Recommenders trained on one shopping category usually cannot help another, because they identify items by arbitrary ID numbers that mean nothing outside their own catalog. A recent fix is to give every item a short "semantic ID": a sequence of discrete codes computed from its text description, so that items from different catalogs share one code vocabulary and one sequence model can be trained across all of them. This fix has a weakness. Different domains describe their goods in different words. Lab supplies, pantry food and art materials have little vocabulary in common, so a tokenizer that groups items by their text tends to place each domain's items in its own corner of code space. The vocabulary looks shared but is not used in common, and behavioral knowledge has no common symbols to travel through. This is the old lexical gap from search, where a query and a relevant document share no words, now showing up inside the item tokenizer.
We propose to measure this effect directly and then repair it. The repair lets the recommendation loss reshape the coarse, shared level of the item codes across domains. As a result, items that play the same behavioral role, such as a consumable that is re-bought or an add-on bought after a main purchase, can receive the same coarse code even when their descriptions share no words. Fine code levels stay private to each domain. The claim to test is that cross-domain gain in generative recommendation depends on how many codes are actually used by more than one domain, and that behavior-shaped codes keep this sharing even when the domains' vocabularies diverge.
Core Claim and How It Differs From Prior Work
Two papers supply the method, and their gap motivates the project. Generative Multi-Target Cross-Domain Recommendation (GMC) works with domains that share no users and no items. It quantizes LLM text embeddings with a residual quantizer (a tokenizer that assigns each item a coarse code, then progressively finer codes for what remains), trains one sequence-to-sequence model on merged domains, and adds small per-domain adapters. Its tokenizer is frozen before recommendation training. It also uses a domain-aware contrastive loss that clusters items within a domain, which, if anything, pushes domains apart in code space. Differentiable Semantic ID for Generative Recommendation (DIGER) shows that freezing the tokenizer creates an objective mismatch. Its remedy is to train the codes jointly with the recommender, using Gumbel-noise exploration with uncertainty decay to avoid codebook collapse, which is the failure where most items end up on a few codes. DIGER was tested only within single domains.
Our proposal is not simply DIGER run on merged domains. It differs in three ways. First, we introduce a diagnostic: per-level cross-domain code sharing, the fraction of each target domain's items whose code at a given level is also held by source-domain items. Second, we split gradient flow by level. Collaborative gradients from all domains reshape the coarse shared level, while fine levels are learned per domain. Third, we add a domain-mixing term that penalizes coarse codes whose usage comes from a single domain. The exploration noise is decayed per domain, so the busiest domain cannot monopolize the hot codes.
RecBase builds a unified concept-ID vocabulary across 15 domains, but its tokenizer is also static and it requires pretraining from scratch on 35 million interactions. We stay at GMC's scale.
Cross-Domain Policy Adaptation by Capturing Representation Mismatch constrains the design. It argues that forcing invariance across very different domains imports misleading source patterns. Adversarial invariance as in SER and shared central weights as in STAR take the opposite stance. We take PAR's warning seriously in two ways. Sharing is encouraged only at the coarse level, and the strength of the mixing term is swept to find the point where sharing starts to hurt.
Research Program
The model follows GMC's pipeline. Item text from each Amazon domain is embedded with a frozen LLM encoder and quantized with three residual levels. A T5-style model predicts the next item's code sequence over merged user histories, and per-domain low-rank adapters are added on top. The changes are confined to the tokenizer and its training signal.
The first level's codebook is shared across domains and receives DIGER-style soft gradients from the recommendation loss summed over all domains. Each domain gets its own exploration-noise schedule: the noise scale shrinks as that domain's own loss falls, and noise applies only to codes that are overused within that domain. The mixing term is the negative entropy of the domain distribution among items assigned to each coarse code. It is weighted by $\lambda$ and is applied only to codes used by at least a minimum number of items, so rare codes are not forced to mix. Levels two and three use per-domain codebooks with the same exploration schedule. This keeps the fine distinctions within a domain while letting the coarse level carry behavioral roles across domains.
The analysis has two parts. The first is the sharing diagnostic, computed for frozen GMC codes, for naive merged-DIGER codes, and for our codes. The second is a controlled lexical-shift study. We rewrite one source domain's item text into a deliberately different register, either jargon-heavy or colloquial, using an LLM. Interaction sequences are kept fixed, so behavior is unchanged and only the vocabulary moves. If our account is right, the rewrite should lower cross-domain code sharing and target accuracy for frozen text tokenizers. Behavior-shaped coarse codes should be much less affected.
Baselines
The do-nothing reference is a single-domain sequential recommender trained on each target domain alone, using both an ID-based model in the SASRec style and a frozen-tokenizer TIGER-style generative model. Both appear as baselines in GMC's evaluation on these Amazon domains. Any cross-domain method has to beat these, or transfer is not paying for itself.
The strongest existing method is GMC itself, re-run on its own five Amazon domains (Scientific, Pantry, Instruments, Arts, Office) with its released settings. GMC reports a 25.41% NDCG@5 improvement over its strongest baseline on Pantry. The file gives no absolute GMC values for the other domains, so we will report our reproduced absolute numbers next to that relative figure. To count as beating GMC, our method must improve NDCG@5 and Recall@10 on most targets, and especially on the smallest domains, with paired significance over five seeds. For reference on the single-domain gain from differentiable codes, DIGER reports Recall@10 of 0.0683 against 0.0610 for the two-stage pipeline on Amazon Beauty. A cross-domain gain should be of at least comparable relative size to be worth the added machinery.
The key ablation is naive merged-DIGER. It uses one fully shared differentiable codebook at every level, a single global noise schedule, and no mixing term. This isolates whether the gain comes from per-level splitting, per-domain decay and mixing, or merely from making the codes differentiable. Two further ablations remove the mixing term (setting $\lambda=0$) and freeze the coarse level while leaving fine levels differentiable.
Every system will use the same text encoder, the same code length and codebook sizes, the same backbone and adapter rank, an equal number of training steps, and the same hyperparameter search budget on validation splits.
Why not just…?
Why not just run DIGER on the merged domains? This is the most direct threat, and it is our main ablation. Our worry is that a single shared noise schedule lets the highest-volume domain drive which codes count as hot. The resulting codes would then track that domain's behavior, which is the noise-pollution failure noted for multi-scenario models. The sharing diagnostic and the per-domain accuracy breakdown will show whether this happens. If merged-DIGER matches our method, we will report that the gain comes from differentiability alone and drop the extra components.
Why not just normalize item text into a common vocabulary before tokenizing? One could have an LLM rewrite every item description into domain-neutral attribute language and then run frozen GMC. This is cheap and could close much of the lexical gap. It is also a real threat, so we include it as a baseline with equal LLM budget. There are two reasons to expect it to fall short. ECLASS-Augmented Semantic Product Search found that LLM rewriting stripped details that mattered and hurt retrieval, and text normalization still cannot tell the tokenizer which items play the same behavioral role. The lexical-shift study settles the question: if rewritten-then-frozen codes recover sharing and accuracy as well as behavior-shaped codes, the simpler fix wins.
Why not just add category paths as structured metadata? ECLASS-style taxonomy augmentation raised Hit_Rate@5 on expert queries from 31.4% for BM25 to 94.3% in ECLASS-Augmented Semantic Product Search. Amazon category paths are available for free, so we test a GMC variant whose input text is prefixed with the category path. Category paths differ across domains by construction, though, so they cannot create cross-domain sharing on their own. We expect them to help within domains and leave the sharing diagnostic unchanged.
Why not keep separate codebooks per domain with adapters? This is the domain-isolation extreme. We run it to locate where sharing helps and where, as PAR warns, it hurts.
Decisive Checks
The project succeeds if three things hold. First, across target domains and methods, the gain from cross-domain training correlates positively with coarse-level code sharing. Second, under the controlled vocabulary rewrite, frozen tokenizers lose both sharing and accuracy, while behavior-shaped coarse codes keep most of both. Third, the full method beats both GMC and merged-DIGER on the smallest target domains without losing accuracy on the largest.
The project fails, and the hypothesis is rejected, if sharing does not predict gain, or if text normalization matches our method under the rewrite. A sweep over $\lambda$ that shows accuracy falling at high mixing would map the boundary between helpful sharing and the forced invariance PAR warns against. That boundary is a result in its own right.
References
- Generative Multi-Target Cross-Domain Recommendation
- Differentiable Semantic ID for Generative Recommendation
- RecBase: Generative Foundation Model Pretraining for Zero-Shot Recommendation
- Cross-Domain Policy Adaptation by Capturing Representation Mismatch
- Review-Based Domain Disentanglement without Duplicate Users or Contexts for Cross-Domain Recommendation
- One Model to Serve All: Star Topology Adaptive Recommender for Multi-Domain CTR Prediction
- ECLASS-Augmented Semantic Product Search for Electronic Components