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…
When 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…
The Netflix homepage consists of rows, each with a theme and a set of titles, and it used to be assembled by several stages such as a row model and an entity model. GenPage encodes the user's viewing history, device, time of day and so on as tokens forming a…
Because attention cost grows with the square of sequence length, generative recommenders usually consume only short behavior sequences. GLASS, from Kuaishou and Tsinghua, uses the long-term history in two places: before generating the first SID level, it…
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…
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…
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…
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…
Mathlib is the math library of the Lean proof assistant. Queries from humans and coding agents searching it have median lengths of only 16 and 15 tokens, while the theorem paraphrases used to train earlier retrievers have a median of 82 tokens. Lean Finder V2…
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…
Generative recommenders emit item-ID tokens left to right, so temporary diversity at inference can only rewrite one position mid-decode and often produces IDs that do not exist in the catalog. ReGen uses a bidirectional Transformer to predict all 4 ID tokens…
Taobao search CTR models mainly remember historical clicks through item IDs, so the same item with a different main image is almost indistinguishable to the model. FineSID turns each item image into 30 discrete codes across four VAR tokenizer scales from 1×1…
Generative recommenders use beam search to emit item-ID tokens one by one, and tail items have low generation probability and are often pruned in the first one or two steps: on Amazon Industrial & Scientific, MiniOneRec's 50 candidates cover only 11.94% of…
Personalized reward models score replies per user, but each user labels only a few dozen preference pairs, so most unlabeled replies go unused. Pairwise loss only constrains the score gap under the same prompt, so scores from different prompts can shift as a…
When a sequential recommender encodes each historical behavior, some behaviors converge after one layer while others need several stacked layers to stabilize. In a three-layer SASRec, 76.1% of behaviors are stable after the first layer and 6.7% need the…
Perplexity releases 0.6B and 9B multi-vector retrievers that encode text, images, and document screenshots: each document token stores a 128-d vector, and scoring sums MaxSim of each query vector against document vectors. Both models distill token-by-token…
When rewriting a query vector with first-round retrieved documents or an LLM-written hypothetical answer (HyDE), the usual approach averages feedback vectors and interpolates with the original query. DIFF argues that each piece of feedback only supports…
Korea University and KAIST have Claude Code (Claude Opus 5.5) design retrieval tools for a search agent: it reads each question's execution log and score, jointly edits pre-generated index files, executable search scripts, and usage instructions, and after 10…
Northeastern University and Tsinghua find that in retrieve-while-reasoning RAG, later sub-queries keep drifting away from the original question along the direction earlier sub-queries already moved, and the effect grows with step count. MAPS therefore makes…
Parlant used to ask an LLM whether each customer-service rule is relevant every turn—about 69,000 tokens for 36 rules. S&Q instead has an LLM offline write 10 triggering user messages and 10 easily confused counterexamples per rule and trains a logistic…
IBM builds ClariEdge, a ~3,900-sample multi-turn RAG clarification set that specially includes hard negatives that "look ambiguous but the referent is already fixed by context," then LoRA-tunes small Granite models to decide whether to clarify and to list…
An independent researcher compares "add 2:4 sparsity" versus "drop one precision tier" on Llama-3.1-8B-Instruct and Qwen3-8B under matched storage bits (including sparse indices) and matched nominal compute tiers, with measurements on L40S, H100, and RTX PRO…
SGLang, NVIDIA, and AMD use a plugin framework to plug specialized fused kernels into SGLang, cutting inter-kernel waits and cross-GPU communication at low concurrency. On AgentX replays of DeepSeek-V4.1-Flash + DSpark, per-user P90 generation rises from…
Hisense's RCSD continues training a speculative-decoding draft model that was distilled with HASS, using DPO or GRPO: sequences from the target alone are "good," sequences alternating target and draft are "bad," and the GRPO variant scores by acceptance rate.…
Scitix.AI's UlaCache maps every GPU KV pool and CPU memory into one logical address space so GPU and CPU each store different history, and removes MLA's cross-GPU duplication under TP; the GPU pool is shared across requests with a bucketed parallel LRU for…
An LLM evolves executable semantic-index programs while a tiny retriever screens them. With the target training recipe fixed, NQ320K Recall@100 rises from 85.82% to 90.57%.
ECHO makes 100K-event histories affordable by retaining reusable original records before candidate-aware retrieval. A full-history teacher supervises selection, while retrieval and aggregation share relevance parameters. Kuaishou reports a 0.613% increase in…
KuaFu avoids repeatedly encoding the same content across users and profiling tasks. It caches compact item representations, assembles them chronologically and uses staged training to preserve evidence. Tencent reports a 1.37% overall GMV increase against a…
Generative retrieval need not encode every available feature up front. AGR requests intent, location, or behavioral evidence while decoding item identifiers, then resumes generation with a compact observation. A Kuaishou online experiment reports a 0.205%…
CTR self-distillation can be replaced with a computable disagreement penalty. FLMNet keeps gates on clean inputs and distills independently trained towers into one serving model. The distilled model reaches 0.8169 AUC on Criteo.
DuoLM precomputes one vector per item and uses a personalized ranker with user history and feedback. Shared history computation and a specialized runtime reduce online work. Throughput reaches about 26,400 candidate scores per second on one H100.
Encode articles separately, then learn which ones matter. LSNR combines weighted sparse article vectors with a user-level semantic expansion, improving MIND NDCG@10 by about 11% over STAR.
Predictable Scaling evaluates ranking-model changes with reduced-scale proxies and estimates their full-scale effects using historical learning curves. Training compute falls by 88% and 79% on two industrial ranking models.
Gryphon-v2 encodes user history once, generates semantic item IDs, and ranks the resolved items with a small shared-state head. A larger teacher supplies training targets but stays out of serving. Replacing the Yandex Music cascade increased active users by…
TransX caches reusable behavior memory and separates sequence encoding from candidate-conditioned scoring. Offline prefill and nearline refresh keep history processing off the request path. Its base model raised CTR by 6.02% over the production baseline in an…
RecSoup makes accuracy–diversity trade-offs adjustable by reusing objective-specific checkpoints. It prunes task-vector updates and learns preference-dependent merge coefficients. Parameter construction was about 165–228 times faster than a hypernetwork…
DualGCN gives interaction-free items a representation through semantic-anchored graph distillation. A separate student learns directions, while inference restores vector magnitude. On Sports cold items, Recall@20 exceeds DropoutNet by 20.8%.