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

EvoRec · Reflection-Driven Self-Evolution

Mengting Li*, Long Zhang*, Hao Jiang†, Sheng Yu, Fei Pan, Peng Jiang

Kuaishou TechnologyKuaishou Local Services Ads Recommendation Foundation Model Team

* Equal contribution  ·  † Corresponding author

▤ Paper⌘ Code
Comparison of conventional cascaded recommendation and EvoRec
Recommendation should learn from outcomes. EvoRec replaces a fixed cascade with an autonomous loop of need understanding, planning, execution, feedback, reflection, and repair.
+15.7%Hit@50
+4.6%Revenue
+3.5%CTCVR
+14.5%High-intent revenue

Case Study

An anonymized real episode: follow how user evidence becomes six local-service needs, two recall routes, a missed conversion, and an ERROR-NEED repair.

👤Long-term MemoryUser profile & stable habitsProfile · preferences · demand · intent · click to expand⌄
user_profile

31–40-year-old working mother in corporate finance, living in Jianshui, Yunnan; quality-oriented and balances family, career, and aspirational consumption.

interest_preference

Strong: short dramas, cars, and finance · Medium: child education · Weak: entertainment; prefers comeback and family-ethics stories.

consumption_behavior

Value-conscious family shopper with paid short-drama habits, skincare brand loyalty, and content-driven or impulsive purchases.

local_life_demand

Medical care · financial services · auto sales/service · local dining · beauty and nail care.

user_query_intent

Throat diseases · score 3

09-14 22:07 · P3S · SID …9248Debt relief through rapid-marriage service
09-14 10:27 · P3S · SID …1578Guidance for resolving a frozen bank card
09-13 15:28 · P3S · SID …2555Used-car trade-in and subsidy information
09-13 13:40 · P3S · SID …4203Professional marriage-law consultation
09-13 00:16 · P3S · SID …3728Eye-care and cosmetic treatment process
▱
Working Memory
Current recommendation episode

Active evidence: recent P3S sequence

Pending output: prioritized local-service needs

AgentNeed
Inference
Inferred needs
HighENT clinicExplicit search + click257
MedBeauty eye-careSingle direct signal172
MedUsed-car tradingSingle direct signal171
HighMarriage legal serviceRepeated evidence257
HighFinancial lendingRepeated evidence257
LowShort-drama contentLong-term interest only86
Route 01Behavior ContinuationContinue demonstrated interests
Success1,800 returned
Budget · 1,800
Route 02Need SatisfactionPriority-weighted allocation · 3 : 2 : 1
Success1,200 returned
Budget · 1,200
Planning AgentAllocate · Select · FuseBC 60% · NS 40%
Retrieval tool pool
Item-CFUser-CFEmbeddingBM25OneReason
→
Candidate mergeSnake-order interleaving3,000 recalled · deduplicate lists
→
Rank
Base fixedSASRec αTwo-Tower 1−α
BC1,800 recalled
NS1,200 recalled
→
Merge & deduplicate3,000 candidatesSnake-order fusion
→
Rank
BasefixedSASRecαTwo-Tower1−α
Top recommendations
Recommendation outcome
  1. BC1,800 candidatesSuccess
  2. NS1,200 candidatesSuccess
  3. Σ3,000 returnedRanked
  4. ···
  5. ×Deposit-free phone rentalMissed by both routesSuccessfully recalled · Rank 23
Dynamic Reflection TreeMatch the observed failure to a reusable error pattern
Observed feedbackConverted item is absent from the ranked list
↓
ERROR-MEMORYERROR-NEEDERROR-PLANERROR-EXEC
↓
Evidence missingERROR-NEED-001Evidence present but need omittedPriority mismatch
Matched strategyCross-check latent acquisition needs before retrieval planning
ERROR-NEED-001Evidence present but omittedMissing need: low-threshold phone rentalRepair: jointly model cash-flow pressure, price sensitivity, and durable-goods acquisition.Repair inserted into Need Agent and Plan · ready to replayReplay verified · repaired recommendation succeeds
1Reflection NodeERROR-NEED-001

Recover the omitted phone-rental need

→
2Working MemoryWrite repair evidence

Cash-flow pressure · price sensitivity

→
3Need AgentReason again

Infer low-threshold phone rental

→
4Re-runRetrieve & Rank

Execute the repaired plan

→
✓Correct ResultSuccessfully recalled

Phone rental · Rank 23

Memory contextLong-term profile, short-term behavior, and working context jointly support need inference.
Updated EvoRec case study from a missed need through hierarchical diagnosis to repair and replay
Paper schematic. This full reflection round illustrates hierarchical ERROR-NEED diagnosis, personalized repair, and replay verification; the interactive episode above reinserts the repair, reruns recommendation, and verifies a successful hit.

Abstract

Conventional recommender systems follow cascaded retrieval, pre-ranking, and ranking pipelines with predefined models and offline-optimized configurations, limiting their ability to adapt to evolving user needs and observed outcomes. Recent agentic recommenders introduce reasoning and tool use but largely focus on isolated stages, lacking both end-to-end decision control and systematic reflection on recommendation failures.

We propose EvoRec, a unified agentic recommendation framework with dynamic reflection that integrates end-to-end autonomous recommendation with reflection-driven self-evolution. A Need Agent infers structured needs from user memory, a Planning Agent adaptively allocates retrieval budgets and selects tools and ranking-fusion weights, and an Acting Executor deterministically executes the plan. A Dynamic Reflection Tree then localizes failures, generates personalized repairs, verifies them through replay, and evolves reusable strategies without updating agent parameters. EvoRec improves Hit@50 by 15.7% over the strongest baseline; a seven-day online A/B test increases advertising revenue by 4.6% and CTCVR by 3.5%.

Method

Updated EvoRec architecture with autonomous recommendation and a Dynamic Reflection Tree
Figure 2. EvoRec unifies autonomous recommendation and reflection-driven self-evolution. The upper path performs need inference, adaptive planning, and deterministic execution; the Dynamic Reflection Tree performs hierarchical error diagnosis, user-level repair, and system-level tree evolution.

Results

Performance comparison of recommendation methods
Table 1. EvoRec achieves the best results across all accuracy metrics, improving Hit@10 by 19.0%, Hit@50 by 15.7%, and NDCG@10 by 17.5% over the strongest baseline.

Online Evaluation

Online A/B results within the deployment scope
Table 6 · Seven-day online A/B test. Revenue rises by 4.6% overall and 14.5% for high-intent users.
Online CTCVR comparison across eight industry categories
Figure 4 · Category-level online results. CTCVR improves across all eight anonymized industry categories.

Ablations & Analysis

Ablation comparing agentic decisions with static variants
Table 2 · Agentic vs. static decisions. Removing adaptive decisions reduces Hit@50 by 3.1%–13.4%.
Retrieval-tool ablation results
Table 3 · Retrieval-tool collaboration. All tools contribute, with Item-CF and User-CF providing the largest gains.
Ablation of ranking-signal fusion strategies
Table 4 · Ranking-signal fusion. Agentic per-user fusion raises Hit@50 from 23.28% to 23.90%.
Ablation of Dynamic Reflection Tree components
Table 5 · Dynamic Reflection Tree. Personalization, evolution, memory-error handling, and verification all contribute.
Relative Hit@50 gain over accumulation and evaluation rounds
Figure 5 · Evolution over rounds. As verified reflective experience accumulates, EvoRec's relative Hit@50 gain grows to 2.9% by day D+4, while variants without tree evolution or ERROR-MEMORY plateau.

Contact

Contact

Please contact us.

Kuaishou Local Services Ads Recommendation Foundation Model Team
First authorMengting Limengtingli000@gmail.com
Team leaderHao Jiangjianghao11@kuaishou.com

BibTeX

@misc{li2027evorec,
  title     = {EvoRec: A Unified Self-Evolving Agentic Recommender with Dynamic Reflection Trees},
  author    = {Li, Mengting and Zhang, Long and Jiang, Hao and Yu, Sheng and Pan, Fei and Jiang, Peng},
  year      = {2027}
}