Section 1: 📊 Trend Analysis

• 🔥 Agentic recommendation enters a "self-evolving" phase: Two 5/5 papers today — E-GenRM (Taobao) and EvoRec (Kuaishou) — point the same direction. Neither settles for one-shot agent reasoning. Both add a reward model or reflection tree to close the loop, so the system keeps correcting itself from user feedback without parameter updates. Industry has moved from "can agents work" to "how do agents evolve themselves." This is the main thread to follow this year.

• 💡 "Utility alignment" in generative recommendation is now a shared pain point: UAGR (Kuaishou), HERO (Alibaba 1688) and GRC (Alibaba) attack the same problem from different angles — generation likelihood drifts from downstream business utility, so high-value candidates get pruned too early in beam search. The fixes converge on injecting multi-objective utility into SID construction, decoding scores, or distillation routing. Pure semantic IDs are no longer enough; utility signals have to move upstream.

• 📈 Experiment measurement and training-system infrastructure are heating up together: Microsoft's SSR metric shows that under generative search, "click ≠ success" — good abandonment accounts for 53.2% of cases. SparseLift works at the compiler level to lift 100B sparse-parameter training throughput by 16.5%. As model-side gains flatten, measurement precision and training efficiency become the new source of industrial upside.

Section 2: 📋 Quick Glance

• Alibaba/Taobao builds a 27B generative reward model for open-ended e-commerce shopping agents. It constructs long-horizon trajectories from real shopping behavior and combines SFT with rubric-guided RL to sharpen preference judgments. E-GenRM-Bench preference accuracy hits 87.40% (6.97pp above the strongest baseline), and a 14-day Taobao A/B shows effective shopping assistance rate +1.39pp. ↗

• Kuaishou proposes EvoRec, a unified self-evolving agentic recommendation framework. Three agents — Need, Planning, Acting — cover the full decision chain, and a dynamic reflection tree distills replay-verified fixes into cross-user policies. Local services ads Hit@50 +15.7%; 7-day online ad revenue +4.6%, CTCVR +3.5%. ↗

• Kuaishou releases SparseLift, a sparse-path compiler. It rewrites opaque sparse nodes in the training graph into analyzable subgraphs, then uses an incremental cost model and a constraint-aware planner to generate execution plans automatically. Steady-state throughput on a 100B sparse-parameter industrial task rises 16.5%; on DLRM it is 1.26× faster than TorchRec. ↗

• Kuaishou proposes UAGR, which injects multi-objective utility into semantic ID construction and beam decoding scores. It supports incremental utility updates without growing SID length or vocabulary. Fully deployed at Kuaishou, with online video watch time +0.206%. ↗

• Microsoft/Simplex AI proposes the SSR session-success metric for generative AI search, training a model on human labels to detect good abandonment online. Balanced accuracy is 0.75 on 6,904 labeled queries; across ~4,000 production experiments its sensitivity is 2.5× that of Sessions per User. ↗

• Alibaba proposes HOB, a cross-channel auto-bidding framework that makes marginal cost computable and alignable across heterogeneous auction mechanisms. It models free wins and price uncertainty with a zero-inflated exponential distribution. Live on a large commercial DSP, GMV +3.0% while satisfying ROAS constraints. ↗

• Tencent proposes CADRE, a multi-agent relevance judgment framework. Three expert agents share a frozen backbone, a Verifier guides debate to resolve disagreement, and the result distills into a lightweight student model for deployment. Offline and online A/B on an e-commerce search ads platform both improve relevance classification and satisfaction metrics. ↗

• Alibaba/Tsinghua proposes ChoirRec, which uses LLM semantics to build cross-activity user groups and enriches sparse embeddings of low-activity users with group-level priors, fused adaptively through two channels. Taobao offline GAUC +1.16%, online orders +7.24%. ↗

• CAS/Baidu systematically compares LLM labels against click data in LTR: click supervision wins on high-frequency queries and document-quality signals, LLM labels win on mid/low-frequency queries and semantic matching. They propose frequency-aware multi-objective learning to fuse both, improving ranking across all query frequencies. ↗

• Alibaba/Wuhan Univ proposes MUSE, a multimodal lifelong interest modeling framework. GSU retrieves with lightweight cosine similarity; ESU does rich multimodal sequence modeling with ID fusion. On Taobao display ads it supports 100K-length behavior sequences and releases the first ultra-long-sequence multimodal dataset. ↗

• Alibaba fully deploys the LEAPS plugin in Taobao AI search. Its Broaden-and-Refine paradigm expands query combinations upstream and filters noise downstream with a CoT relevance verifier. Serving hundreds of millions of MAU since August 2025, the non-invasive architecture preserves short-text retrieval performance. ↗

• JD/CASIA proposes OSMF, an ad image generation framework. Product-aware adaptive grouping plus Group-DPO aligns each user group's click preferences, and they release the GAIP dataset with 600K groups and 40M users. SOTA offline and online, easing the group-level suboptimality of one-size-fits-all generation. ↗

• Kuaishou/BIT proposes UniMVT, a unified causal framework that disentangles marketing-intervention confounders and jointly reconstructs debiased base CTR and multi-value coupon intensity response curves. It beats baselines on prediction accuracy and calibration across synthetic and industrial datasets, and online A/B confirms gains in coupon distribution metrics. ↗

• Alibaba/Wuhan Univ proposes GRC, a reflection-and-correction decoding framework for generative recommendation. Decoding splits into draft generation, multi-granularity reflection and correction, with GRPO optimizing the full trajectory. Up to +15.74% over 6 SOTA baselines, online ad revenue +1.79% with controlled latency. ↗

• Tencent/WeChat proposes WeSearch, a unified model-centric retrieval framework. It first retrieves query-relevant clusters, then ranks atomic documents within each cluster, decoupling pruning from fine-grained relevance modeling. Fully deployed in WeChat search: CTR +0.68%, article clicks per user +2.27%, retrieval resource cost down 41%. ↗

• Meituan/Tianjin Univ proposes SA-RSQ, a sparse residual soft quantization framework. Top-K sparse routing plus softmax weights store compact tuples, decoupling per-item storage from codebook dimension. One-week A/B on Meituan food-delivery ads: CTR +2.51%, CPM +3.66%. ↗

• Amazon summarizes ad CVR auxiliary supervision practice, supplementing sparse conversion labels with diverse user events and metadata signals. It proposes linearized labels, a split architecture, and the LaCo noise-robust label combination. Production ranking models show measurable gains, and the authors stress transferability to the recommendation domain. ↗

• Alibaba proposes HERO, a diversity-aware on-policy distillation framework. Heterogeneous teachers route by popularity, and cold-item SID paths are injected to ease popularity bias. A 15-day A/B on the 1688 homepage "Guess What You Like" yields GMV +3.412%. ↗

• Unknown proposes DCD, a post-click destination personalization framework. It evolves from two-task conversion prediction to a single-task policy model, trained with A2C-inspired soft labels. Large-scale web-vs-app A/B on a major ad platform shows per-click personalization beats static routing. ↗

• Alibaba/NJU systematically evaluates LLMs' ability to identify implicit related touchpoints in conversion attribution, exposing a wide gap between collaborative-filtering rules and user semantic intent. LLM attribution labels used to augment industrial CVR training deliver significant offline gains. ↗

• Waseda Univ proposes PACT, a model-agnostic defense framework. It pools multi-generator, multi-strategy rewrite attacks and trains for score consistency to resist LLM promotional rewrites. It achieves the lowest promotion rate across nine settings, up to 3.0pp below the strongest baseline, with zero extra inference latency. ↗

• Unknown identifies a "first-step coverage bottleneck" in generative recommendation beam search: every pruned top-10 path is lost at the first step. Reallocating budget lifts Search Recall@10 by 0.96–10.90pp at Beam-50 and cuts matching-output decoding latency by about 6.4%. ↗

• Unknown releases NumEmbBench, the first open benchmark for CTR numerical embeddings. It groups 18 configurations into six mechanism classes and evaluates them uniformly across 3 datasets, 9 backbones and 5 seeds. DEER ranks best on average, but each of the six methods has strengths, and accuracy does not correlate with inference cost. ↗

• Unknown proposes SIDRec, an end-to-end recommendation-aware semantic ID learning framework. It self-updates fused content representations and collaborative embeddings, and uses MMoE-style joint optimization so recommendation supervision directly guides differentiable quantization. It leads on AUC/GAUC across three datasets, with up to 5.15% relative gain. ↗

• Unknown proposes IncVNM, an interactive recommender that incrementally learns a user's VNM utility function from ordinal click feedback, with entropy-regularized sampling to balance exploration and exploitation. In rolling-window experiments it clearly beats expected-revenue, mean-variance and discounted-UCB baselines across response error rates. ↗

• CAS/Baidu proposes LLM-QL, which uses query-likelihood maximization as an auxiliary task to strengthen dense retrieval backbones, with attention blocking and document corruption as two components. It beats other LLM retrievers both in-domain on MS MARCO and out-of-domain on BEIR. ↗

• CityU/NTU/Sydney proposes MUSE, a multi-teacher reasoning distillation framework. It builds a diverse reasoning pool, expands it with student self-exploration, and filters trajectories by Rank-Gain utility for two-level contrastive learning. Three real datasets confirm effectiveness and robustness. ↗

• SJTU proposes RecRSI, a mechanism-hypothesis-driven framework for automatic recommendation algorithm discovery. MechSpace derives realizable primitives, and Research Line iteratively validates hypotheses. Across four initial recommenders and four datasets, NDCG@5 rises 8.48% on average and HR@5 6.88%. ↗

• Northeastern Univ/Tencent proposes PACERec, an ID-free progressive adaptation framework. Frequency-weighted flow matching learns a shared velocity field that transfers frequent-item interaction context to the long tail. It beats baselines on overall and long-tail recommendation across three datasets and three sequence backbones. ↗

• Istanbul Tech systematically studies how fine-tuning data preprocessing affects LLM recommendation utility and fairness, building a controlled pipeline of SFT + catalog-constrained decoding + prompt augmentation. On MovieLens-1M, raw titles lift HR@10 from 10.67% to 12.47%, and fairness effects vary by attribute. ↗

• Unknown proposes ACRC, average conformal risk calibration. It averages first, then truncates, to restore valid activation of recommendation rules under risk control. Across six datasets, post-fold net utility improves 0.010226 over raw, and Jester independently confirms +0.011025±0.001082. ↗

• Unknown studies fairness in value-weighted ranking. It defines group complaint disparity and introduces complaint penalties and group multipliers to explicitly price complaint reduction. On multiple public datasets, group-aware penalties close complaint gaps for some cohorts but establish no significant average advantage. ↗

• Tsinghua proposes CL-DGFM, a dynamic graph foundation model. A dynamic interaction tokenizer encodes interaction memory and ordered transitions as tabular input, so tabular foundation models can be reused without a graph encoder. It is competitive on TGB/TGB-Seq and supports direct transfer with a frozen foundation model. ↗

• PKU/Kuaishou proposes CURec, a collaborative-perspective content understanding framework. It first trains an LLM to give recommendation-aligned reasoning, then uses recommendation-model scores as reward for RL refinement of rationales. It beats existing methods on public benchmarks, narrowing the semantic-collaborative gap. ↗

• Tsinghua/Utah moves fair ranking from exposure fairness to context-dependent revenue fairness. DIDRF exploits the quadratic structure of revenue-fairness violations to derive state-aware scoring rules. In semi-synthetic experiments calibrated on ad and e-commerce logs, revenue fairness consistently beats baselines while preserving ranking quality. ↗

• CityU/Huawei/DUT proposes R²-Searcher, a retrieval-reasoning boundary calibration framework. Query tokens guide fine-grained evidence modeling and retrieval reflection, and R²PO reinforcement learning jointly optimizes both. It leads on answer accuracy and retrieval-reasoning quality across seven multi-hop QA benchmarks. ↗

• Unknown proposes SA-OPSD, search-aware on-policy self-distillation. Evolving boundary sampling prioritizes critical search states, and confidence-decoupled branch distillation separates target calibration from sibling-relation supervision. It performs strongly on Industrial/Office/Instruments. ↗

Section 3: 📰 Daily Digest

1. E-GenRM: Generative Reward Modeling for Open-Ended E-Commerce Shopping Agents

2. EvoRec: A Unified Self-Evolving Agentic Recommender with Dynamic Reflection Trees

3. SparseLift: Automatically Optimizing Opaque Sparse Paths in Industrial-Scale Recommendation Model Training

4. Utility-Aware Generative Recommendation: Utility-Enriched Semantic Codes and Utility-Guided Decoding

5. Measuring User Success in Randomized Search Experiments: Interpreting Good Abandonment in Generative AI Search

🎯 Today's Theme: Reward Signals for Agentic Recommendation — Generative RM or Verifier Judgment?

Trajectory-level or evidence-level: reward granularity decides whether the RM can locate errors