Generating an item crosses three gates: codebook, model distribution, and legal catalog¶

Updated 2026-10-30 · 62 topic entries / 7 representative comparisons

TIGER converts items to hierarchical codes and learns the next code sequence. COBRA adds a conditioned dense vector after sparse codes. PLUM continues pretraining an existing LLM in behavior and SID space. STATIC changes how legal decoding executes without changing representation; Reward Guided Decoding tilts existing support toward business preferences. Locate the changed gate before assuming an omitted item can be recovered. PinGER separates historical inputs from SID supervision and treats prefix diversity, hierarchical DPO, and catalog legality separately.

Mechanisms · Trade-offs · Cross-paper questions · Updates

How to read this card

The table compares representative methods. It is not a shared-benchmark leaderboard. Teaching examples are not measured results.

Mechanism comparison¶

Axes: Item representation: discrete codes, sparse–dense cascades, and LLM continual pretraining · What learned distributions and inference search change independently · Legal catalog constraints versus business rewards · Prior support, beam budgets, and small-scale evidence

Paper and source	Mechanism	Benefits and trade-offs
Recommender Systems with Generative Retrieval · Original source Sections in the Chinese source: 背景与主要改动, 核心公式, 本地复现	RQ-VAE quantizes content residuals in stages and adds a collision-resolution token to the item SID. An encoder–decoder autoregressively predicts the next item's code sequence.	Shared semantic prefixes share parameters, but code collisions and early path errors affect reachability. The local experiment sums token log probabilities over legal item IDs, not production beam throughput.
Sparse Meets Dense: Unified Generative Recommendations with Cascaded Sparse-Dense Representations · Original source Sections in the Chinese source: 背景与主要改动, 核心公式, 本地复现	Generates a sparse code, then a dense query conditioned on code and history. Cascades beam code scores with dense similarity.	Adds distinctions lost by quantization but introduces a dense index, fusion weights, and two-stage budgets. If a target code never enters the beam, the dense stage generally cannot recover it.
PLUM: Adapting Pre-trained Language Models for Industrial-scale Generative Recommendations · Original source Sections in the Chinese source: 背景与主要改动, 核心公式, 本地结果	Initializes an existing MoE LLM, aligns it with behavior and video metadata through continual pretraining, then retrieves via next-SID prediction. SID-v2 constrains reconstruction, quantization, and co-occurrence.	Shifts capacity from large ID tables to the backbone but depends on pretraining recipes and resources. Online SID generation avoids dot-product indexes, not SID mapping, catalog maintenance, or decoding costs.
Vectorizing the Trie: Efficient Constrained Decoding for LLM-based Generative Retrieval on Accelerators · Original source Sections in the Chinese source: 背景与主要改动, 核心公式, 结果与边界	Compiles legal paths into accelerator-resident CSR transitions. Runs full-vocabulary log-softmax before masking invalid tokens to negative infinity and tracks beam prefixes and node states together.	Changes constraint execution and legality without creating new preferences. Renormalizing over each legal subset does not preserve original model path scores.
Reward Guided Decoding for Generative Recommendation · Original source Sections in the Chinese source: 背景与主要改动, 核心公式, 复现边界	Fixes the generative prior P0 and tilts it by exp(R/β) to obtain Q. KL temperature controls deviation without retraining the generator.	Rewards reorder existing support, not paths with zero prior mass. A proxy reward is not online value; the local MovieLens construction does not measure production gains.
PailitaoMMSearch2.0 Technical Report: Scaling Generative Retrieval to Unified End-to-End Multimodal Search · Original source Sections in the Chinese source: 背景与主要改动, 核心公式	Maps joint image/text requests to SID sets and aligns multi-intent, mixed-budget training with Sample–Beam serving.	The complete multi-model, multi-stage system and private production experiments are too expensive to reproduce. No local placeholder model.
PinGER: Token-Efficient Generative Retrieval with Semantic IDs · Original source Sections in the Chinese source: 背景与主要改动, 核心公式, 高成本完整实验	Interleaves SID and dense inputs during training while blocking cross-item SID reads. Drops past SIDs during serving. Applies LCP-depth penalties per beam, with DPO over all SID prefixes and an NTP anchor.	Unlike TIGER's per-item historical SID inputs. Prefix diversity does not provide STATIC's legality guarantee and may reduce offline recall. The complete tokenizer and serving stack are not local.

Trade-off analysis¶

Check item reachability before ranking quality¶

TIGER and PLUM determine item-to-SID mappings and step probabilities. COBRA recovers detail within nominated sparse regions. Unresolved collisions, pruned prefixes, or catalog omissions cannot be fixed simply by increasing reward weight. Measure token prediction, valid item coverage, and final ranking separately. A wider beam is not proof of a better codebook.

Sources: Recommender Systems with Generative Retrieval · PLUM: Adapting Pre-trained Language Models for Industrial-scale Generative Recommendations · Sparse Meets Dense: Unified Generative Recommendations with Cascaded Sparse-Dense Representations · Reward Guided Decoding for Generative Recommendation

Legality and preference are different scores¶

STATIC decides which tokens can continue a prefix to a valid item. Reward Guided Decoding tilts the allowed distribution toward a reward. Negative infinity defines the feasible set; exponential tilting is a soft preference. A finite stock-violation penalty can still admit a high-reward invalid item. Keep constraint compilation, model probability, and business value distinct.

Sources: Vectorizing the Trie: Efficient Constrained Decoding for LLM-based Generative Retrieval on Accelerators · Reward Guided Decoding for Generative Recommendation

October 29: New method and source changes¶

PailitaoMMSearch2.0 directly decodes multi-intent image/text requests into item SID sets, scaling serving with data tools, continual pretraining, mixed-budget SFT, and intent-balanced RL. Sample–Beam aligns inference sampling and beam candidate generation with training and display needs respectively. Larger models or longer reasoning do not guarantee improvement on every intent. Compare on fixed candidate sets, budgets, and production blending rules rather than model names. The full system is not reproduced locally.

Sources: Recommender Systems with Generative Retrieval · PailitaoMMSearch2.0 Technical Report: Scaling Generative Retrieval to Unified End-to-End Multimodal Search

History inputs, candidate coverage, and legal paths have separate costs¶

TIGER represents historical items as multilevel SIDs. PinGER keeps levelwise SID supervision but restricts the backbone so later items see only historical dense inputs. Those dense inputs can still fuse SID information. PinGER's LCP penalty spreads beam slots across prefixes; it does not exclude nonexistent SIDs. STATIC compiles legal transitions. Greater spread has a trade-off: at λ=3 the paper reports higher distinct@1 and lower R@100. Do not attribute full-stack online lift to local mask or DPO components.

Sources: PinGER: Token-Efficient Generative Retrieval with Semantic IDs · Recommender Systems with Generative Retrieval · Vectorizing the Trie: Efficient Constrained Decoding for LLM-based Generative Retrieval on Accelerators

Cross-paper questions¶

Try answering before expanding the explanation. Each question relates at least two papers.

Why can't COBRA's dense vector guarantee recovery of items missed early in a TIGER-style code path?¶

Show answer and example

Answer: COBRA conditions dense queries on sparse codes, usually reranking within regions the beam nominated. If a target code is absent, the dense stage does not automatically search all omitted regions. It restores fine-grained distinction within a quantized region rather than repairing arbitrary recall misses.

Illustrative example: If the first beam keeps only running shoes and prunes photography, perfect dense ranking of shoes will not retrieve a camera.

Sources: Sparse Meets Dense: Unified Generative Recommendations with Cascaded Sparse-Dense Representations · Recommender Systems with Generative Retrieval

How should we interpret a conflict between STATIC's legal mask and high reward in Reward Guided Decoding?¶

Show answer and example

Answer: The mask defines eligible results. Reward expresses preferences within that set. Exponential tilting still multiplies the prior, so neither prior-zero nor mask-zero results can reappear. Replacing a mask with a finite penalty creates a different objective that allows violations to be offset.

Illustrative example: An out-of-stock item may have high predicted revenue but is still excluded by the catalog. Subtracting a little value does not guarantee exclusion.

Sources: Vectorizing the Trie: Efficient Constrained Decoding for LLM-based Generative Retrieval on Accelerators · Reward Guided Decoding for Generative Recommendation

Both TIGER and PLUM predict SIDs. Why doesn't the same next-token loss make them the same training recipe?¶

Show answer and example

Answer: TIGER builds a pipeline from content quantization and sequence generation. PLUM inherits an LLM, aligns language and collaborative spaces using behavior, text, and ASR continual pretraining, and adds co-occurrence and progressive-depth reconstruction in SID-v2. Shared outputs do not equal shared initialization, data, or representation learning. Fix catalog and budget when comparing.

Illustrative example: Two models predict three-level codes. One learns only behavior sequences; the other first sees mixed metadata and behavior. Calling both final losses cross-entropy does not erase that difference.

Sources: Recommender Systems with Generative Retrieval · PLUM: Adapting Pre-trained Language Models for Industrial-scale Generative Recommendations

Why can't Pailitao's complete online gain be attributed causally to Sample–Beam alone?¶

Show answer and example

Answer: The online test measures a complete deployment with training data, model, presentation, and blending. A controlled offline ablation supplies different evidence about one module. Local work is documentation only, with no executable interface.

Illustrative example: If transactions rise after changing training data, the model, and blending, the increase cannot all be assigned to the beam recipe simply because it also changed.

Sources: Recommender Systems with Generative Retrieval · PailitaoMMSearch2.0 Technical Report: Scaling Generative Retrieval to Unified End-to-End Multimodal Search

Can PinGER prefix diversity replace STATIC's legal mask or guarantee higher TIGER-style offline recall?¶

Show answer and example

Answer: No. LCP penalties spread slots across parent prefixes as a soft preference. STATIC determines which paths reach valid items. Wider subtree spread can lower recall for one held-out item, as PinGER's table shows. Removing SIDs from inputs, adding diversity, and ensuring legality are different interventions.

Illustrative example: Three semantic categories may each receive beam slots while still containing nonexistent catalog leaves. Removing illegal leaves also does not guarantee retention of the target.

Sources: PinGER: Token-Efficient Generative Retrieval with Semantic IDs · Recommender Systems with Generative Retrieval · Vectorizing the Trie: Efficient Constrained Decoding for LLM-based Generative Retrieval on Accelerators

Update log¶

Updates record new approaches, changed trade-offs, and follow-up evidence.

2026-10-30 · Reviewed PinGER's three masking rules, per-level LCP penalty, and prefix-wise anchored DPO. Compared historical SID cost with TIGER and legal transitions with STATIC, retaining the diversity/recall trade-off and limits of full-stack online evidence.

Affected: PinGER: Token-Efficient Generative Retrieval with Semantic IDs · Recommender Systems with Generative Retrieval · Vectorizing the Trie: Efficient Constrained Decoding for LLM-based Generative Retrieval on Accelerators

2026-10-29 · PailitaoMMSearch2.0 directly decodes multi-intent image/text requests into item SID sets, scaling serving with data tools, continual pretraining, mixed-budget SFT, and intent-balanced RL. Sample–Beam aligns inference sampling and beam candidate generation with training and display needs respectively. Larger models or longer reasoning do not guarantee improvement on every intent. Compare on fixed candidate sets, budgets, and production blending rules rather than model names. The full system is not reproduced locally.

Affected: PailitaoMMSearch2.0 Technical Report: Scaling Generative Retrieval to Unified End-to-End Multimodal Search

2026-10-27 · Initial version: compared TIGER, COBRA, PLUM, STATIC, and Reward Guided Decoding, with analysis and questions on item representations, learned distributions, and inference search.

Covers 60 topic entries; see the full list below.

All 62 topic entries

The full catalog follows. The main comparison covers the representative papers listed above.

Paper-by-paper questions · All deep dives