01
BPR
Recommender Systems Began as Scoring Machines
Start with implicit feedback, positive–negative pairs, and pairwise loss to establish the classical paradigm: learn s(u,i), then rank.In progress
Meta · Multimodal LLMs
Editorial series · 00–20
How Recommender Systems Learned to Say the Answer: Generative Recommendation in 20 Papers
From BPR’s scoring function to OneReason’s recommendation reasoning
How did recommender systems evolve from scoring every item to directly generating items, lists, explanations, and even an explicit reasoning process?
阅读中文版 →Prologue · published
What Does Generative Recommendation Actually Generate?
Before the papers, distinguish language, item IDs, complete lists, and a process that merely looks like reasoning.
Start with the prologue →The story in one map
One question splits into two branches, then converges again.
BPRScoring & ranking
GRU4Rec → SASRec → BERT4RecBehavior sequences
05–15What does the model generate?
P5 → LLaRAGenerate language answers
DSI → ETEGRecGenerate item identifiers
HSTU → OneRec → OneReasonUnification, scale & recommendation reasoning
The 20-paper roadmap
Every paper is motivated by what the previous one could not solve.
01–04
Prehistory: from scoring to sequences
First learn to compare items, then learn that behavior has an order.
02
GRU4Rec
A User Is Not a Point, but a Story in Motion
Preferences change, so recommendation begins to model the order in which behavior unfolds.In progress
03
SASRec
History Is Long, but Only a Few Steps May Matter
Replace recurrence with causal self-attention and let the model find which past actions matter for the next one.In progress
04
BERT4Rec
Treat Items as Words and Shopping Histories as Sentences
Item IDs become tokens, histories become sentences, and Cloze-style masked modeling enters recommendation.In progress
05–09
The LLM branch: recommendation as language
Unify tasks and transfer knowledge—while exposing the gap between language fluency and preference understanding.
05
P5
Can Every Recommendation Task Be Rewritten as a Sentence?
Unify rating, sequential recommendation, and explanation as text-to-text instead of separate task heads.In progress
06
M6-Rec
One Model for Retrieval, Ranking, Explanation, and Creation
Move from academic unification toward an industrial foundation model through prompt tuning, open tasks, and efficient inference.In progress
07
TALLRec
A General-Purpose LLM Does Not Naturally Understand “Preference”
Expose the domain gap between language competence and recommendation competence, and explain why LoRA/SFT helps.In progress
08
ReLLa
Fitting in Context Does Not Mean the Model Understands It
Long histories placed directly in a prompt may hurt, motivating behavioral retrieval and relevant-history selection.In progress
09
LLaRA
How Do World Knowledge and Collaborative Behavior Meet?
Classical recommenders know co-occurrence and LLMs know semantics; this paper aligns the two representation spaces.In progress
10–15
The Semantic ID branch: generate the item itself
From generating document IDs to designing recommendation tokenizers, items become decodable tokens.
10
DSI
What If Retrieval Generated IDs Instead of Looking Them Up?
The ancestor of generative retrieval: make document IDs output tokens and learn P(ID|query) directly.In progress
11
TIGER
Give Every Item a Semantic Address
Introduce RQ-VAE, Semantic IDs, token-by-token decoding, and their benefits for retrieval scale and cold start.In progress
12
GPTRec
Top-K Is Not the Only Answer: A List Can Be Written Item by Item
Use SVD sub-IDs and Next-K generation to optimize list-level objectives such as diversity.In progress
13
LC-Rec
Semantically Similar Does Not Mean Liked by the Same People
Separate language semantics from collaborative semantics through quantized codes and alignment objectives.In progress
14
LETTER
A Good ID Must Understand Content, Co-occurrence, and Balance
Examine the tokenizer problem through hierarchical semantics, collaborative regularization, and codebook collapse.In progress
15
ETEGRec
The Tokenizer Should Not Be Merely Offline Preprocessing
Jointly optimize the tokenizer and recommender through dual alignment objectives and alternating optimization.In progress
16–20
Industrial convergence: unification, scale, and reasoning
The two branches meet in long histories, unified retrieval-ranking, deployment constraints, and recommendation reasoning.
16
HSTU
Do Recommender Models Have Scaling Laws?
Meta reframes recommendation as sequential transduction and studies long histories, industrial efficiency, and scale.In progress
17
OneRec
Can One Model Retrieve and Rank in a Single Pass?
Use session-wise generation, Sparse MoE, and DPO to unify the multi-stage retrieve–rank pipeline.In progress
18
MTGR
Should We Really Throw Away Every Traditional Cross Feature?
Challenge the pure generation route with DLRM cross features, user-level compression, and deployment cost.In progress
19
OneRec-Think
Should Recommenders Think Before They Answer?
Combine item–text alignment, recommendation CoT, and preference rewards to pursue explicit, controllable reasoning.In progress
20
OneReason
Looking Like Reasoning Does Not Mean Understanding Preference
Reframe recommendation reasoning as perception plus cognition and leave the series with a genuinely open question.In progress
One reading protocol
Approachable to beginners, reproducible for engineers, and debatable for paper readers.
Beginner layer
A familiar example, the previous model’s limitation, and the new idea in one sentence.
Engineering layer
Inputs, outputs, tensor shapes, training data, inference, and complexity.
Paper layer
One central equation, one decisive result, one ablation, and what remains unproven.
One toy world across the series
The same user and items reveal how each generation changes.
The fixed history is “tennis → climbing → badminton → swimming.” Every article uses a before-and-after diagram, a data flow, an equation map, and an evidence figure.
🎾→
🧗→
🏸→
🏊