-2 and level-3 category tokens and only then the three SID levels, and when building SIDs it gives every code an exposure-weighted capacity limit. Once the category is generated, each later SID level only has to choose within that category. Offline HR@50 is 22.6% higher (relative) than OneRec-V2; online, against the old multi-stage pipeline, orders rose 4.76% and GMV 5.60%.
4.60
KuaishouWhen an LLM serves as a recommendation foundation model, generating a natural-language reasoning trace (CoT) before outputting the item's semantic ID (SID) makes the output long and cuts throughput. Kuaishou's OneLatent inserts 3 learnable special tokens between the user context and the target SID: training first makes the model reproduce a teacher LLM's reasoning text after these 3 tokens, then switches to predicting the SID directly after them; at serving time only these 3 tokens are fed and no reasoning text is generated. The teacher writes reasoning from three angles: inferring intent, scoring the clicked item, and explaining why it was clicked. On the industrial dataset, SID@64 is 17.44% higher than OneReason-CoT's reason-then-answer mode; but against the strongest baseline trained on the same teacher reasoning and answering directly, the gain is only 1.58%, so most of the improvement comes from the teacher reasoning data. The reasoning slots are fixed at 3, so decoding does not grow with reasoning length, and throughput is 17.49x that of explicit reasoning. A local-services advertising A/B test estimates a 9.6% revenue lift.
4.60
Google's Discover Ads team found that the offline Poisson Loss of its conversion model (pConvs) correlated only 0.1 with online advertiser value (Delivered Value: conversions per click times the advertiser's target cost per acquisition, tCPA, summed over clicks), so models that improved offline often did not gain once launched. The authors weight each click in the loss by its tCPA, and pair every low-traffic experiment with a retrained production model serving at the same traffic as its control. With tCPA weighting, the offline loss measures exactly the conversions advertisers are willing to pay for. Offline-to-online correlation rose to 0.65, and low-to-high-traffic experiment correlation rose from 0.53 to 0.93.
4.15
Recommendation teams often use a different library for ranking, contextual bandits, uplift and sequential recommendation, each with its own offline evaluation interface. Intuit's open-source scikit-rec splits business logic, scoring strategy and the underlying model into three layers (Recommender, Scorer, Estimator), and five recommendation paradigms share one evaluate() interface, which includes six off-policy evaluators such as IPS, SNIPS and DR whose results match Open Bandit Pipeline within floating-point error. On the same logged data, a bandit policy and a ranking model can be compared directly with the same evaluators. An internal version has powered 15 use cases in TurboTax and QuickBooks for more than three years under a sub-100 ms latency requirement.
4.85
Semantic Retrieval
Hybrid retrieval fuses BM25 and dense retrieval into one ranking, and most search engine and RAG framework docs default to reciprocal rank fusion (RRF). The authors keep a per-query ledger on 7 BEIR datasets: the best fusion method, score interpolation with weight 0.7 on the dense score, beats dense-only retrieval on mean nDCG@10 on every dataset, yet still ranks 8–21% of queries worse than dense-only. These losses concentrate on queries where only one retriever finds a relevant document; the other retriever then contributes only noise, and fusion pushes relevant documents that were already ranked high back down. Scaling the dense encoder from 22M to 596M parameters does not remove this loss. RRF beats dense-only only on datasets where the two retrievers are roughly equal, so the fusion method should be chosen from the measured gap between the two retrievers on the target corpus.
4.55
RAG Pipeline
RAG chunking schemes are usually compared on one static corpus, but knowledge bases keep changing. Each chunk is embedded on its own, so when an edit moves a chunk boundary, unchanged text must be re-embedded too. The authors measure this on 120 pairs of adjacent Chinese and English Wikipedia revisions, reusing an old vector only when a chunk's embedding input is byte-identical: non-overlapping fixed-size chunking recomputes 57–68% of the new version's embedding input, paragraph-first recursive chunking 38–56%, and content-defined chunking (CDC, which picks cut points from local text fingerprints) 43–54%, while evidence recall under a 2048-token retrieval budget differs by less than 2 percentage points among the three. When 1% of text is inserted at the start of an article, fixed-size chunking recomputes everything, because every window after the insertion point shifts, so no chunk matches the old version. For knowledge bases that change often, a chunker's recompute fraction should be reported alongside its retrieval quality.
4.25
LEAD STORY ·
GoogleGoogle Agent traffic is multi-step, shares long prefixes and arrives in bursts. Behind an ordinary round-robin load balancer, replicas recompute the same prompts and accelerator memory fills up with duplicate KV cache. llm-d 1.0, released at the PyTorch Conference by Google together with Red Hat, IBM and others, is an open-source distributed inference stack that runs on Kubernetes and organizes a set of vLLM and SGLang model servers into one serving pool: the router tracks which replicas hold each cached prefix and picks a replica using load and predicted latency; when accelerator memory fills up, KV cache is moved first to host memory and then to SSD; prefill and decode are split into two pools. Version 1.0 adds flow control that queues requests by priority and fairness, and rolling upgrades that do not interrupt sessions. On an agentic coding workload on Ironwood TPU, prefix-aware routing plus host-memory offload cut median time-to-first-token by more than 85% and delivered 2.4x the total throughput of a plain Kubernetes Service. The gain comes mainly from sending a session's requests back to the replica that already holds its context, which removes repeated prefill.
3.80
DrugCLIP embeds protein pockets and small molecules into one vector space, which turns drug screening into vector retrieval: one campaign queries 500M molecules with over 20K pockets, keeps 500K candidates per pocket, and needs 99% mean recall relative to exact search. With answers this large, an IVF index must probe almost every list and a graph index must track at least K candidates, so little of the library can be skipped, and the bottleneck becomes maintaining each query's K candidates. SweepANN scans instead: INT8-compressed molecule vectors stay resident in the memory of 4 GPUs, and each block is read only once for a whole batch of pockets; it first takes each pocket's top K over a sample of blocks and uses the lowest of those scores as a cutoff, after which only rows above the cutoff enter the buffer. Four RTX PRO 6000 GPUs finish 26,562 pockets in about 61 seconds at 99.6% mean recall; GPU IVF configurations that reach the recall target take 8.2–24x as long.
RecSys 2026
Intuit