Semantic IDs for Joint Generative Search and Recommendation

Alexandre Tamborrino; Ali Vardasbi; Edoardo D'Amico; Enrico Palumbo; Francesco Fabbri; Gustavo Penha; Hugues Bouchard; Marco De Nadai; Max Lefarov; Shawn Lin

REVIEW 2 major objections 3 minor 1 cited by

Semantic IDs for Joint Generative Search and Recommendation

T0 review · 2 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A single Semantic ID space built from a jointly fine-tuned bi-encoder delivers strong performance in both search and recommendation.

desk verdict A relevant empirical design paper whose central claim is unverifiable from the abstract alone; worth a serious look if the full evaluation holds up. read the letter →

arxiv 2508.10478 v1 pith:QUIG7IBN submitted 2025-08-14 cs.IR

Gustavo Penha , Edoardo D'Amico , Marco De Nadai , Enrico Palumbo , Alexandre Tamborrino , Ali Vardasbi , Max Lefarov , Shawn Lin

show 3 more authors

Timothy Heath Francesco Fabbri Hugues Bouchard

This is my paper · ORCID

classification cs.IR

keywords SemanticIDsgenerativesearchandrecommendationbi-encoderembeddingsdiscreteitemcodesunifiedrepresentationjointfine-tuningLLM-basedrecommendersystems

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Generative models that handle both search and recommendation need a way to represent items. This paper investigates how to build Semantic IDs — discrete codes derived from item embeddings — so that a single unified model performs well at both tasks. The central claim is that a bi-encoder fine-tuned jointly on search and recommendation, followed by a unified Semantic ID space, offers an effective trade-off. If true, this resolves a key design choice for the next generation of generative recommender systems.

What carries the argument

The central mechanism is the Semantic ID: a discrete code sequence obtained from item embeddings. The paper's key choice is the bi-encoder, a two-tower embedding model fine-tuned on both search and recommendation losses, whose output embeddings are quantized into a unified Semantic ID vocabulary. This shared vocabulary is what lets a single generative decoder condition on the same items for both tasks.

What would settle it

A direct experiment on a public joint dataset with both relevance judgments and user interaction data, where a task-specific Semantic ID scheme beats the unified scheme on both search and recommendation metrics under identical conditions, would falsify the central claim.

Watch

Extended reading notes

Core claim

The paper proposes and compares several strategies for constructing Semantic IDs: task-specific embedding models versus cross-task approaches, and separate versus shared Semantic ID tokens per task. It finds that taking a bi-encoder fine-tuned on both tasks to produce item embeddings, then building one unified Semantic ID space, gives strong performance in search and recommendation simultaneously. This is the discovery: the same discrete item representation can serve both tasks if the embedding backbone is jointly trained, rather than specialized.

Load-bearing premise

The claim that a jointly fine-tuned bi-encoder plus unified Semantic ID space is the effective trade-off rests on the chosen evaluation benchmarks fairly representing both search and recommendation and on the comparison being made against the strongest alternative Semantic ID strategies.

Editorial extensions

If this is right

A unified generative search-and-recommendation model can rely on a single Semantic ID vocabulary, simplifying the item representation layer.
Jointly fine-tuning the embedding backbone on both tasks is preferable to using task-specific embeddings for a joint model.
The comparison of task-specific versus cross-task ID construction provides a design principle: a shared ID space beats separate per-task tokens in this setting.
Strong performance on both tasks suggests that semantic grounding of IDs helps generalization across tasks.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same trade-off logic may transfer to other pairs of tasks sharing an item catalog, such as ranking and filtering, or browsing and open-ended question answering.
A testable implication is that a jointly fine-tuned bi-encoder should also outperform a task-specific embedding backbone on a held-out third task, since the unified ID space may encode more transferable structure.
The finding points toward generative systems whose item IDs are not arbitrary but derived from semantics, which could improve cold-start behavior for new items.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You asked for a take on arXiv:2508.10478. Here it is, with the caveat that only the abstract was available, so this is a judgment of the research question and the framing, not the evidence.

The paper tackles a real, live design question: how should you build Semantic IDs for a unified generative search-and-recommendation model? The abstract compares task-specific versus cross-task embedding strategies and shared versus per-task ID token spaces. That is a genuinely useful comparison, and the claim — jointly fine-tuning a bi-encoder and then building a unified Semantic ID space gives a good trade-off — is plausible and well scoped. If the full paper delivers a clean systematic comparison, it would give practitioners concrete guidance instead of the usual hand-waving.

That said, the abstract contains no datasets, no metrics, no baselines, no error bars. The phrase “effective trade-off” is doing all the work. There is no way to tell whether the joint-fine-tuned bi-encoder wins because it is genuinely better or because the baselines are weak, or because the evaluation benchmarks happen to favor one configuration. The mild circularity — the embedding backbone is trained on the same two tasks used for evaluation — is standard practice and not by itself a flaw, but it means the paper needs to show evaluation on held-out data or at least be careful about overfitting to the target objectives.

I can't flag any internal contradiction or in-principle flaw, because there's no substance to inspect. The stress-test note's concern about unnamed benchmarks is precisely the right one. The paper's value depends entirely on the rigor of the experiments, and that's hidden.

For whom is this? Anyone designing unified generative retrieval or recommendation systems, especially those considering Semantic IDs. It is not a theory paper and it won't change how we think about the problem, but it could settle a practical design question.

My recommendation: send it to peer review if the full paper has a real evaluation with named datasets, strong baselines, and sensible held-out splits. A serious referee should check whether the comparison set includes the strongest prior Semantic ID methods and whether the joint fine-tuning advantage survives across benchmarks. As it stands, the abstract alone doesn't earn a citation from me, but the work deserves a chance.

If the full paper delivers what the abstract promises, this is a useful contribution to an active area. If the experiments are narrow, it's a one-setting observation. That's exactly what a referee should decide.

Referee Report

2 major / 3 minor

Summary. The paper argues that in a unified generative model for search and recommendation, item representations should be constructed as Semantic IDs derived from embeddings, and that the best strategy is to fine-tune a bi-encoder jointly on both tasks before building a unified Semantic ID space. The abstract reports that this approach offers an 'effective trade-off' and strong performance in both tasks, but it provides no datasets, metrics, baselines, or numerical results. The full text was not available to the referee, so this assessment is limited to the abstract and the claims stated there.

Significance. If the central claim is substantiated, the paper would provide a practical design principle for unified generative search-and-recommendation systems: use a single jointly fine-tuned embedding backbone and a shared Semantic ID vocabulary. The comparison of task-specific and cross-task Semantic ID construction strategies is a useful contribution. However, because the abstract contains no experimental evidence, the significance is entirely conditional on an evaluation that is not visible to the reader. The claimed trade-off cannot currently be assessed for magnitude, statistical reliability, or generality.

major comments (2)

[Abstract] The central empirical claim—that a joint fine-tuned bi-encoder plus unified Semantic ID space provides an 'effective trade-off'—is stated without any supporting evidence. No datasets, metrics, baselines, or numerical results are given. As an empirical claim, its validity cannot be inspected from the abstract, and the absence is load-bearing: the paper's contribution is precisely this experimental finding. The full text may contain the necessary details, but they are not present in the manuscript as reviewed.
[Abstract] The evaluation protocol is unspecified with respect to the fine-tuning/evaluation overlap. The bi-encoder is fine-tuned on both search and recommendation tasks, and the final model is presumably scored on those same task families. The abstract does not state whether evaluation is performed on held-out data, held-out tasks, or tasks not directly optimized for. If the tasks overlap with the training objectives, the 'effective trade-off' may partly reflect fitting to the evaluation targets rather than generalization. This concern needs to be addressed with a clear description of the experimental split.

minor comments (3)

[Abstract] The term 'Semantic ID' is used without definition or a pointer to prior work, which may confuse readers outside the immediate area.
[Abstract] The phrase 'effective trade-off' is vague; naming at least one quantitative comparison (e.g., relative performance drop versus a single-task baseline) would strengthen the abstract.
[Abstract] The abstract would benefit from naming the benchmark datasets and the main baseline(s) to give readers a concrete sense of the scope of the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity identifiable from abstract-only text; no equations, derivations, or self-citations to audit.

full rationale

The paper is available only as an abstract; no equations, derivations, or cited prior results are present. The central claim is an empirical finding about Semantic ID construction strategies, not a derivation from first principles. The bi-encoder is fine-tuned on search and recommendation and later evaluated on those tasks, which is a standard supervised learning setup rather than a circular reduction: the abstract gives no indication that the reported 'strong performance' is measured on the same training instances used to fit the embeddings, and the comparison of strategies is conducted under a common protocol. Without a specific equation or self-citation that forces the conclusion to equal its inputs, no circular step can be exhibited under the stated hard rules. The absence of dataset and metric details is an evidence limitation, not circularity. Therefore the circularity score is 0.

Assumptions & free parameters 3 free parameters · 2 assumptions · 0 invented entities

The central claim rests on two unverified premises that the abstract itself does not establish: that the embedding space from a jointly fine-tuned bi-encoder survives discrete quantization into Semantic IDs without losing the information needed for both tasks, and that the (unnamed) evaluation protocol fairly measures joint search and recommendation quality. Three classes of fitted quantities are load-bearing: the bi-encoder weights, the Semantic ID codebook, and the generative model weights; none are specified in the abstract.

free parameters (3)

Bi-encoder item embedding parameters = not reported (abstract only)
Item embeddings come from a bi-encoder fine-tuned on both search and recommendation data; every compared Semantic ID strategy depends on this fit.
Semantic ID construction hyperparameters (codebook size, codes per item, quantization levels) = not reported (abstract only)
The discrete codes are obtained from embeddings via an unstated quantization procedure; its capacity controls how much semantics survive into the IDs.
Unified generative model weights = not reported (abstract only)
The joint search-and-recommendation generative model is trained on the Semantic IDs; its parameters are fitted to the same two tasks.

assumptions (2)

domain assumption Quantization of jointly fine-tuned item embeddings into discrete Semantic IDs preserves enough information for both search relevance and recommendation compatibility
The entire strategy depends on the embedding space being a lossy-but-sufficient substrate for both tasks; the abstract presents this as the finding to test, not as an established fact.
domain assumption The evaluation protocol (unnamed datasets, metrics, and baselines) is a fair, representative measurement of joint search and recommendation performance
The 'effective trade-off' is comparative and empirical; without named benchmarks, this premise is unverified.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Semantic IDs for Joint Generative Search and Recommendation." pith.science (2026). https://pith.science/paper/QUIG7IBN

@misc{pith2026250810478,
  author       = {Pith},
  title        = {Pith review of: Semantic IDs for Joint Generative Search and Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QUIG7IBN}},
  note         = {Machine review of arXiv:2508.10478}
}

read the original abstract

Generative models powered by Large Language Models (LLMs) are emerging as a unified solution for powering both recommendation and search tasks. A key design choice in these models is how to represent items, traditionally through unique identifiers (IDs) and more recently with Semantic IDs composed of discrete codes, obtained from embeddings. While task-specific embedding models can improve performance for individual tasks, they may not generalize well in a joint setting. In this paper, we explore how to construct Semantic IDs that perform well both in search and recommendation when using a unified model. We compare a range of strategies to construct Semantic IDs, looking into task-specific and cross-tasks approaches, and also whether each task should have its own semantic ID tokens in a joint search and recommendation generative model. Our results show that using a bi-encoder model fine-tuned on both search and recommendation tasks to obtain item embeddings, followed by the construction of a unified Semantic ID space provides an effective trade-off, enabling strong performance in both tasks. We hope these findings spark follow-up work on generalisable, semantically grounded ID schemes and inform the next wave of unified generative recommender architectures.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

HCGRec: Hint-Conditioned Generative Recommendation with Semantic IDs
cs.IR 2026-08 conditional novelty 6.0 of 10

HCGRec supplies the shortest target-prefix hint for hard training instances during GRPO post-training, then optimizes only the unhinted suffix, reducing zero-advantage rollout groups and improving sequential recommend...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.