Many recommenders now describe each catalog item with a short sequence of discrete codes, called a semantic ID. The codes come from vector quantization, which rounds each item's continuous embedding to the nearest entry in a learned codebook. Because a new item's codes can be computed from its text description alone, this is attractive for the cold-start problem: items with no clicks can still be recommended. It also suggests a cheap form of explainability. If every code carried a human-readable name, the recommendation would be explained by the names of the codes it was generated from.
The catch is that the code is an approximation of the item. The distance between an item's embedding and the codebook entry it was rounded to is the quantization error. That error is precisely the part of the item the explanation does not describe. We propose to make this observation operational. Name each code in the vocabulary users write in their reviews, rather than in catalog category labels. Then treat each cold item's per-level quantization residual as a calibrated signal for which parts of its explanation can be stated, which should be hedged, and which should be withheld. Finally, train the quantizer so that residual error is smallest along the directions the names depend on.
Core Claim
In a generative recommender that scores items only through their code sequence, an explanation built from code names is faithful to the model by construction, since the model saw nothing else. Whether the explanation is true of the item is a separate question. Our claim is that the gap between "faithful to the model" and "true of the item" is governed by quantization error, and specifically by the residual measured along the directions that determine a code's name.
This yields three testable predictions. First, a cold item's per-level residual predicts whether that level's descriptor will later be confirmed by the item's own reviews. Second, naming codes from user review language rather than catalog taxonomy closes a vocabulary mismatch: catalog metadata and user phrasing for the same concept often share no words. Yang et al. found that 20.3% of correct answers share zero content words with the question in WikiQA, and catalog text versus reviews should show a similar gap. Third, a quantization penalty restricted to name-relevant directions raises explanation truthfulness on cold items without the ranking loss that full tag supervision incurs.
The closest prior work divides along a tension the file identifies directly. HiD-VAE supervises its codebook with LLM-extracted category tags, which makes the codes interpretable and cuts code collisions on Amazon Beauty from 21.2% to 2.1%. However, it ties the codes to frozen item-side taxonomies and evaluates only warm sequential settings. DIGER argues the opposite: codes optimized for ranking should drift away from content reconstruction. It prevents codebook collapse with annealed Gumbel noise, but its codes are opaque indices.
We do not force codes to equal names. Instead, we let the codebook train for ranking, attach names afterward from user language, and use the residual to decide how much of each name an item has earned. LAMIA supplies a useful structural choice: parallel aspect codes quantized by K-Means, which generalize to cold items. It produces no rationales and leaves its post-hoc clustering error unpenalized. Our method targets that residual explicitly.
Research Program
Base system. We build on a multi-aspect quantizer: either LAMIA-style parallel aspects or a residual quantizer with $L$ levels. Item embeddings come from a text encoder over metadata. A cold item $i$ with embedding $z_i$ is assigned code $k_\ell$ at aspect or level $\ell$, with residual
$$r_{i,\ell} = \lVert P_\ell (z_{i,\ell} - e_{\ell,k_\ell}) \rVert_2 .$$
Here $e_{\ell,k}$ is the $k$-th codebook entry at level $\ell$. $P_\ell$ is a projection onto the "naming subspace," the span of a linear probe trained to predict a code's name descriptors from embeddings.
Naming codes in user vocabulary. We borrow the mechanism from MINDER. MINDER indexes documents under pseudo-queries written to mirror how users search, which is what lets it reach Recall@100 of 78.7 on MS MARCO against 57.2 for SEAL. We apply the same idea to codes. For each code, we gather the reviews of the warm items assigned to it. We extract the descriptors that are over-represented in those reviews relative to sibling codes, and keep them only if they are absent from the code's majority catalog category, so the name adds user-side vocabulary. The code's name is the top few such descriptors, for example "lightweight, doesn't feel greasy" rather than "Skin Care > Moisturizer".
Residual-aware training. We add a penalty that pulls embeddings onto their codes only within the naming subspace:
$$\mathcal{L} = \mathcal{L}_{\text{rec}} + \lambda \sum_{\ell} \lVert P_\ell (z_{\ell} - \mathrm{sg}[e_{\ell,k_\ell}]) \rVert_2^2 .$$
$\mathcal{L}_{\text{rec}}$ is the generative recommendation loss, $\mathrm{sg}$ is stop-gradient, and $\lambda$ weights the term. This is the vertex-forcing idea from hashing (Deep Supervised Hashing with Triplet Labels, Unsupervised Triplet Hashing), restricted to a subspace. The rest of the embedding stays free to serve ranking, which is the freedom DIGER argues for. DIGER-style annealed noise is retained to prevent codebook collapse.
Gated explanation at inference. For each recommended cold item, we calibrate per level a threshold on $r_{i,\ell}$ against held-out truthfulness, using isotonic calibration on warm items that were once cold. Each level's name is then stated, hedged, or withheld. Withholding is the analogue of answer triggering in question answering, deciding that no valid answer exists, where WikiQA's best F1 was only about 31.6%. We expect abstention to be the hard part and plan to measure it directly.
Baselines
The do-nothing reference has two parts. For ranking, it is popularity fallback. GRECS reports that collaborative baselines such as NeuMF reach near 0% cold-item coverage on Amazon Beauty (Graph Reasoning for Explainable Cold Start Recommendation). For explanation, it is showing the cold item's catalog category. Beating this reference means higher cold-item ranking and higher truthfulness than the category string at equal explanation length.
The strongest interpretable-codes competitor is HiD-VAE. Its tag-aligned codes are directly comparable explanations, and it is the only system in the file that makes semantic IDs readable. We retrain it on our cold split with the same text encoder, codebook size, and hyperparameter search budget as our method, and use its taxonomy tags as explanations. Its reported collision rate of 2.1% on Beauty serves as a sanity check for our reproduction. Beating it means statistically higher truthfulness on cold items, especially in the low-lexical-overlap stratum, at no worse cold Recall@10.
The strongest accuracy competitor for opaque discrete codes is DIGER, reported at Recall@10 0.0683 and NDCG@10 0.0372 on Amazon Beauty versus 0.0610 and 0.0331 for a two-stage pipeline (DIGER). That result is on its own sequential split, not a strict cold split. We found no reported cold-split absolute result on Lacuna for DIGER or HiD-VAE on Beauty, so all methods will be re-run on one shared split. Our ranking target is parity with DIGER, within its run-to-run variance, rather than a gain.
The continuous cold-start reference is MARec. It aligns metadata and collaborative similarity in closed form and reports +8.4% to +53.8% cold-start gains over prior methods. We run it as an upper reference for cold ranking without quantization and pair it with nearest-liked-item explanations.
Three ablations isolate the key ingredients. The first replaces residual gating with random gating at matched abstention rates. The second replaces it with gating by code popularity. The third sets $P_\ell$ to the identity, which applies full vertex forcing.
Why not just…?
Why not just use HiD-VAE's category tags as the explanation? Tags are written in catalog vocabulary and are supervised into the codes, which is exactly the tension DIGER identifies: tag supervision constrains ranking. Our stratified evaluation tests whether tags fail specifically on items whose metadata shares little vocabulary with their later reviews. If tags match user-named codes in truthfulness across all strata, the vocabulary part of our claim fails. In that case we would report the residual gating as a standalone contribution layered on HiD-VAE.
Why not just prompt an LLM with the cold item's metadata to write the explanation? This is the real threat to the truthfulness metric. An LLM reading a description may produce descriptors that later reviews confirm. It is not, however, faithful to the recommender. Beyond Ranking Accuracy shows that LLM rationales need an explicit masking test before they can be trusted to reflect predictive signal, and ColdRAG reports 5% to 10% out-of-domain outputs for ungrounded LLM rankers. We therefore score both truthfulness and faithfulness. The prediction is that the LLM wins or ties on truthfulness alone but fails the code-deletion faithfulness test, while our method is competitive on both. If the LLM also passes faithfulness, which it cannot by design because it never sees the codes, the distinction becomes moot. We would report that honestly.
Why not just gate on cosine similarity between the item and its nearest warm neighbor? This is a cheaper confidence signal and needs no projection. It is included as a gating baseline. Our hypothesis is that the full-space residual or neighbor similarity is diluted by ranking-only dimensions, and that the projected residual $P_\ell$ predicts truthfulness better. If the unprojected signal calibrates equally well, we drop the projection and keep the simpler method.
Proposed Experiments
Datasets and split. We use Amazon Beauty and Amazon CDs and Vinyl, as used by GRECS and LAMIA, plus Yelp, as used by DIGER, because all three carry review text. We use a temporal split. Items first appearing after a cutoff are strictly cold at training time. The reviews they accumulate afterward are held out as ground truth written in users' own words. We stratify cold items by content-word overlap between their metadata and their future reviews, mirroring the WikiQA statistic.
Ranking. We report cold Recall@10, NDCG@10, and cold-item coverage, following GRECS, which reports 46% coverage on Beauty. Warm-item metrics are reported too, since cold-specialized models often degrade on warm items.
Truthfulness. For each stated descriptor, we check whether the item's future reviews support it. Support is judged by an off-the-shelf entailment model, with a human audit of a random sample to validate the judge.
Faithfulness. We adapt Cao et al.'s masking protocol to codes. We swap the named code for its nearest sibling and measure the change in rank. Explanations are counted as faithful when the swap moves the item.
Decisive checks. The first decisive check is whether the residual at level $\ell$ predicts descriptor truthfulness better than random, popularity, or neighbor-similarity signals. This is measured by AUROC and by risk-coverage curves for abstention. If it does not, the core claim is false. The second is whether the restricted penalty raises truthfulness on cold items in the low-overlap stratum while keeping Recall@10 within DIGER's variance. If truthfulness gains come only with full vertex forcing and a ranking loss, the DIGER–HiD-VAE trade-off stands unresolved, and we would say so.
References
- WikiQA: A Challenge Dataset for Open-Domain Question Answering
- HiD-VAE: Interpretable Generative Recommendation via Hierarchical and Disentangled Semantic IDs
- Differentiable Semantic ID for Generative Recommendation
- STORE: Streamlining Semantic Tokenization and Generative Recommendation with A Single LLM
- Multiview Identifiers Enhanced Generative Retrieval
- Deep Supervised Hashing with Triplet Labels
- Unsupervised Triplet Hashing for Fast Image Retrieval
- Graph Reasoning for Explainable Cold Start Recommendation
- MARec: Metadata Alignment for cold-start Recommendation
- Beyond Ranking Accuracy: Evaluating LLM-Cited Feature Rationales for Next Basket Repurchase Recommendation
- Cold-Start Recommendation with Knowledge-Guided Retrieval-Augmented Generation