arXiv subjects

Otmane Sakhi

Publications and source records attributed to Otmane Sakhi.

At least 19 recordsLinked to original sources

arXiv2026

Shapley-based Data Valuation for LLM Alignment via Sequential Preference Optimization

Data valuation is a natural framework for understanding which preference datasets matter most when aligning a Large Language Model (LLM) using multiple sources. The standard game-theoretic approach assigns each dataset a contribution score via the Shapley value. In practice, however, Shapley-based valuation is computationally prohibitive because it requires fine-tuning a separate model for every possible coalition of preference datasets, i.e., an exponential number of alignments. We address this challenge for a broad family of preference-optimization objectives, including DPO and IPO, that learn directly from log-policy ratios with respect to a reference policy. We introduce Sequential Preference Optimization, an offline procedure that applies existing preference optimization methods sequentially, source by source, updating the current policy after each dataset. Under exact optimization, this procedure yields an additive composition rule in reward space and an equivalent arithmetic composition rule in policy space. This observation enables an efficient approximation of the Shapley value: we train one model per preference dataset and reconstruct coalition policies at inference time from the singleton models, reducing the required alignments from exponential to linear in the number of sources. Leveraging this property, we compute Shapley values for several real-world preference datasets and reveal how each source drives model alignment.

cs.LG↗

arXiv2026

Exploiting Similarities in A/B Testing with Off-Policy Estimation

We study A/B testing, the standard protocol for measuring the performance gain of a new decision system relative to a baseline. Traditional A/B testing treats both systems as black boxes, ignoring potential similarities between them. In practice, however, new and baseline systems are rarely radically different and often share significant structure, which can be captured by their propensities to make similar decisions. We show that in such cases, the commonly used difference-in-means estimator, though unbiased, is statistically suboptimal. Leveraging off-policy estimation, we introduce a family of A/B testing estimators that exploit the propensities of the tested systems to achieve improved concentration properties. This family is flexible enough to be tailored to practical decision-making. The resulting estimators are simple, robust to propensities misspecification, substantially more accurate when the tested systems exhibit similarities, and gracefully fall back to the difference-in-means estimator when such similarities are absent. Our theoretical analysis and empirical studies confirm their efficiency and practicality.

stat.ML↗

arXiv2026

Off-Policy Learning in Large Action Spaces: Optimization Matters More Than Estimation

Off-policy evaluation (OPE) and off-policy learning (OPL) are foundational for decision-making in offline contextual bandits. Recent advances in OPL primarily optimize OPE estimators with improved statistical properties, assuming that better estimators inherently yield superior policies. Although theoretically justified, this estimator-centric approach neglects a critical practical obstacle: challenging optimization landscapes. In this paper, we provide theoretical insights and empirical evidence showing that current OPL methods encounter severe optimization issues, particularly as the action space grows. We show that estimator-aware policy parametrization can mitigate, but not fully resolve, optimization challenges. Building on this, we explore simpler weighted log-likelihood objectives and demonstrate that they enjoy substantially better optimization properties and still recover competitive, often superior, learned policies. Our findings emphasize the necessity of explicitly addressing optimization considerations in the development of OPL algorithms for large action spaces.

stat.ML↗

arXiv2026

Learning to Bid in Repeated Second-Price Auctions with Dynamic Values and Aggregated Feedback

We study the problem of learning to bid when the bidder's value is dynamic, i.e., when the current value depends on past outcomes. Specifically, we consider a bidder participating in repeated second-price auctions whose value depends on the time elapsed since their last successful bid, with auctions arriving in continuous time and only aggregated feedback revealed at the end of the horizon. Such a bidder must (1) balance the immediate benefit of winning the current auction against its impact on future values and (2) learn unknown environmental parameters. We derive regret bounds for a class of learning methods that combine plug-in estimators with a differential-equation characterization of the optimal policy, and show that a specific confidence bound algorithm learns the optimal policy with a near optimal regret of $\widetilde{O}(\log N)$ for piecewise linear primitives, and $\widetilde{O}(N^{1/3})$ for general, smooth primitives, achieving these regrets without explicit randomization. These theoretical results are supported by numerical experiments.

cs.LG↗

arXiv2026

Self-Consistency via Marginal Sharpening

Inference-time sampling can elicit strong reasoning abilities from language models without additional training. Existing power-sampling methods do so by sharpening the distribution over full generated outputs, favoring completions that are individually likely under the model. We argue that this is the wrong object to target for reasoning: a completion entangles a reasoning trace with a final answer, whereas what matters is whether an answer is supported by many plausible reasoning paths. We therefore shift the target from the full-output distribution to the sharpened answer marginal, making self-consistency an inference-time objective rather than a post-hoc voting criterion. Surprisingly, this marginal target admits an efficient approximation: we propose a simple, purely autoregressive parallel sampling algorithm that approximately samples from the sharpened answer marginal, eliciting stronger performance than standard power sampling on mathematics and coding benchmarks while being orders of magnitude faster.

cs.LG↗

arXiv2026

Off-Policy Learning to Reason Works Because It Is More Pessimistic Than You Think

Large scale reinforcement learning has become a central tool for improving reasoning in large language models. At this scale, generation is often lagged or asynchronous, so updates are performed on data collected by older policies. This makes learning inherently off-policy. Most existing approaches nevertheless remain rooted in PPO-style trust-region objectives, treating training as approximately on-policy and using importance weights to correct distribution mismatch. These corrections can introduce high variance, destabilize optimization, and accelerate entropy collapse. Recent work suggests an alternative: rather than correcting the mismatch, one can embrace off-policy data and remove importance weights, often yielding stronger algorithms. In this paper, we provide an intuitive construction of off-policy objectives that include successful off-policy objectives and show that their effectiveness can be understood through implicit pessimism: they optimize toward target policies that are more conservative than their nominal objectives suggest. This perspective explains why some particular implementation choices improve stability: they implicitly control the effective target distribution. We then propose a principled modification that stabilize this induced distribution and improve off-policy learning.

cs.LG↗

arXiv2026

Sequential Off-Policy Learning with Logarithmic Smoothing

Off-policy learning enables training policies from logged interaction data. Most prior work considers the batch setting, where a policy is learned from data generated by a single behavior policy. In real systems, however, policies are updated and redeployed repeatedly, each time training on all previously collected data while generating new interactions for future updates. This sequential off-policy learning setting is common in practice but remains largely unexplored theoretically. In this work, we present and study a simple algorithm for sequential off-policy learning, combining Logarithmic Smoothing (LS) estimation with online PAC-Bayesian tools. We further show that a principled adjustment to LS improves performance and accelerates convergence under mild conditions. The algorithms introduced generalise previous work: they match state-of-the-art offline approaches in the batch case and substantially outperform them when policies are updated sequentially. Empirical evaluations highlight both the benefits of the sequential framework and the strength of the proposed algorithms.

stat.ML↗

arXiv2026

RecoAtlas: From Semantic Plausibility to Set-Level Utility in LLM Recommendation Agents

LLM recommendation agents increasingly produce structured recommendation reports: sets of items accompanied by natural-language justifications. Yet existing evaluations often reduce this setting to reranking small shortlisted candidate sets or judge reports mainly by semantic plausibility. We introduce Recommendation Atlas (Agentic Tool-Level Assessment for Shopping), or RecoAtlas, a benchmark and toolkit for evaluating shopping agents with behavior-grounded metrics. RecoAtlas complements held-out interaction metrics with learned utility proxies for relevance, complementarity, and diversity derived from interaction data, while separately measuring semantic coherence and explanation quality. Its controlled tool environment exposes agents to either semantic, behavior-aligned, or faulty tools, enabling diagnosis of whether performance gains arise from stronger reasoning, better signals, or more effective tool-use policies. Across controlled experiments, we show that RecoAtlas exhibits key properties of a meaningful benchmark for agentic systems: performance scales with model capacity and test-time compute, improves with stronger and better-aligned tools, degrades under noisy or misaligned signals, and reveals that semantic plausibility does not necessarily capture behavior-grounded utility. RecoAtlas provides a foundation for developing and evaluating shopping assistants that optimize not only for plausible recommendations, but also for coherent, behaviorally grounded recommendation sets.

cs.IR↗

arXiv2025

RoiRL: Efficient, Self-Supervised Reasoning with Offline Iterative Reinforcement Learning

Reinforcement learning (RL) is central to improving reasoning in large language models (LLMs) but typically requires ground-truth rewards. Test-Time Reinforcement Learning (TTRL) removes this need by using majority-vote rewards, but relies on heavy online RL and incurs substantial computational cost. We propose RoiRL: Reasoning with offline iterative Reinforcement Learning, a family of lightweight offline learning alternatives that can target the same regularized optimal policies. Unlike TTRL, RoiRL eliminates the need to maintain a reference model and instead optimizes weighted log-likelihood objectives, enabling stable training with significantly lower memory and compute requirements. Experimental results show that RoiRL trains to 2.5x faster and consistently outperforms TTRL on reasoning benchmarks, establishing a scalable path to self-improving LLMs without labels.

cs.LG↗

arXiv2025

Non-Linear Counterfactual Aggregate Optimization

We consider the problem of directly optimizing a non-linear function of an outcome, where this outcome itself is the sum of many small contributions. The non-linearity of the function means that the problem is not equivalent to the maximization of the expectation of the individual contribution. By leveraging the concentration properties of the sum of individual outcomes, we derive a scalable descent algorithm that directly optimizes for our stated objective. This allows for instance to maximize the probability of successful A/B test, for which it can be wiser to target a success criterion, such as exceeding a given uplift, rather than chasing the highest expected payoff.

stat.ML↗

arXiv2025

Offline Contextual Bandit with Counterfactual Sample Identification

In production systems, contextual bandit approaches often rely on direct reward models that take both action and context as input. However, these models can suffer from confounding, making it difficult to isolate the effect of the action from that of the context. We present \emph{Counterfactual Sample Identification}, a new approach that re-frames the problem: rather than predicting reward, it learns to recognize which action led to a successful (binary) outcome by comparing it to a counterfactual action sampled from the logging policy under the same context. The method is theoretically grounded and consistently outperforms direct models in both synthetic experiments and real-world deployments.

cs.LG↗

arXiv2024

Logarithmic Smoothing for Pessimistic Off-Policy Evaluation, Selection and Learning

This work investigates the offline formulation of the contextual bandit problem, where the goal is to leverage past interactions collected under a behavior policy to evaluate, select, and learn new, potentially better-performing, policies. Motivated by critical applications, we move beyond point estimators. Instead, we adopt the principle of pessimism where we construct upper bounds that assess a policy's worst-case performance, enabling us to confidently select and learn improved policies. Precisely, we introduce novel, fully empirical concentration bounds for a broad class of importance weighting risk estimators. These bounds are general enough to cover most existing estimators and pave the way for the development of new ones. In particular, our pursuit of the tightest bound within this class motivates a novel estimator (LS), that logarithmically smooths large importance weights. The bound for LS is provably tighter than its competitors, and naturally results in improved policy selection and learning strategies. Extensive policy evaluation, selection, and learning experiments highlight the versatility and favorable performance of LS.

stat.ML↗

arXiv2024

Probabilistic Rank and Reward: A Scalable Model for Slate Recommendation

We introduce Probabilistic Rank and Reward (PRR), a scalable probabilistic model for personalized slate recommendation. Our approach allows off-policy estimation of the reward in the scenario where the user interacts with at most one item from a slate of K items. We show that the probability of a slate being successful can be learned efficiently by combining the reward, whether the user successfully interacted with the slate, and the rank, the item that was selected within the slate. PRR outperforms existing off-policy reward optimizing methods and is far more scalable to large action spaces. Moreover, PRR allows fast delivery of recommendations powered by maximum inner product search (MIPS), making it suitable in low latency domains such as computational advertising.

cs.IR↗

arXiv2023

Fast Slate Policy Optimization: Going Beyond Plackett-Luce

An increasingly important building block of large scale machine learning systems is based on returning slates; an ordered lists of items given a query. Applications of this technology include: search, information retrieval and recommender systems. When the action space is large, decision systems are restricted to a particular structure to complete online queries quickly. This paper addresses the optimization of these large scale decision systems given an arbitrary reward function. We cast this learning problem in a policy optimization framework and propose a new class of policies, born from a novel relaxation of decision functions. This results in a simple, yet efficient learning algorithm that scales to massive action spaces. We compare our method to the commonly adopted Plackett-Luce policy class and demonstrate the effectiveness of our approach on problems with action space sizes in the order of millions.

cs.LG↗

arXiv2023

Fast Offline Policy Optimization for Large Scale Recommendation

Personalised interactive systems such as recommender systems require selecting relevant items from massive catalogs dependent on context. Reward-driven offline optimisation of these systems can be achieved by a relaxation of the discrete problem resulting in policy learning or REINFORCE style learning algorithms. Unfortunately, this relaxation step requires computing a sum over the entire catalogue making the complexity of the evaluation of the gradient (and hence each stochastic gradient descent iterations) linear in the catalogue size. This calculation is untenable in many real world examples such as large catalogue recommender systems, severely limiting the usefulness of this method in practice. In this paper, we derive an approximation of these policy learning algorithms that scale logarithmically with the catalogue size. Our contribution is based upon combining three novel ideas: a new Monte Carlo estimate of the gradient of a policy, the self normalised importance sampling estimator and the use of fast maximum inner product search at training time. Extensive experiments show that our algorithm is an order of magnitude faster than naive approaches yet produces equally good policies.

cs.IR↗

arXiv2023

PAC-Bayesian Offline Contextual Bandits With Guarantees

This paper introduces a new principled approach for off-policy learning in contextual bandits. Unlike previous work, our approach does not derive learning principles from intractable or loose bounds. We analyse the problem through the PAC-Bayesian lens, interpreting policies as mixtures of decision rules. This allows us to propose novel generalization bounds and provide tractable algorithms to optimize them. We prove that the derived bounds are tighter than their competitors, and can be optimized directly to confidently improve upon the logging policy offline. Our approach learns policies with guarantees, uses all available data and does not require tuning additional hyperparameters on held-out sets. We demonstrate through extensive experiments the effectiveness of our approach in providing performance guarantees in practical scenarios.

stat.ML↗

arXiv2022

Offline Evaluation of Reward-Optimizing Recommender Systems: The Case of Simulation

Both in academic and industry-based research, online evaluation methods are seen as the golden standard for interactive applications like recommendation systems. Naturally, the reason for this is that we can directly measure utility metrics that rely on interventions, being the recommendations that are being shown to users. Nevertheless, online evaluation methods are costly for a number of reasons, and a clear need remains for reliable offline evaluation procedures. In industry, offline metrics are often used as a first-line evaluation to generate promising candidate models to evaluate online. In academic work, limited access to online systems makes offline metrics the de facto approach to validating novel methods. Two classes of offline metrics exist: proxy-based methods, and counterfactual methods. The first class is often poorly correlated with the online metrics we care about, and the latter class only provides theoretical guarantees under assumptions that cannot be fulfilled in real-world environments. Here, we make the case that simulation-based comparisons provide ways forward beyond offline metrics, and argue that they are a preferable means of evaluation.

cs.IR↗

arXiv2020

Improving Offline Contextual Bandits with Distributional Robustness

This paper extends the Distributionally Robust Optimization (DRO) approach for offline contextual bandits. Specifically, we leverage this framework to introduce a convex reformulation of the Counterfactual Risk Minimization principle. Besides relying on convex programs, our approach is compatible with stochastic optimization, and can therefore be readily adapted tothe large data regime. Our approach relies on the construction of asymptotic confidence intervals for offline contextual bandits through the DRO framework. By leveraging known asymptotic results of robust estimators, we also show how to automatically calibrate such confidence intervals, which in turn removes the burden of hyper-parameter selection for policy optimization. We present preliminary empirical results supporting the effectiveness of our approach.

stat.ML↗