EPIC: Explicit Posterior Item Conditioning for Semantic ID Diffusion Recommendation

• An SID is a compact tuple of discrete codes for an item, with related items able to share codes.Source: §1 · p. 1
• Unlike autoregressive SID generation [15], masked diffusion uses bidirectional context and flexible position resolution [10] [12] [16] [17], but still predicts tokens rather than complete items.Source: §1 · pp. 1–2
• Token-level denoising leaves competition among feasible complete items implicit, so a locally plausible commitment can eliminate a promising item before reranking.Source: §1 · p. 2
• EPIC recomputes a personalized posterior over feasible items from backbone context and recent interactions, then projects it to unresolved SID positions.Source: §1 · p. 2
• Candidate-conditioned transition evidence and frontier-aware learning support hypotheses using recent interactions and focus item supervision where candidates compete.Source: §1 · p. 2
• EPIC consistently improves full-catalog next-item recommendation over strong baselines on four Amazon benchmarks.Source: §1 · p. 2

• Semantic-ID recommenders generate item identifiers, with TIGER assigning items semantic identifiers through residual quantization and autoregressively generating them [15].Source: §2 · p. 2
• Masked SID methods recover tokens using bidirectional dependencies or parallel and adaptive decoding [17] [10] [12] [16], but retain token-level supervision and prediction.Source: §2 · p. 2
• Item-level alternatives score items directly [30], rerank generated candidates [21], or guide autoregressive prefixes with a pre-generation ranking [23]; EPIC updates evidence for arbitrary masked-diffusion states before the next commitment.Source: §2 · p. 3

3 Preliminaries and Problem Formulation

• Sequential recommendation ranks the next catalog item from a user’s chronologically ordered interaction history.Source: §3 · p. 3
• Each item is represented by a tuple of discrete codes; different items may share a tuple, and a generated tuple identifies one item only when its catalog preimage is a singleton.Source: §3 · p. 3

3.1 Masked Semantic ID Generation

• Masked SID diffusion trains on target tuples with sampled positions replaced by MASK, then iteratively resolves positions from a fully masked state.Source: §3.1 · p. 3
• At inference, the frozen token-level backbone uses its predictions to select codes and positions for resolution.Source: §3.1 · p. 3

3.2 Partial States as Feasible Candidate Sets

• A partial state induces a feasible candidate set of catalog items whose SIDs match its resolved positions; masked positions need not form a prefix.Source: §3.2 · p. 3
• Feasibility means SID consistency, not personalized relevance; multiple items can remain indistinguishable, including after all codes are resolved when identifiers collide.Source: §3.2 · p. 3

3.3 The Token–Item Inference Gap

• Position-wise token distributions and feasible-set membership do not provide a normalized, personalized ranking over complete candidates, motivating an actionable item posterior during denoising.Source: §3.3 · p. 3

4.1 Framework Overview

• At each denoising state, EPIC scores up to M feasible candidates using complete-item and history representations, then projects the normalized posterior to SID codes before the next transfer decision.Source: §4.1 · p. 3
• The design is additive: new item-level components enter through a residual path while reusing the frozen backbone’s existing outputs.Source: §4.1 · pp. 3–4
• Personalized transition evidence is fused with backbone token logits through an ambiguity-aware gate to produce the refined token posterior and next state.Source: Figure 2 · p. 4

4.2.1 Feasible candidates and retained support.

• Token likelihood proposes and prunes feasible candidates, while the learned posterior ranks retained hypotheses without counting token evidence twice.When the feasible set has at most M items, the retained support covers the full set.Source: §4.2.1 · p. 4

4.2.2 Structured item and history representations.

• A shared encoder represents candidates and recent items from complete SIDs, while a lightweight GRU summarizes the retained history window.The encoder reuses frozen code embeddings and adds no item vocabulary; items with the same complete SID remain tied.Source: §4.2.2 · p. 4

4.2.3 Personalized transition field and item posterior.

• Candidate-conditioned transition evidence lets different feasible items draw support from different parts of recent history, and the normalized posterior represents competition across the retained items.The posterior combines direct context relevance with transition evidence; token likelihood forms the candidate support but is excluded from posterior energy.Source: §4.2.3 · p. 5

4.2.4 Frontier-aware learning.

• Direct item supervision applies on the identification frontier, where 2 to M feasible candidates compete, and structured smoothing is strongest under heavy masking before annealing toward exact-item discrimination.The frontier excludes singleton states and ensures supervision covers every feasible candidate; only the item-level adapter is updated.Source: §4.2.4 · p. 5

4.3.1 Item-to-token marginalization.

• The complete-item posterior is marginalized onto each unresolved SID position, preserving joint item competition and assigning mass only to codes used by retained items.The feedback is the exact positional marginal of the item posterior.Source: §4.3.1 · p. 5
• A position-specific signed gate conditions item feedback on item- and token-level uncertainty, allowing it to promote or suppress evidence as denoising proceeds.The gate's zero-initialized final layer makes feedback an exact initial no-op.Source: §4.3.1 · p. 5
• EPIC recomputes the candidate set, retained support, and posterior at each step, so item evidence changes the denoising trajectory rather than reranking completed outputs or carrying a recurrent posterior.The computations reuse current backbone states and logits, with no additional backbone forward pass.Source: §4.3.1 · pp. 5–6

5.1 Datasets and Evaluation Protocol

• Four 5-core Amazon Review categories are evaluated with each user's final interaction held out for testing, the penultimate for validation, and earlier interactions for training.The task is full-catalog next-item recommendation.Source: §5.1 · p. 6

5.2 Baselines and Metrics

• Evaluation compares full-catalog Recall@K and NDCG@K under a unique-SID item-level protocol, with ambiguous SIDs receiving no credit.The baselines span item-ID sequential recommenders, SID generative models, and masked-diffusion recommenders; adjusted paired comparisons are significant across all 16 dataset–metric combinations.Source: §5.2 · p. 6

5.3 Training and Inference Details

• EPIC adopts LLaDA-Rec's bidirectional Transformer and four-step transfer schedule, freezes the backbone, and optimizes only the item-reasoning adapter.Matched interventions share the tokenizer, decoder, beam width, candidate budget, and completion rule; hyperparameters are selected using validation NDCG@5.Source: §5.3 · p. 6

6.1 Overall Performance

• EPIC achieves the best mean in all 16 metric–dataset combinations, with gains of 1.2%–16.8% over the second-best result in each column.Every gain remains statistically significant after Holm correction across all 16 comparisons (p_adj < 0.05).Source: §6.1 · pp. 6–7, Table 2 · p. 7

6.2 Does Item Evidence Preserve Hypotheses During Denoising?

• Inline inference improves NDCG@5 over matched post-hoc inference by 15.7%, 26.6%, and 26.8% on Beauty, Sports, and Toys, respectively.Delaying feedback until step 2 reduces NDCG@5 by 4.3%–9.3%; restricting evidence to conditional completion reduces it by 8.6%–14.9%.Source: §6.2 · pp. 7–8
• Inline inference yields 315/371/360 net trajectory hits on Beauty/Sports/Toys and recovers 261/337/366 top-5 hits while losing 107/106/113.Among recovered cases, 25/48/49 targets are absent from every final control beam and the control top-128 pool, so post-hoc reranking cannot recover them.Source: §6.2 · p. 8

6.3 What Drives the Improvement?

• Removing transition memory causes a 15.13% NDCG@5 drop; replacing users’ histories reduces NDCG@5 by 12.4%–17.9%, while reversing history order reduces it by 1.5%–4.1%.A generic item posterior without transition memory is neutral relative to the decode-matched control.Source: §6.3 · p. 8
• On Beauty, token-coupled posterior constructions trail the independently learned scorer by 3.7% and 2.3% validation NDCG@5; learned history weighting improves over uniform aggregation by 2.2%–6.4%.The compared constructions are a product-of-experts prior and leave-one-position-out cavity messages.Source: §6.3 · p. 8

6.4 When Is EPIC Effective, and at What Cost?

• Resolving two SID positions gives the best Beauty NDCG@5 (.04385) and is 1.72× faster than resolving all four; EPIC improves every populated feasible-set stratum, though ambiguity alone does not ensure a large gain.The largest Toys point contains only 34 users and is descriptive; positive gain at |C| = 1 reflects benefits accumulated at earlier states.Source: §6.4 · p. 9
• The adapter trains 0.391M parameters, 94.3% fewer than full LLaDA-Rec fine-tuning, while inference latency rises 34.2%, from 93.0 to 124.8 ms/user.The training comparison excludes pretraining the frozen backbone, shared by both systems; resolving two SID positions provides a lower-latency operating point than the complete trajectory.Source: §6.4 · p. 9, Table 3 · p. 9

6.5 Case Study: A Hard Rescue

• In the audited Beauty hard rescue, inline feedback keeps the target available through denoising and finishes at rank 3, while control prunes it from every final beam.The inline ranks are 4, 6, and 7 before rank 3, yielding NDCG@5 of .50; control ranks the target 27, 175, and 77 before pruning it.Source: §6.5 · p. 9, Figure 7 · p. 9

7 Conclusion

• EPIC compares feasible candidates using user-specific item evidence and projects their posterior back into token generation while candidates remain reachable.Source: §7 · pp. 9–10
• EPIC improves all 16 dataset–metric combinations across four Amazon benchmarks.Matched inline–post-hoc comparisons favor applying item evidence during denoising over using the same scorer only after generation.Source: §7 · p. 10
• Analyses attribute the gains primarily to user-specific transition evidence guiding intermediate decoding decisions rather than generic candidate scoring.Future work could use approximate candidate construction, adapt candidate budgets across denoising steps, and distinguish catalog items sharing the same SID.Source: §7 · p. 10

Ethical Considerations

• The study uses publicly available, de-identified Amazon review benchmarks and involves neither new data collection nor human-subject intervention.Source: Ethical Considerations · p. 10
• Historical interactions and item metadata may encode biases, which repeated recommendations and feedback loops could amplify to disadvantage long-tail items or underrepresented users.Source: Ethical Considerations · p. 10
• Deployment evaluation should extend beyond ranking accuracy to subgroup performance, catalog exposure, diversity, and robustness to sparse histories.The authors also call for consent, data minimization and security, ongoing bias audits, user controls, and mechanisms to correct harmful recommendations.Source: Ethical Considerations · p. 10