acceptodds
Under review as a conference paper at ICLR 2027

ELGAR: Exposure-Ledger Guided Adaptive Routing for Fair Generative Recommendation

Abstract

Generative recommendation has recently attracted growing attention, with Semantic-ID((SID)-based generative recommenders emerging as a major paradigm. They formulate recommendation as autoregressive item generation, where each item is represented by a sequence of discrete semantic tokens. Despite strong preference modeling, these models tend to over-expose popular items, leading to severe item-side unfairness. Existing methods face two limitations: misalignment, as improvements in tokenization or generation do not directly ensure alignment with the target item-exposure distribution; and rigidity, as offline corrections do not adapt to evolving exposure deficits across requests. We propose ELGAR, an Exposure-Ledger Guided Adaptive Routing framework for fair generative recommendation. ELGAR combines explicit group routing before semantic pruning with an Exposure Ledger that tracks cumulative exposure across requests. Together, they enable adaptive group-level allocation while preserving user relevance across different SID-based backbones. Experiments on four datasets and three backbones demonstrate improved fairness while maintaining recommendation performance. Our code is available at https://anonymous.4open.science/r/ELGAR-Anon-FB51.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Generative Recommendation, Popularity Bias, Large Language Models.

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_FBdxOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ELGAR addresses popularity-calibrated item-side fairness in Semantic-ID (SID) based generative recommendation: exposure across fixed popularity groups is calibrated to their shares of raw training interactions. The method prepends a popularity-group token to each existing SID and fuses the backbone's group logits with a ledger-conditioned fairness head at the first decoding step. The Exposure Ledger records normalized group-exposure deviations and a clipped exposure-progress indicator. Training constructs label-driven exposure episodes from shuffled mini-batches and combines the generative objective, target-group supervision, and an expected one-step fairness-gain objective. During inference, a persistent ledger is updated from the displayed Top-K items after each request block, while each beam continuation remains within its selected group. Experiments on Beauty, Toys, Sports, and KuaiRec*, using TIGER, LETTER, and MiniOneRec, report lower exposure mismatch and higher Tail NDCG@10 than the corresponding vanilla backbones.

Critical strengths
  1. The intervention point is well motivated for SID-based generation. Placing group allocation before semantic-prefix pruning gives the method access to groups whose items might otherwise disappear before post-hoc reranking. ELGAR retains the existing SID mappings and combines request-dependent relevance with accumulated exposure information at this control point. This is a useful design for bringing exposure allocation into generative retrieval, rather than correcting only the candidates that survive decoding. Target-group supervision also provides a concrete mechanism for balancing the fairness signal with user relevance.

  2. The empirical evaluation covers four datasets and three backbones, including both encoder-decoder and decoder-only models. The reported results demonstrate a meaningful utility-calibration trade-off: on Beauty with TIGER, MGU@10 decreases from 0.1827 to 0.0703 and Tail NDCG@10 increases from 0.0018 to 0.0054, while overall NDCG@10 changes from 0.0303 to 0.0298. The fusion-loss ablation and the random-routing comparison further illustrate why exposure matching alone is insufficient. The granularity and alternative Head-Tail boundary analyses also provide useful evidence about the dependence on the chosen fairness configuration.

Critical weaknesses
  1. The experiments do not isolate the contribution of adaptive exposure feedback, which is central to the paper. The w/o Stateful Decoding ablation in Table 2 sets lambda_dec = 0 and removes the entire fairness-head contribution, including its request-dependent component, rather than only removing ledger feedback. The random-routing baseline in Appendix I is useful, but it discards request-dependent group relevance as well as learned feedback; outperforming it in NDCG does not establish the benefit of the proposed ledger mechanism over a relevance-aware exposure controller. A comparison that retains the relevance and fairness heads while freezing or removing only the ledger input, together with a simple relevance-plus-exposure-deficit controller at the same group-token step, would better separate the effects of group-token generation, additional supervision, and adaptive feedback. This matters because the current evidence supports the complete system more clearly than its distinctive feedback mechanism.
  2. The gap between training and inference needs further validation. Training resets exposure counts for each mini-batch and updates them using ground-truth items, while inference accumulates exposure from generated Top-K lists. It is unclear whether the learned feedback remains reliable over longer recommendation sequences or under different request orders.
Questions for authors
  1. To assess the contribution of adaptive feedback, can you provide a control that retains the group-token generator, the fairness head, and relevance supervision while freezing or removing only the ledger input? A comparison with a relevance-aware deficit controller applied at the same routing step would also be informative. Please use comparable validation budgets and report the utility-MGU trade-off, since the random-routing baseline already achieves lower MGU at lower utility.

  2. Could you clarify the intended state-sufficiency claim in Appendix J.2? With pi = (0.8, 0.2) and H = 20, counts n = (17, 4) and n' = (170, 40) produce the same ledger, (1/105, -1/105, 1), but Eq. (8) gives the larger fairness gain to Head in the first case and Tail in the second. Is the clipped ledger intended as an approximate state representation, and does retaining information about the exposure total materially change the results?

  3. What request ordering and ledger-reset protocol are used for validation and test? To assess the training-to-serving transfer, can you report prefix-wise exposure mismatch and utility under multiple permutations of the same held-out requests, with variability across runs?

10 comments
AuthorsOpenReview ↗

Response to Reviewer FBdx - Part 1/6 (W1 & Q1)

Thank you for recognizing ELGAR’s strengths and for your constructive suggestions; we added controlled experiments, theoretical analysis, and multi-permutation evaluations to address all points below.

W1 & Q1: Isolate adaptive exposure feedback using frozen-ledger controls (retaining the fairness head) and a relevance-aware deficit controller under comparable validation budgets.

R1:

We appreciate the reviewer’s suggestion to isolate the contribution of adaptive exposure feedback through fixed-ledger and relevance-aware controls. To separate adaptive exposure feedback from static group bias and rule-based deficit heuristics, we evaluated ELGAR by freezing or removing only the ledger input (four static controls) and comparing against a relevance-plus-exposure-deficit controller (Dynamic bias controller) at the same group-token step on Beauty with TIGER. ELGAR achieves tighter local fairness alignment than all four static controls across both request settings, and outperforms the Dynamic bias controller under nonstationary shifts without sacrificing overall utility. Although both ELGAR and the Dynamic bias controller track cumulative exposure deficits, ELGAR fuses ledger states with user representations via a neural fairness head instead of applying a rigid linear offset, enabling smoother adaptation when request mixes shift.

We compare ELGAR with five control policies: Static Bias, fixed Head-, Tail-, and zero-deficit ledgers, and the Dynamic bias controller. All policies reuse the same trained Beauty-TIGER checkpoint without retraining, preserving the group-token generator, fairness head, and training supervision. Static Bias retains request-dependent fairness scores and replaces the ledger contribution with a constant Tail-group logit offset bb; the three fixed-ledger controls keep the ledger input fixed to states representing Head underexposure, Tail underexposure, or target alignment. The Dynamic bias controller retains the same request-dependent scores but replaces the learned ledger contribution with an offset proportional to the current cumulative Tail-exposure deficit, scaled by αctrl\alpha_{\mathrm{ctrl}}. Both bias controls act at the same group-token step as ELGAR. Here, λdec\lambda_{\mathrm{dec}} is the inference-time weight controlling the strength of the ledger-guided fairness intervention. All policies and parameters were tuned on the validation set. Statistics were computed per nonstationary order, then averaged over five orders from five random seeds (2026-2030).

We conduct the experiments on Beauty with TIGER and test the effectiveness in two situations. Default uses the original test order in our paper; we add a Nonstationary situation in which we reorder the same requests into eight alternating Head- and Tail-leaning blocks using visible user histories to test adaptation to changing request mixes.

Evaluation metrics. As each stream contains multiple requests, we use Full-stream MGU@10 to measure aggregate exposure deviation from the target across all requests. Mean window MGU@10 averages the deviations computed within complete trailing 500-request windows; lower values mean closer average alignment as request mixes change.

Table R1. Routing-control comparison on Beauty-TIGER under Default and Nonstationary request orders.

Within each Test order setting, bold marks the best value and underlining‾\underline{\text{underlining}} marks the second-best value for each metric. ↑\uparrow/↓\downarrow indicates whether higher/lower is better.

Table R1.1. Default.

PolicyNDCG@10 ↑\uparrowTail NDCG@10 ↑\uparrowFull-stream MGU@10 ↓\downarrowMean window MGU@10 ↓\downarrow
ELGAR (λdec=1.2\lambda_{\mathrm{dec}}=1.2)0.028930.00706‾\underline{0.00706}0.01465‾\underline{0.01465}0.01863‾\underline{0.01863}
Static Bias (b=0.9b=0.9)0.029300.007300.016570.02687
Fixed Head deficit (λdec=0\lambda_{\mathrm{dec}}=0)0.029930.004180.120720.14897
Fixed Tail deficit (λdec=0.4\lambda_{\mathrm{dec}}=0.4)0.02936‾\underline{0.02936}0.006650.017060.02457
Fixed zero deficit (λdec=2\lambda_{\mathrm{dec}}=2)0.028390.003610.148420.15122
Dynamic bias controller (αctrl=3\alpha_{\mathrm{ctrl}}=3)0.029070.006640.014000.01842

Bold: best; underline‾\underline{\text{underline}}: second best.

AuthorsOpenReview ↗

Response to Reviewer FBdx - Part 2/6 (W1 & Q1)

W1 & Q1: Isolate adaptive exposure feedback using frozen-ledger controls (retaining the fairness head) and a relevance-aware deficit controller under comparable validation budgets.

R1(continued):

Table R1.2. Nonstationary.

PolicyNDCG@10 ↑\uparrowTail NDCG@10 ↑\uparrowFull-stream MGU@10 ↓\downarrowMean window MGU@10 ↓\downarrow
ELGAR (λdec=1.1\lambda_{\mathrm{dec}}=1.1)0.028930.007110.01959‾\underline{0.01959}0.02750
Static Bias (b=0.7b=0.7)0.02954‾\underline{0.02954}0.006050.037760.05438
Fixed Head deficit (λdec=0\lambda_{\mathrm{dec}}=0)0.029930.004180.120720.14808
Fixed Tail deficit (λdec=0.4\lambda_{\mathrm{dec}}=0.4)0.029360.00665‾\underline{0.00665}0.017060.05542
Fixed zero deficit (λdec=2\lambda_{\mathrm{dec}}=2)0.028390.003610.148420.14843
Dynamic bias controller (αctrl=2\alpha_{\mathrm{ctrl}}=2)0.029050.006490.020460.03939‾\underline{0.03939}

Bold: best; underline‾\underline{\text{underline}}: second best.

1. Comparison with static controls: ELGAR consistently delivers tighter local exposure alignment than all four static policies, especially when request mixes alternate. Compared with the four static controls, ELGAR achieved the lowest full-stream and mean window MGU in Default (0.01465 vs. 0.01657 and 0.01863 vs. 0.02457), and lowered nonstationary full-stream MGU to 0.01959 (vs. 0.03776 for Static Bias). Fixed Tail deficit had slightly lower full-stream MGU (0.01706 vs. 0.01959), but much higher mean window MGU (0.05542 vs. 0.02750). Furthermore, ELGAR achieved 17.6% higher Tail NDCG with 2.0% lower overall NDCG than Static Bias on nonstationary streams.

2. Comparison with the Dynamic bias controller: Both policies leverage dynamic exposure ledgers to achieve comparable aggregate alignment under stationary traffic, but ELGAR's neural ledger fusion yields substantially stronger local stability and tail accuracy under nonstationary shifts. Because the Dynamic bias controller also tracks the cumulative exposure deficit, it matches ELGAR on full-stream MGU (0.01400 vs. 0.01465) and mean window MGU (0.01842 vs. 0.01863) in Default, though ELGAR achieves 6.2% higher Tail NDCG with only 0.49% lower overall NDCG. However, under Nonstationary shifts, the Dynamic bias controller degrades markedly on Mean window MGU@10 (0.03939 vs. 0.02750, a 43.2% higher local deviation than ELGAR) and yields 8.7% lower Tail NDCG. This gap stems from how each method utilizes the ledger state: a linear deficit rule applies a uniform score offset regardless of user preferences, causing lag or over-correction when request distributions alternate, whereas ELGAR fuses the ledger with request representations through a neural network to adjust group routing smoothly across shifting user contexts.

We have added this analysis to Section 4.4 and provided the full setup and metric trajectories in Appendix M of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer FBdx - Part 3/6 (Q2)

Q2: Clarify Appendix J.2's state-sufficiency claim under the (17,4)(17, 4) vs. (170,40)(170, 40) counterexample and assess the impact of omitting total exposure counts.

R2:

We thank the reviewer for the concrete counterexample and for asking us to clarify the state-sufficiency claim in Appendix J.2. The counterexample occurs when adding one Tail exposure overshoots the target and yields a larger deviation than adding one Head exposure. To formalize this, we derived the exact analytical conditions under which omitting the exposure total changes the preferred one-step group action. We confirm that the clipped ledger is an approximate state representation past the warm-up horizon, and prove that the preferred action depends on the exposure total only within this narrow overshoot regime. Outside this regime, the clipped ledger preserves the exact exposure-imbalance direction without needing the cumulative exposure total.

The two count vectors have the same exposure proportions, with Head over-exposed and Tail under-exposed. However, Eq. (8) selects the group that leaves the smaller deviation after adding one exposure, and the effect of this addition depends on the total exposure count NexpN_{\mathrm{exp}}. At n=(17,4)\mathbf{n}=(17,4), adding one Tail exposure overshoots the target and produces a larger deviation than adding one Head exposure. At n′=(170,40)\mathbf{n}'=(170,40), adding one Tail exposure reduces the deviation without overshooting. The precise conditions are as follows.

Exact condition for disagreement when G=2G=2

Let Nexp=nH+nT>0N_{\mathrm{exp}}=n_H+n_T > 0, d=ρH−πHd=\rho_H-\pi_H, and πH>1/2\pi_H > 1/2. Since ρT−πT=−d\rho_T-\pi_T=-d,

D=2|d|,\qquad r(g)=2|d|-2|d'_g|. \end{equation}$$ The deviations after adding one Tail or Head exposure are $$\begin{equation} d'_T =d-\frac{\rho_H}{{N_{\mathrm{exp}}+1}}=\frac{{N_{\mathrm{exp}}d-\pi_H}}{{N_{\mathrm{exp}}+1}}, \end{equation}$$ $$\begin{equation} d'_H = d+\frac{\rho_T}{{N_{\mathrm{exp}}+1}} = \frac{{N_{\mathrm{exp}} d+1-\pi_H}}{{N_{\mathrm{exp}}+1}}. \end{equation}$$ Therefore, $$\begin{equation} |d'_H|^2-|d'_T|^2 = \frac{{2N_{\mathrm{exp}} d+1-2\pi_H}}{{(N_{\mathrm{exp}}+1)^2}}. \end{equation}$$ Eq. (8) strictly prefers Head if and only if $$\begin{equation} d < \frac{2\pi_H-1}{{2N_{\mathrm{exp}}}}. \end{equation}$$ Consequently, when Head is over-exposed ($d > 0$), **a strict disagreement with the underexposure direction occurs if and only if $d$ falls inside the narrow range** $$\begin{equation} \boxed{ 0 < d < \frac{2\pi_H-1}{{2N_{\mathrm{exp}}}}. } \end{equation}$$ **At equality, both actions yield the same gain. Outside this range ($d \ge \frac{2\pi_H-1}{2N_{\mathrm{exp}}}$), Tail remains the preferred action.** This disagreement happens only when adding Tail overshoots the target severely. Specifically, adding Tail crosses the target ($d'_T < 0$) whenever $d < \pi_H / N_{\mathrm{exp}}$. Within this overshoot region, two regimes arise: - **Severe overshoot ($0 < d < \frac{2\pi_H-1}{2N_{\mathrm{exp}}}$):** The Tail overshoot $|d'_T|$ exceeds half the step separation $\frac{1}{2(N_{\mathrm{exp}}+1)}$, leaving a larger deviation than adding Head. Thus, Head is preferred. - **Mild overshoot ($\frac{2\pi_H-1}{2N_{\mathrm{exp}}} \le d < \frac{\pi_H}{N_{\mathrm{exp}}}$):** Adding Tail crosses the target slightly, but still produces a smaller deviation than adding Head. Thus, Tail remains preferred. When $d < 0$, Head is under-exposed. Because $(2\pi_H-1)/(2N_{\mathrm{exp}}) > 0$, the condition $d < (2\pi_H-1)/(2N_{\mathrm{exp}})$ always holds, so Head is always preferred without disagreement.
Reviewer_ZboROpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes ELGAR, an Exposure-Ledger Guided Adaptive Routing framework designed to mitigate popularity bias in Semantic-ID (SID) based generative recommendation. The authors identify that autoregressive SID generation amplifies popularity bias because early pruning disproportionately eliminates long-tail items. To address this, ELGAR prepends a popularity-group token to each item's existing SID, forcing group-level allocation to occur as the first decoding decision before semantic pruning. An Exposure Ledger tracks cumulative deviations from a training-interaction reference distribution and conditions a dedicated fairness head during both training and stateful inference. A fused objective combines generative relevance, target-group supervision, and one-step fairness gains. Extensive experiments across four datasets and three backbones demonstrate that ELGAR significantly reduces exposure misalignment while preserving overall recommendation utility.

Critical strengths

Strengths:

1) The core idea of intervening at the very first autoregressive step via a prepended group token is highly elegant. By making group allocation the initial routing decision, the method effectively prevents the irreversible pruning of underexposed item groups inherent in standard SID beam search.

2) The Exposure Ledger mechanism provides a principled, stateful approach to fairness. Unlike static reweighting or post-hoc reranking, dynamically tracking realized exposure discrepancies allows the model to adaptively correct imbalances as serving requests arrive.

3) The paper offers exceptional theoretical depth for a recommendation systems submission. Interpreting the fairness regularization as a generalized MAP estimator with an exposure-alignment prior provides strong mathematical grounding for the proposed loss functions.

4) The empirical evaluation is remarkably comprehensive, spanning multiple datasets and distinct architectural backbones (encoder-decoder vs. decoder-only). The inclusion of fine-grained distribution alignment visualizations and rule-based random routing comparisons thoroughly validates the design choices.

Critical weaknesses

Weaknesses:

1) The reliance on predefined, discrete popularity groups limits the granularity of fairness control. As shown in the ablation where increasing routing cardinality degrades NDCG, the method struggles to balance fine-grained exposure alignment with utility preservation, potentially failing to protect individual long-tail items within broad groups.

2) The training-time ledger construction uses shuffled mini-batches to simulate short exposure episodes, which fundamentally mismatches the persistent, sequential nature of real-world serving streams. Although the authors argue for state sufficiency over trajectory matching, this discrepancy may lead to suboptimal router calibration when deployed in continuous production environments.

3) While prepending a group token preserves the original semantic codes, it alters the length and structure of the generation sequence. This structural modification might subtly interfere with the pre-trained positional encodings or attention dynamics of the backbone models, yet the paper lacks probing analyses to confirm that within-group semantic representations remain entirely unperturbed.

4) The framework assumes a static training-interaction reference distribution. In practical deployments where item catalogs and user interest distributions shift continuously over time, a fixed reference target may become obsolete, and the paper does not discuss mechanisms for dynamically updating the ledger's baseline without full retraining.

Questions for authors

Questions:

1) Given that finer routing granularity (e.g., G=10) leads to progressive NDCG degradation, how would the framework perform if the group token was replaced by a continuous latent fairness embedding rather than discrete categorical tokens?

2) How robust is the Exposure Ledger to sudden distribution shifts, such as viral trends causing abrupt spikes in specific item popularity, and can the horizon parameter H be adapted online to handle such non-stationarity?

3) Since the fairness intervention only occurs at the first decoding step, could a multi-armed bandit formulation replace the deterministic ledger updates to better handle the exploration-exploitation trade-off between exposing tail groups and maintaining immediate relevance?

10 comments
AuthorsOpenReview ↗

Response to Reviewer ZboR - Part 1/7 (W1 & Q1)

Thank you for your positive assessment of ELGAR and for your constructive feedback; we address all points below, grouping closely related items.

W1 & Q1: Address fairness granularity limits at larger routing cardinalities and evaluate replacing discrete group tokens with continuous latent fairness embeddings.

R1:

Thank you for this insightful observation. To examine the granularity of fairness control in W1 and the continuous latent fairness embedding alternative in Q1, we evaluated group-level utility across finer routing cardinalities and implemented a continuous prefix-conditioning variant. Our experiments yield two qualitative conclusions: (1) coarse routing best preserves tail recommendation quality, whereas overly fine routing fragments the fixed beam budget and hurts generalization; and (2) replacing discrete categorical tokens with a continuous latent fairness embedding worsens utility degradation under fine-grained routing.

These findings support keeping the initial fairness allocation coarse and selecting routing granularity on validation data to balance exposure alignment with overall and tail utility. A practical strategy is to start with a small number of groups and increase granularity when the additional exposure alignment justifies the observed utility cost. This follows the coarse-to-fine organization of SID decoding: the initial route allocates exposure broadly, while subsequent semantic decisions refine item selection. The tested continuous alternative improves exposure alignment but does not resolve the utility cost of finer initial routing.

To examine W1 in detail, we report overall, Head, and Tail NDCG@10 as routing cardinality GrouteG_{\mathrm{route}} increases with TIGER across three datasets (see Reviewer ZboR Table R1.1 to R1.3).

Table R1. Effect of routing cardinality GrouteG_{\mathrm{route}} on overall, Head, and Tail NDCG@10 with TIGER across three datasets.

Table R1.1. Beauty.

GrouteG_{\mathrm{route}}NDCG@10 ↑\uparrowHead NDCG@10 ↑\uparrowTail NDCG@10 ↑\uparrow
20.02910.0418‾\underline{0.0418}0.0069
30.02810.04090.0057‾\underline{0.0057}
50.0283‾\underline{0.0283}0.04190.0047
80.02740.04020.0050
100.02690.03960.0050

Bold: best; underline‾\underline{\text{underline}}: second best.

Table R1.2. Sports.

GrouteG_{\mathrm{route}}NDCG@10 ↑\uparrowHead NDCG@10 ↑\uparrowTail NDCG@10 ↑\uparrow
20.01610.02420.0023
30.01390.0204‾\underline{0.0204}0.0028‾\underline{0.0028}
50.0142‾\underline{0.0142}0.0204‾\underline{0.0204}0.0036
80.01220.01810.0020
100.01350.02030.0019

Bold: best; underline‾\underline{\text{underline}}: second best.

Table R1.3. Toys.

GrouteG_{\mathrm{route}}NDCG@10 ↑\uparrowHead NDCG@10 ↑\uparrowTail NDCG@10 ↑\uparrow
20.0261‾\underline{0.0261}0.03430.0129
30.0261‾\underline{0.0261}0.0344‾\underline{0.0344}0.0128‾\underline{0.0128}
50.02650.03540.0122
80.02380.03270.0095
100.02550.03430.0115

Bold: best; underline‾\underline{\text{underline}}: second best.

Although the trends are not strictly monotonic, the finest routing configurations (Groute∈{8,10}G_{\mathrm{route}} \in \{8,10\}) yield lower overall and Tail NDCG@10 than Groute=2G_{\mathrm{route}}=2 on all three datasets, while Head NDCG@10 also generally declines. Tail NDCG@10 peaks at Groute=2G_{\mathrm{route}}=2 on Beauty and Toys, and at Groute=5G_{\mathrm{route}}=5 on Sports.

We attribute this trade-off to candidate-space fragmentation and increased routing difficulty: partitioning items into more groups increases group classification complexity, spreads the fixed beam budget (20) across more initial prefixes, tightens within-group search capacity, and reduces per-group training support. Moreover, distributional comparisons on Beauty (Appendix Figures 8 and 9) show that larger GrouteG_{\mathrm{route}} aligns exposure closer to the training reference while yielding smoother profiles against held-out test targets without closer shape alignment, indicating that enforcing fine-grained training targets constrains test-time generalization.

AuthorsOpenReview ↗

Response to Reviewer ZboR - Part 2/7 (W1 & Q1)

W1 & Q1: Address fairness granularity limits at larger routing cardinalities and evaluate replacing discrete group tokens with continuous latent fairness embeddings.

R1(continued):

Because our work targets group-level exposure calibration instead of individual item-level exposure fairness, group partitions cannot be omitted and must remain as the discrete basis for distribution alignment. Therefore, to isolate the effect of continuous embeddings raised in Q1, we remove discrete group tokens from both user histories and target SIDs. Two small networks map the user representation and exposure-ledger state to 128-dimensional vectors, which are fused through a weighted sum and layer normalization to form a continuous fairness prefix for SID decoding. A learned bridge maps this prefix to routing probabilities over the GrouteG_{\mathrm{route}} candidate groups, which guide constrained beam search within valid group subtrees without feeding discrete group tokens into the decoder.

We evaluated this alternative on Beauty with TIGER at Groute∈{2,10}G_{\mathrm{route}}\in\{2,10\}, using seed 2026, the original training budget and checkpoint-selection rule, and a global beam size of 20. Checkpoints were selected by validation NDCG@20 at λdec=0\lambda_{\mathrm{dec}}=0, and we report test results at the prespecified λdec=1\lambda_{\mathrm{dec}}=1, consistent with the granularity experiment above (see Reviewer ZboR Table R2). MGU is evaluated using the same fixed ten groups for all rows, while Tail NDCG uses the original Head/Tail partition.

Table R2. Comparison of discrete categorical tokens and continuous latent fairness embeddings on Beauty with TIGER.

Routing representationGrouteG_{\mathrm{route}}NDCG@10 ↑\uparrowMGU@10 ↓\downarrowTail NDCG@10 ↑\uparrow
Discrete categorical token20.029100.054510.00694
Continuous latent fairness embedding20.02739‾\underline{0.02739}0.043060.00382
Discrete categorical token100.026930.01154‾\underline{0.01154}0.00497‾\underline{0.00497}
Continuous latent fairness embedding100.020620.008820.00340

Bold: best; underline‾\underline{\text{underline}}: second best.

Overall, continuous prefix conditioning achieves tighter reference alignment at fine granularity but incurs substantially steeper utility loss than discrete group tokens. Increasing GrouteG_{\mathrm{route}} from 2 to 10 reduces NDCG@10 by 24.7% with continuous conditioning, compared with 7.5% with discrete tokens. At Groute=10G_{\mathrm{route}}=10, the continuous variant lowers MGU@10 by 23.6% relative to discrete categorical tokens, but reduces overall NDCG@10 by 23.4% and Tail NDCG@10 by 31.7%. Across the shared validation grid λdec∈{0,0.1,0.2,0.5,1}\lambda_{\mathrm{dec}}\in\{0,0.1,0.2,0.5,1\}, no continuous Groute=10G_{\mathrm{route}}=10 setting reached the common utility floor of 95% of discrete Groute=2G_{\mathrm{route}}=2's validation NDCG@10 at zero decoding weight. A separate direction for future investigation is to retain coarse initial routing and introduce continuous fairness control at deeper SID decoding stages.

AuthorsOpenReview ↗

Response to Reviewer ZboR - Part 3/7 (W2 & Q2)

W2 & Q2: Address the mini-batch training vs. sequential serving mismatch, ledger robustness to sudden popularity bursts, and online adaptation of HH.

R2:

Ledger-state coverage during training and inference

Thank you for pointing out that the gap between short exposure episodes and persistent serving streams warrants further investigation. To verify whether shuffled mini-batches cover serving-time conditions for reliable router calibration, we compared the distributions of bounded ledger states (exposure-share deviation and warm-up progress) observed during training and full-stream inference on Beauty with TIGER. We find that inference ledger states fall entirely within the range of states observed during training, so the fairness head never extrapolates out-of-distribution. Because the ledger passes bounded normalized features instead of raw stream length, short exposure episodes cover both cold-start and steady-state regimes without replicating a full serving trajectory.

Specifically, during training, the ledger resets in each mini-batch (at most 256 exposures), whereas during inference, it accumulates throughout the entire test stream (223,630 exposures across 22,363 requests, approximately 870 times the length of a training episode).

Because the ledger passes only two bounded normalized features, namely the Head exposure-share deviation d=ρH−πHd=\rho_H-\pi_H and the warm-up progress min⁡(Nexp/H,1)\min(N_{\mathrm{exp}}/H,1) with Nexp=∑g=1GngN_{\mathrm{exp}}=\sum_{g=1}^{G} n_g and H=20H=20, distributional consistency depends on (d,progress)(d,\text{progress}) coverage instead of episode length.

Training covers diverse states of both over- and under-exposure. Shuffled mini-batches generate 657,065 states in total. Of these, 92.2% have progress equal to 1, with dd ranging from −0.35-0.35 to +0.20+0.20 and Head over-exposure (d>0d > 0) and under-exposure (d<0d < 0) each accounting for approximately half. The remaining 7.8% are cold-start states (progress<1\text{progress} < 1).

Table R3. Ledger states observed during training.

StateCountShareRange of dd
Progress<1\text{Progress} < 1 (cold start)51,4007.8%[−0.80,+0.20][-0.80,+0.20]
Progress=1\text{Progress} = 1605,66592.2%[−0.35,+0.20][-0.35,+0.20]

Inference states fall entirely within the range covered by training. The first two inference requests are in the cold-start phase (progress<1\text{progress} < 1), which is represented by 7.8% of training states. From the 3rd request on, the progress is 1 and dd quickly settles around +0.07+0.07: 99.8% of the states lie in [0.05, 0.10)[0.05,\,0.10) and the remaining 0.2% in [0, 0.05)[0,\,0.05). Training contains 52,548 and 254,430 states in these two intervals, respectively. All post-warm-up inference states lie within the central 99% interval of the training distribution of dd. The settled value agrees with the MGU@10 of 0.0703 reported in the paper (with two groups, MGU equals ∣d∣|d|).

Table R4. Distribution of d=ρH−πHd=\rho_H-\pi_H during training and inference at progress=1\text{progress} = 1.

Interval of ddTraining countTraining shareInference share
<−0.10< -0.109,7611.6%0
[−0.10,−0.05)[-0.10,-0.05)48,1097.9%0
[−0.05,0)[-0.05,0)233,50538.6%0
[0,0.05)[0,0.05)254,43042.0%0.2%
[0.05,0.10)[0.05,0.10)52,5488.7%99.8%
≥0.10\ge 0.107,3121.2%0

Although [0.05,0.10)[0.05,0.10) lies between the 90th and 99th percentiles of the training distribution (8.7% of training states) outside its densest mode, the fairness head never extrapolates out-of-distribution during the continuous 22,363-request test stream. The corresponding distribution histograms and convergence trajectory are provided in Appendix N.1, Fig. 18 and Fig. 19.

Reviewer_RVbkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes ELGAR (an Exposure-Ledger Guided Adaptive Routing framework) that addresses popularity-calibrated exposure fairness in semantic-ID generative recommendation. It prepends a popularity-group token to each item identifier and uses a cumulative exposure ledger to adjust group-routing logits before semantic decoding. Via experiments on four datasets and three backbones, the paper shows that the proposed framework generally reduces exposure mismatch and improves tail-item recommendation while keeping overall recommendation quality close to the backbone.

Critical strengths
  • The problem studied here is well motivated and addresses a concrete generative-recommendation limitation.
  • The experimentation in this paper is quite comprehensive. The paper evaluates multiple datasets and architectures, includes component ablations and routing-granularity studies, and explicitly examines the fairness–utility trade-off.
Critical weaknesses
  • It is unclear whether the exposure state actually drives the gains. Setting lambda_{dec}=0 removes the entire fairness correction at inference time, so it does not isolate the benefit of using a changing ledger. Could the authors add a fixed-ledger ablation or a tuned static group-logit offset to show whether cumulative-exposure adaptation adds anything beyond a learned group bias?
  • FairCo and P-MMF also adjust allocation based on accumulated exposure, but the paper only discusses them in related work. I wonder if some comparisons can be added?
Questions for authors

See weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer RVbk - Part 1/2 (W1)

Thank you for recognizing ELGAR’s strengths and for your constructive questions; we address them below.

W1: Isolate the gains of cumulative-exposure adaptation beyond a static group bias via fixed-ledger or tuned static-offset ablations.

R1:

To isolate whether the changing exposure state drives ELGAR’s gains beyond a learned group bias, we added a fixed-ledger ablation and a tuned static group-logit offset on Beauty with TIGER. Our results show that cumulative-exposure adaptation is essential for local exposure alignment and tail recommendation quality under shifting request mixes. Static offsets and fixed ledgers apply rigid corrections that fail when user popularity preferences fluctuate, whereas ELGAR’s updating ledger tracks accumulated exposure to prevent local over- and under-exposure.

Among the four control policies, Static Bias retains request-dependent fairness scores and replaces the ledger contribution with a constant Tail-group logit offset bb; the three fixed-ledger controls keep the ledger input fixed to states representing Head underexposure, Tail underexposure, or target alignment. Here, λdec\lambda_{\mathrm{dec}} is the inference-time weight controlling the strength of the ledger-guided fairness intervention. All policies were tuned on the validation set. Statistics were computed per nonstationary order, then averaged over five orders from five random seeds (2026–2030).

We conduct the experiments on Beauty with TIGER and test the effectiveness in two situations. Default uses the original test order in our paper; we add a Nonstationary situation in which we reorder the same requests into eight alternating Head- and Tail-leaning blocks using visible user histories to test adaptation to changing request mixes. Other datasets and backbones exhibit a similar trend.

Evaluation metrics. As each stream contains multiple requests, we use Full-stream MGU@10 to measure aggregate exposure deviation from the target across all requests. Mean window MGU@10 averages the deviations computed within complete trailing 500-request windows; lower values mean closer average alignment as request mixes change.

Table R1. Comparison of ELGAR with static-bias and fixed-ledger controls on Beauty with TIGER.

Table R1.1. Default.

PolicyNDCG@10 ↑\uparrowTail NDCG@10 ↑\uparrowFull-stream MGU@10 ↓\downarrowMean window MGU@10 ↓\downarrow
ELGAR (λdec=1.2\lambda_{\mathrm{dec}}=1.2)0.028930.00706‾\underline{0.00706}0.014650.01863
Static Bias (b=0.9b=0.9)0.029300.007300.01657‾\underline{0.01657}0.02687
Fixed Head deficit (λdec=0\lambda_{\mathrm{dec}}=0)0.029930.004180.120720.14897
Fixed Tail deficit (λdec=0.4\lambda_{\mathrm{dec}}=0.4)0.02936‾\underline{0.02936}0.006650.017060.02457‾\underline{0.02457}
Fixed zero deficit (λdec=2\lambda_{\mathrm{dec}}=2)0.028390.003610.148420.15122

Bold: best; underline‾\underline{\text{underline}}: second best.

Table R1.2. Nonstationary.

PolicyNDCG@10 ↑\uparrowTail NDCG@10 ↑\uparrowFull-stream MGU@10 ↓\downarrowMean window MGU@10 ↓\downarrow
ELGAR (λdec=1.1\lambda_{\mathrm{dec}}=1.1)0.028930.007110.01959‾\underline{0.01959}0.02750
Static Bias (b=0.7b=0.7)0.02954‾\underline{0.02954}0.006050.037760.05438‾\underline{0.05438}
Fixed Head deficit (λdec=0\lambda_{\mathrm{dec}}=0)0.029930.004180.120720.14808
Fixed Tail deficit (λdec=0.4\lambda_{\mathrm{dec}}=0.4)0.029360.00665‾\underline{0.00665}0.017060.05542
Fixed zero deficit (λdec=2\lambda_{\mathrm{dec}}=2)0.028390.003610.148420.14843

Overall, ELGAR maintains consistently tight exposure alignment across both stationary and nonstationary streams, whereas static controls suffer severe local exposure drift when request mixes alternate. In Default, ELGAR achieves the lowest full-stream MGU@10 (0.01465 vs. 0.01657 for Static Bias) and mean window MGU@10 (0.01863 vs. 0.02457 for Fixed Tail deficit). Under Nonstationary streams, ELGAR cuts mean window MGU@10 by 49.4% compared with Static Bias (0.02750 vs. 0.05438) and by 50.4% compared with Fixed Tail deficit (0.02750 vs. 0.05542). Although Fixed Tail deficit attains a slightly lower full-stream MGU@10 (0.01706 vs. 0.01959), its local window deviation is more than twice ELGAR's because a constant offset cannot track alternating Head- and Tail-leaning blocks.

ELGAR’s window MGU standard deviation was also lower than both controls’ in all five nonstationary orders, averaging 0.01816 compared with 0.03800 for Static Bias and 0.02224 for Fixed Tail deficit, alongside higher Tail NDCG@10 (0.00711 vs. 0.00605) at a 2.0% lower overall NDCG@10.

We have added this analysis to Section 4.4 and provided the full setup and metric trajectories in Appendix M of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer RVbk - Part 2/2 (W2)

W2: Provide empirical comparisons against exposure-accumulating post-processing baselines, FairCo and P-MMF.

R2:

To compare ELGAR with post-processing baselines that adjust allocation from accumulated exposure, we evaluated FairCo and P-MMF on Beauty with TIGER under the same beam width and alignment target. ELGAR achieves substantially stronger exposure alignment and tail recommendation accuracy than both baselines while keeping overall utility comparable. Post-hoc rerankers underperform in generative recommendation because autoregressive beam search may prune most tail SID prefixes early, leaving too few tail candidates to promote. In contrast, ELGAR intervenes at the prepended popularity-group token before semantic decoding to preserve tail reachability.

In our setup, both post-processors re-rank TIGER’s 20 beam-search candidates into the final Top-KK list (beam width 20, matching TIGER and ELGAR). Both baselines reuse our Head/Tail item partition (mapped to fairness groups in FairCo and providers in P-MMF) and use the training-interaction shares πg\pi_g as the alignment reference. Following their standard formulations, we tune FairCo’s proportional gain KP∈{0,10−9,…,100}K_P \in \{0, 10^{-9}, \ldots, 10^0\} and P-MMF’s trade-off weight λ∈{10−3,…,10}\lambda \in \{10^{-3}, \ldots, 10\} and step size ηpmmf∈{10−4,…,1}\eta_{\mathrm{pmmf}} \in \{10^{-4}, \ldots, 1\} (T=256T=256, K=10K=10, momentum α=0.1\alpha=0.1) on the validation set, selecting the configuration that minimizes validation MGU@10 while retaining at least 95% of vanilla TIGER’s validation NDCG@10.

Table R2. Comparison of ELGAR, FairCo, and P-MMF on Beauty with TIGER.

DatasetMethodHR@5 ↑\uparrowHR@10 ↑\uparrowNDCG@5 ↑\uparrowNDCG@10 ↑\uparrowTail NDCG@5 ↑\uparrowTail NDCG@10 ↑\uparrowMGU@5 ↓\downarrowMGU@10 ↓\downarrow
BeautyTIGER0.03610.05590.02400.0303‾\underline{0.0303}0.00090.00180.18830.1827
Beauty+ FairCo0.0360‾\underline{0.0360}0.05510.0242‾\underline{0.0242}0.03040.00130.00200.17860.1701
Beauty+ P-MMF0.03610.05480.02430.0303‾\underline{0.0303}0.0017‾\underline{0.0017}0.0041‾\underline{0.0041}0.1763‾\underline{0.1763}0.1074‾\underline{0.1074}
Beauty+ ELGAR (Ours)0.03590.0556‾\underline{0.0556}0.02340.02980.00370.00540.09620.0703

Bold: best; underline‾\underline{\text{underline}}: second best.

Empirically, FairCo and P-MMF remain close to vanilla TIGER in exposure mismatch, whereas ELGAR cuts MGU@10 by 58.7% relative to FairCo (0.0703 vs. 0.1701) and by 34.5% relative to P-MMF (0.0703 vs. 0.1074), while raising Tail NDCG@10 by 170.0% and 31.7% with minimal change in overall NDCG@10 (0.0298 vs. 0.0304 for FairCo and 0.0303 for P-MMF). Unlike full-catalog scoring in conventional recommendation, generative beam search restricts post-processors to the 20 surviving candidates, leaving pruned tail items unrecoverable. Even under P-MMF’s budget masking, the reranker must fill the remaining Top-KK slots from over-budget Head groups once valid tail candidates in the beam are exhausted. In contrast, ELGAR adjusts group-routing logits at the prepended popularity-group token before semantic decoding, expanding the set of reachable tail items instead of reordering an already pruned list.

We added this comparison and discussion to Appendix L of the revised paper.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal and Score Re-evaluation

Dear Reviewer RVbk,

Thank you again for your time and constructive feedback. We have tried our best to address all of your concerns in our rebuttal. If our responses have resolved your questions, we kindly ask if you would consider re-evaluating your score. Please let us know if you have any further questions, as we would be glad to answer them.

Best regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.