Balancing Multiple Objectives in Generative Recommendations with Adaptive Decoding

TL;DR: We introduce Stochastic Primal-Dual Decoding (SPDD), a lightweight inference-time method that steers generative recommenders toward auxiliary objectives, without retraining. In a large-scale A/B test on playlist ranking, SPDD delivered +1.8 percentage points more auxiliary-objective gain than fixed-weight generation, with no loss in user consumption.

Context

Recent advancements in generative AI have transformed how we produce and interact with content. Naturally, this success has motivated a growing interest in applying generative models to recommender systems: building ordered lists of items such as songs, videos, products, or articles. Because these models generate one item at a time, with each selection informing what comes next, they can capture sequential dependencies that conventional ranking approaches miss.

In practice, relevance is rarely the only goal. Recommender systems often need to balance multiple objectives simultaneously, and ideally do so while maintaining a great user experience. This becomes especially interesting when objectives evolve over time and retraining a large generative model is costly.

Why multiobjective generation is different

Many multiobjective techniques were developed for non-generative ranking systems. They first score items independently and then re-rank the finished list to improve on another objective.

That assumption breaks in autoregressive recommendation. Placing an item in one position changes the model’s relevance scores for the next position; the second selection changes the scores for the third, and so on. A post-processing method that rearranges items only after generation therefore acts on scores that may no longer describe the final sequence.

An alternative is to combine relevance with an auxiliary objective, such as a target over item characteristics, directly during generation, using a fixed weighted average at each step. But this applies the same balance between objectives from first item to last. It cannot react to how things are going: it may not push hard enough when the auxiliary objective is falling behind its target, or keep pushing after the target has already been met.

Multiobjective behaviour can also be incorporated directly into model training. However, this tightly couples the model to a particular set of objectives. When those objectives change, retraining or fine-tuning a large model can be expensive and operationally slow.

This raises a practical question: can we steer a generative recommender towards additional objectives during generation, while preserving both its sequential structure and relevance?

A decoder that adapts as it goes

To address this, we introduce Stochastic Primal-Dual Decoding (SPDD), a lightweight layer that sits on top of an existing generative recommender and adds multiobjective control without modifying or retraining the underlying model.

SPDD frames the generation of a recommendation list as an online constrained optimization problem. At each position, the underlying model provides relevance scores for the available items, and each item also carries a score describing how much it contributes to the auxiliary objective.

The decoder combines these signals using a dynamic weight that reflects the current “urgency” of the auxiliary objective. When the list is falling behind its target, the weight increases and gives more priority to items that help close the gap. Once the target is on track, the weight decreases and lets relevance dominate again.

In optimization terms, choosing the next item is the primal decision, while the adaptive weight is a dual variable associated with the auxiliary objective. After every selection, SPDD updates this variable based on the remaining gap: how much progress has been made, how much is still required, and how many positions remain.

Figure 1. The Stochastic Primal-Dual Decoder sits on top of a pre-trained transformer, combining the model’s relevance scores with constraint-related scores before selecting each next item.

Crucially, this mechanism operates entirely at generation time. It can be added to existing autoregressive recommenders and adjusted as objectives evolve, without changing model parameters or training a new system.

Theoretical guarantees for practical control

A single step-size parameter controls how aggressively the decoder reacts: larger values make it more responsive when the auxiliary objective is at risk, while smaller values place more emphasis on relevance. We derive bounds on both the auxiliary-objective gap and the relevance cost, making this trade-off explicit and tunable rather than something that must be discovered empirically. The algorithm also scales linearly with both list length and candidate-set size, making it suitable for low-latency, large-scale deployments.

Results

To demonstrate the generality of SPDD, we evaluated it across three recommendation tasks and three different generative architectures:

• Playlist ranking: Given a user's listening history on Spotify, the task is to order tracks within a personalized playlist. We used approximately 40 billion interactions from 62 million users across more than 42 million tracks, with a SASRec model.

• Home shelf ranking: Spotify's Home page presents rows of content (playlists, artists, podcasts, audiobooks), and the task is to rank these shelves to maximise engagement. We used 5.3 billion shelf impressions from over 500 million user requests, with a TransAct model.

• E-commerce recommendation: Given a user's interaction history, the task is to recommend relevant products. We used the public Amazon Sports and Outdoors dataset, with more than 5 million interactions across over 1 million items, and a LLaMA-based recommender.

In every case, the underlying model was trained for relevance alone; multiobjective control was applied entirely at generation time. While the specific auxiliary objective varies by domain, our formulation is general: it can represent any target defined over item characteristics, such as category allocation, novelty, or diversity.

We compared SPDD with post-generation re-ranking, fixed weighted-average generation, and two multiobjective bandit approaches. Across all three tasks, SPDD achieved a consistent Pareto improvement: for almost any auxiliary-objective target, it delivered stronger user-consumption metrics while meeting a comparable target.

Figure 2. Offline evaluation. SPDD delivers a Pareto-improvement over all baselines across different domains and generative recommenders.

A position-level analysis helps explain why. SPDD makes efficient progress on the auxiliary objective early in the list, then shifts focus back to relevance as soon as the target is on track. This adaptive transition is what gives SPDD its edge.

Validating SPDD in a live system

We ran a large-scale A/B experiment on Spotify with approximately 1 million users across more than 150 countries and 600,000 personalized playlists, comparing a relevance-only control, fixed weighted-average generation, and SPDD.

SPDD increased the auxiliary stream-share metric by +5.44% relative to control, compared with +3.60% for the fixed-weight approach, an additional gain of roughly +1.8 percentage points. It did so without without decreasing user consumption, and produced a smaller skip-rate increase (+0.16% vs. +0.42%).

Figure 3. Online evaluation: SPDD delivers a larger lift in auxiliary stream share than fixed-weight generation, while maintaining consumption and producing a lower skip rate.

Looking ahead

Our work shows that multiobjective control does not need to be permanently encoded in a generative model. By moving that control into a principled generation-time decoder, SPDD combines operational flexibility, theoretical guarantees, and real-world performance — with objectives that can be adjusted without training a new model.

Looking ahead, the framework naturally extends to personalised user-level targets, multiple or non-linear constraints, and non-autoregressive architectures such as diffusion models. As generative recommender systems become more widely deployed, generation-time optimisation provides a practical path toward recommendations that are not only relevant, but also controllable and aligned with multiple objectives.

For more information, please refer to our paper: 

Stochastic Primal-Dual Decoding for Multiobjective Generative Recommender Systems

Dmitrii Moor, Ben Carterette, Senthilkumar Krishnamoorthy, Kyle Kretschman, Denis Beslic, Melissa Yalla, Alice Y Wang, and Mounia Lalmas. 

KDD 2026.