DREAM Technical Report
Bin Zhang, Bowen Zheng, Chao Yi, Chengyu Lai, Dian Chen, Dimin Wang, Gaoyang Guo, Jialin Zhu, Jian Wu, Jing Yu, Jiuning Lin, Lingqing Zhang, Lingyun Zheng, Mao Zhang, Mingming Pan, Ruiquan Lan, Shuai Zhong, Wen Chen, Wendong Zhang, Xiaodong Zhu, Xuan Chen, Xunke Xi, Yifan Lu, Yiheng Wang, Yue Zeng, Yujie Luo, Yuning Jiang, Zhe Hu, Zhibo Xiao, Zihong Huang, Binbin Cao, Bo Zheng, Danning Wang, Dixuan Wang, Ge Fan, Haixia Wu, Han Zhu, Hao Fang, Haoming Chen, Huiping Chu, Jian Wang, Jianjun Wu, Jiawei Wu, Jiaxin Yu, Jingwen Liu, Jinzhe Shan, Kai Meng, Kai Zhang, Keqin Xu, Kewei Zhu, Lang Tian, Leihui Chen, Li Chen, Licheng Xu, Lide Xiao, Ruitong Zhang, Shiyao Peng, Silu Zhou, Tao Wang, Wei Shi, Wenjun Yang, Xiang Chen, Xiang Gao, Xiao Ren, Xu Liu, Xuwen Wang, Yang Li, Yeqiu Yang, Yi Hu, Yichen Yuan, Yinnan Song, Yipeng Yu, Yuan Liu, Yunqi Gao, Zhiliang Huang, Zhujin Gao, Zongyuan Wu
cs.IR
2026-08-10
DREAM overlays intent and meta-control on Taobao's cascade: re-rank lifts IPV 2.06%, Core IPV 2.39%, GMV 0.88%; with fine rank these become 2.71%, 3.06%, and 1.31%.
Industrial recommenders still run as cascaded retrieval, ranking, and re-ranking. The split is stable and fast. It also fragments information across stages, scatters click, conversion, and experience objectives, freezes policy in static rules, and barely sees a session move from browsing to comparison to purchase.
Agentic recommendation work usually covers one slice: memory-based item selection, user simulation, dialogue, or architecture search. The missing piece is a control plane that perceives live intent, writes cross-module strategy, and feeds execution back upstream. DREAM (Developing Recommender Engine with Agentic Methods) adds that plane on top of the existing pipeline. It does not replace the rankers.
Two engines and a dual reward loop.
The Intent Engine folds on-device and server signals into three layers: L0 physical profile, L1 demand (category and scenario), L2 preference (brand, price, decision stage, real-time psychology). On device, F1-F3 encode events and run a small GRU change-point gate; about 15% of behavior is packed as ID-encoded reports. Cloud F4 restores text and decides whether to call the reasoner. The abstract states that the full chain cuts reporting volume to about 8.7%. The online path is a 0.8B Main Agent that emits only insert/update deltas. About 6.3% of requests go asynchronously to 4B context or expert agents; on-policy distillation writes those refinements back. At night a 4B Dreaming pass consolidates the full-day trace with keep, correct, enrich, merge, add, and kill, replacing the incremental daytime state.
The Meta Engine's MetaModel is built on Qwen3 and reasons nearline in three layers. M1 compresses intent into an IPV-oriented or GMV-oriented orientation. M2 emits a six-module enumerated JSON bundle: objective boosts, business protection, category and card-type preferences, experience constraints, position policy. M3 is a deterministic compiler from those semantic actions into bounded service parameters. At request time the system only reads the cache and applies a local override; missing, stale, or out-of-range bundles fall back to defaults. Guardrails cover ranges, allowlists, and traffic caps.
The Reward Dual Loop splits exploration from calibration. Offline, an Evaluator replays logged requests; a strategy gets positive reward only if its mean score beats the default pipeline, so the training signal is binary. Online, live IPV, CVR, and GMV write into Strategy Memory. The LLM stays off the latency-critical path.
On Taobao Homepage Guess You Like, widening the control surface from re-ranking into fine ranking yields these relative lifts over production (absolute counts withheld):
| Config | PV | IPV | Core IPV | GMV | PCTR |
| Re-rank only | +1.03% | +2.06% | +2.39% | +0.88% | +0.76% |
| Re-rank + fine rank | +1.04% | +2.71% | +3.06% | +1.31% | +1.25% |
PV barely moves between the two treatments. The extra stage converts the same exposure more effectively. When the Intent Engine is wired in alone, cognitive recall lifts platform IPV by 0.80% and Core IPV by 0.91%; strategy adaptation lifts IPV by 0.52% with PCTR flat at -0.02%. Inside heuristic recommendation, inquiry-card clicks rise 7.17% and 10.64% for the two intent-consuming treatments.
Intent quality against later behavior, judged by an LLM: overall 71.32 baseline, 78.20 with routing, 84.74 after self-evolution; search-term recall 50.57% to 56.05%. Dreaming on 683 paired users moves intent type 0.583 to 0.680 and subcategory 0.477 to 0.523. Offline 4B replay RL versus a same-size Base: pCTR +2.42%, pIPV +1.38%, pGMV +0.37%, pCVR -0.99%, and valid executable bundles 80.86% to 98.85%.
For teams that already own a stable cascade, an overlay is more shippable than rewriting recommendation as an agent. Old models keep serving. The agent turns bounded knobs, and the default path is the safety net. The device-cloud funnel keeps reports under one tenth of raw behavior; 0.8B handles the live path and 4B sleeps over the day's trace. Homepage lifts on both IPV and GMV show that meta-control is not a zero-sum trade in this setting.
These are one-to-three-point gains on an already strong production baseline. Systems without on-device behavior packs or re-rank knobs should not expect the same percentages.
There is no standalone limitations section. Online tables report only relative lifts; traffic volume, significance, and confidence intervals are absent. Intent quality leans on LLM-as-a-Judge and drops users with thin follow-up behavior, which can bias the sample. The 8.7% reporting figure lives in the abstract; the body only pins F2 at about 15%, with no multiplication table for the full funnel. Replay cannot reconstruct the historical serving state, and the binary reward ignores lift magnitude. The exact MetaModel size used in the A/B is not pinned to a Qwen3 variant in Table 7. On-device micro-signals are Taobao-client specific. Wider control still cannot outrun the rankers underneath.