PERSONALIZATION · 2026-10-23
Personalization — Paper Daily, 2026-10-23
Community search over short user-generated posts, questions, answers and comments. Key metrics are search relevance (nDCG) and CTR. Focus: semantic matching, query understanding, short-text retrieval, ranking models, and LLM applications in search.
4.85
LEAD STORY ·
AlibabaAlibaba 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 to 4×4, tokenizes titles and OCR text by field, and looks them all up like ordinary ID features; the shared network is first trained on query–item relevance labels from a Qwen3.5 teacher, then fine-tuned with a click head on click logs. Offline AUC is 0.25 points above the HiFormer baseline and 0.38 points higher on cold items; a seven-day Taobao main-search A/B test raised GMV by 0.85%.
4.25
USTCGenerative 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 tail targets. CAPE does not retrain the model; at inference it counts which items co-occurred with the user's history in the training data, retrieves 25 extra tail candidates by that count, and adds the count into the ranking score. On LastFM, TIGER's top-10 hit rate on tail targets rises from 3.94% to 17.19%, with 11–16% extra latency per request.
4.20
Duke UniversityGGeorge Washington UniversityPersonalized 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 whole and a single reply cannot be judged with one threshold. PerSPL pushes each labeled pair's mean score toward 0 so scores are comparable across prompts, fits a monotonic score-to-preference map per user, and only pseudo-labels high-confidence unlabeled replies before iterative retraining. On PersonalLLM, unseen-user accuracy is 95.5% versus 94.0% for the strongest baseline LoRe.
4.10
Renmin University of ChinaWhen 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 third. LoopRec lets each historical behavior choose how many layers to stack instead of sharing one depth for the whole request. On the Amazon Software dataset with a SASRec backbone, NDCG@20 rises 5.68% relatively.
3.95
ResearchHub recommends biomedical preprints over 44,937 MedRxiv papers from 2023–2025. Its core retriever PRISM is a dual-encoder: GPT-4o first writes researcher-style queries for each abstract, then contrastive loss plus cross-encoder distillation (MarginMSE) fine-tunes the encoder. On those queries Recall@10 is 0.892, 12.4 points above SPECTER2-Search. Optional post-steps prefer results farther from the user's saved papers, DPP reranking to cut near-duplicates, and ordering from mainstream to specialized by local intrinsic dimension in embedding space.
3.95
Long-horizon assistants must remember user preferences and update them when preferences truly change. Common practice appends or overwrites by rule each turn, so one casual remark can overwrite an old memory. PreCoMem stores a confidence on each memory and computes a surprise score each turn: how far the new content is from retrieved memories and whether they conflict; the score is lowered when retrieval is diffuse. Low scores reinforce existing memory, mid-range scores become low-confidence hypotheses, and only high, reliable scores down-weight old facts. On LongMemEval-S with gpt-4.1 answering, overall accuracy is 89.39% versus 86.16% for EverMemOS.
3.75
AAristotle University of Thessaloniki
SOAR recommends points of interest on Yelp by folding review sentiment and friend ties into the score: each place gets an initial score from public review polarity, then friends' reviews add or subtract. Versus a baseline without reviews or friends, Top-10 recall rises from 47.01% to 56.59% and Top-1 from 13.37% to 16.37%. Expanding friends to friends-of-friends barely changes quality while prediction speed falls from 55.56 to 7.32 requests per second. The paper compares only against this self-built baseline.
3.65
RAG stages—embedding models, chunking, retrieval, reranking, query rewrite, context compression, agentic checks—are usually evaluated alone. LASAGNA runs 450 configs on 8 benchmarks with corpora from 5K to 200K documents to measure how much single-component gains survive when combined. The three strongest retrieval techniques alone would sum to NDCG@10 +47%, but together only +17%; the components that raise retrieval are not the same ones that raise answer faithfulness. A setup that leads on a small corpus trails by 13.7% NDCG@10 at 200K documents, with rank flips starting at 25K.
3.55
MeituanMeituan moved Karpathy's AutoResearch (an agent edits training code, runs experiments, and keeps or rolls back by eval) onto its cloud training platform and rebuilt it three times in three months. V1 triggers a single agent with a cron script; V2 opens a new session each trigger, restores progress from a JSON state file, and no longer waits synchronously for training; V3 has a lead agent split the problem into 3–4 mutually exclusive hypotheses explored by parallel branches and then merged. On two modeling projects, single-task time fell about 85% and 83% versus humans; one model shipped the system's proposals fully, and a multi-objective project beat the human baseline on all 12 prediction targets.
3.40
This is a framework paper with no new model and no experimental numbers. The authors argue that an agent's external memory should be judged by whether later actions improve, not only by whether old facts can be recalled. They classify memory along three axes: where state lives (context, external store, model parameters, …), what form it takes (episodic records, semantic knowledge, procedures), and which write/update policies apply. Evaluation should test retention under distraction, adaptation after environment change, and cross-task transfer, with "keep memory" versus "reset every time" controls so memory gains are separated from simply lengthening context.
NeurIPS 2026