Slate-GLM-OFU: Efficient Logistic Slate Bandit Method

• Slate-GLM-OFU is an algorithmic framework that optimizes decision-making in logistic contextual slate bandits by combining local slot-level optimization with global logistic parameter estimation.
• The approach employs UCB-style slot planning and adaptive confidence-set updates to achieve a regret bound of O(√T) while keeping per-round computation at N^(O(1)).
• Empirical evaluations demonstrate its scalability and superior performance in synthetic and practical settings, maintaining sub-millisecond decision times even in high-dimensional slate spaces.

Slate-GLM-OFU is an algorithmic framework designed for efficient decision-making in the logistic contextual slate bandit problem with bandit feedback. Its architecture combines local slot-level optimization and global joint parameter estimation, achieving $\widetilde{O}(\sqrt{T})$ regret and $N^{O(1)}$ per-round computational cost. This approach enables scalable solutions to environments where, at each round, a slate of $N$ items—each from a large candidate set—must be selected, and a single binary reward is observed, governed by an unknown global logistic model. The method leverages both upper-confidence-bound (UCB) style slot-wise planning and a principled confidence-set update scheme adapted from logistic bandit theory (Goyal et al., 16 Jun 2025).

1. Problem Formulation and Notation

The logistic contextual slate bandit problem is characterized by a sequence of rounds $t=1,2,\ldots,T$, where, at each round:

• $N$ slots are presented, each with a set $\mathcal X_t^i\subset\mathbb{R}^d$ of candidate items.
• The learner selects $\bfx_t^i\in\mathcal X_t^i$ per slot, forming the slate $\bfx_t = (\bfx_t^1,...,\bfx_t^N)$.
• A global logistic model yields a binary reward $y_t\in\{0,1\}$ with probability $\mu(\bfx_t^\top\btheta^*)$, where $N^{O(1)}$0 is unknown and $N^{O(1)}$1.
• Regret is measured against the best slate in hindsight:

$N^{O(1)}$2

A key regularity parameter, $N^{O(1)}$3, captures the nonlinearity of the link function and shapes the regret bounds.

The diversity assumption, which underpins the algorithmic guarantees, imposes slot-level moment conditions: $N^{O(1)}$4 for each slot $N^{O(1)}$5 and round $N^{O(1)}$6.

2. Algorithmic Structure and Core Components

Slate-GLM-OFU operationalizes two main principles:

• Local Planning (Slot-Level Optimization): Each slot is selected independently, using slot-wise UCB bonuses. This circumvents the exponential blowup associated with the combinatorial slate space $N^{O(1)}$7.
• Global Learning (Joint Parameter Estimation): A shared logistic parameter $N^{O(1)}$8 is learned jointly across all slots. The approach adapts the ada-OFU-ECOLog subroutine for slot-wise and slate-wise concentration matrix updates.

Design matrices maintained are:

• The slate-level matrix:

$N^{O(1)}$9

• The slot-level proxies:

$N$0

Optimism is incorporated through slot-wise bonuses:

$N$1

A global confidence set $N$2 maintains the plausible parameter region for $N$3, updated after each reward via adaptive cross-entropy or MLE-style resets, depending on a second-derivative criterion.

3. Computational Workflow and Pseudocode Insights

The algorithm iterates over rounds as follows:

1. For each slot $N$4, select $N$5 maximizing the sum of the estimated reward and its UCB bonus.
2. Pull the formed slate $N$6, observe $N$7.
3. Update parameter estimate $N$8 and design matrices using an adaptive, error-controlled procedure.

The update routine computes auxiliary estimates via cross-entropy minimization. If a local curvature condition is met, it makes a regularized update; otherwise, it accumulates the data into a "reset" set and recomputes an MLE plus confidence set. This scheme ensures computational efficiency while respecting nonlinearity in the logistic link (Goyal et al., 16 Jun 2025).

The update and selection routines are detailed in Algorithm 1 and 2 in the cited source, emphasizing independence at the slot level and global parameter sharing.

4. Regret and Theoretical Guarantees

Under the diversity assumption, the sequence of slot-level design matrices grows linearly. A block-matrix argument shows that the global slate-level matrix and the diagonal of slot-level matrices are multiplicatively close:

$N$9

This ensures that slot-wise UCBs adequately control uncertainty in the global parameter space.

The principal regret theorem guarantees:

$t=1,2,\ldots,T$0

where $t=1,2,\ldots,T$1 hides polylogarithmic quantities and lower-order terms of $t=1,2,\ldots,T$2 for rare "reset" rounds.

Per-round computational complexity is $t=1,2,\ldots,T$3, dominated by $t=1,2,\ldots,T$4 independent scans over candidate sets and (typically) two small convex programs per update, with exponential runtime avoided outside rare adverse rounds.

5. Empirical Evaluation and Practical Performance

Comprehensive experiments validate Slate-GLM-OFU across synthetic and practical settings:

• Contextual Synthetic: For $t=1,2,\ldots,T$5, $t=1,2,\ldots,T$6, $t=1,2,\ldots,T$7, and horizons up to $t=1,2,\ldots,T$8, Slate-GLM-OFU achieves the lowest regret, with Slate-GLM-TS nearly matching global logistic-bandit baselines at 10–100× faster runtime.
• Scalability: As $t=1,2,\ldots,T$9 increases from 3 to 6 (so $N$0 scales to $N$1), global methods incur exponential decision time, while Slate-GLM-OFU remains sub-millisecond.
• Non-contextual Benchmarking: Slate-GLM-OFU demonstrates superior regret relative to adapted versions of MPS and Ordered Slate Bandits.
• Prompt Selection in LLMs: For SST-2 and Yelp Sentiment tasks, selection of $N$2 in-context examples achieves test accuracies of 69–81% (SST-2) and 74–82.5% (Yelp), outperforming random selection baselines.

The method consistently maintains per-round average running times below 1 ms, scaling favorably for large slate spaces (Goyal et al., 16 Jun 2025).

6. Connections, Context, and Extensions

Slate-GLM-OFU builds on advances in contextual bandit algorithms (notably ada-OFU-ECOLog) and extends optimism-driven planning to the exponentially large slate space by exploiting slot-level decomposability under global parameter sharing. The diversity assumption is essential, ensuring slot-level design matrices are well-conditioned relative to the global design.

This suggests a broader applicability to structured bandit settings where combinatorial explosion is a challenge but diversity and shared parameterization can be meaningfully leveraged. A plausible implication is that extensions to more complex feedback models—such as click models or multi-reward settings—could benefit from the local/global division of labor implemented in Slate-GLM-OFU.

7. Significance and Limitations

Slate-GLM-OFU addresses a critical scaling bottleneck in slate bandit optimization by balancing statistical optimality and computational tractability. Its regret bounds match the optimal rates for this class under mild assumptions, and empirical performance confirms its advantage in both accuracy and speed.

Limitations stem primarily from the diversity assumption and the reliance on global logistic parameterization. In scenarios with extreme context dependence across slots, or where per-slot parameters are needed, alternative approaches may be warranted. The method's update mechanism requires careful tuning of regularization and confidence widths but is robust across the tested domains.

Further details, including implementation resources, extended tables, and code, are provided in the source's supplementary material (Goyal et al., 16 Jun 2025).