Jean-Michel Renders

Romain Deffayet, Thibaut Thonet, Jean-Michel Renders et al.

Recent research has employed reinforcement learning (RL) algorithms to optimize long-term user engagement in recommender systems, thereby avoiding common pitfalls such as user boredom and filter bubbles. They capture the sequential and interactive nature of recommendations, and thus offer a principled way to deal with long-term rewards and avoid myopic behaviors. However, RL approaches are intractable in the slate recommendation scenario - where a list of items is recommended at each interaction turn - due to the combinatorial action space. In that setting, an action corresponds to a slate that may contain any combination of items. While previous work has proposed well-chosen decompositions of actions so as to ensure tractability, these rely on restrictive and sometimes unrealistic assumptions. Instead, in this work we propose to encode slates in a continuous, low-dimensional latent space learned by a variational auto-encoder. Then, the RL agent selects continuous actions in this latent space, which are ultimately decoded into the corresponding slates. By doing so, we are able to (i) relax assumptions required by previous work, and (ii) improve the quality of the action selection by modeling full slates instead of independent items, in particular by enabling diversity. Our experiments performed on a wide array of simulated environments confirm the effectiveness of our generative modeling of slates over baselines in practical scenarios where the restrictive assumptions underlying the baselines are lifted. Our findings suggest that representation learning using generative models is a promising direction towards generalizable RL-based slate recommendation.

Seungsu Kim, Jinyoung Choi, Seungmin Baek et al.

We present ProgVLA, a compact vision-language-action (VLA) model designed for reliable robot manipulation under tight compute and memory budgets. The model specifically focuses on efficiently processing long multi-modal sequences by maintaining an explicit representation of task progress over extended horizons. To this end, ProgVLA integrates two key components. First, a multi-modal encoder with a two-stage Perceiver resampling scheme compresses variable-length visual, language, and proprioceptive streams into a fixed set of control-ready context tokens, substantially reducing sequence length while preserving cross-modal grounding. Second, an auxiliary set of progress heads is trained with offline reinforcement learning (RL) objectives to jointly learn critics over normalized remaining-horizon targets. This provides the policy with an internal estimate of task progress and enables advantage- and success-weighted flow-matching imitation learning. On two well-established multi-task robot manipulation benchmarks, a 0.1B-parameter ProgVLA model reaches success rates that are competitive with, and on long-horizon and harder task tiers exceed, substantially larger pretrained baselines. Ablations indicate that the learned context resampler and task-adaptive visual fine-tuning are the largest single contributors, while progress-aware training provides a consistent additional gain that is concentrated on long-horizon and multi-object tasks. We further validate the approach in real-world toy-kitchen environments.

Alberta Longhini, David Emukpere, Jean-Michel Renders et al.

We address the problem of fine-tuning pre-trained generative policies with reinforcement learning (RL) while preserving the multimodality of their action distributions. Existing methods for RL fine-tuning of generative policies (e.g., diffusion policies) improve task performance but often collapse diverse behaviors into a single reward-maximizing mode. To mitigate this issue, we propose an unsupervised mode discovery framework that uncovers latent behavioral modes within generative policies. The discovered modes enable the use of mutual information as an intrinsic reward, regularizing RL fine-tuning to enhance task success while maintaining behavioral diversity. Experiments on robotic manipulation tasks demonstrate that our method consistently outperforms conventional fine-tuning approaches, achieving higher success rates and preserving richer multimodal action distributions.

Philipp Hager, Romain Deffayet, Jean-Michel Renders et al.

Unbiased learning-to-rank (ULTR) is a well-established framework for learning from user clicks, which are often biased by the ranker collecting the data. While theoretically justified and extensively tested in simulation, ULTR techniques lack empirical validation, especially on modern search engines. The Baidu-ULTR dataset released for the WSDM Cup 2023, collected from Baidu's search engine, offers a rare opportunity to assess the real-world performance of prominent ULTR techniques. Despite multiple submissions during the WSDM Cup 2023 and the subsequent NTCIR ULTRE-2 task, it remains unclear whether the observed improvements stem from applying ULTR or other learning techniques. In this work, we revisit and extend the available experiments on the Baidu-ULTR dataset. We find that standard unbiased learning-to-rank techniques robustly improve click predictions but struggle to consistently improve ranking performance, especially considering the stark differences obtained by choice of ranking loss and query-document features. Our experiments reveal that gains in click prediction do not necessarily translate to enhanced ranking performance on expert relevance annotations, implying that conclusions strongly depend on how success is measured in this benchmark.

David Emukpere, Bingbing Wu, Julien Perez et al.

Self-supervised skill learning aims to acquire useful behaviors that leverage the underlying dynamics of the environment. Latent variable models, based on mutual information maximization, have been successful in this task but still struggle in the context of robotic manipulation. As it requires impacting a possibly large set of degrees of freedom composing the environment, mutual information maximization fails alone in producing useful and safe manipulation behaviors. Furthermore, tackling this by augmenting skill discovery rewards with additional rewards through a naive combination might fail to produce desired behaviors. To address this limitation, we introduce SLIM, a multi-critic learning approach for skill discovery with a particular focus on robotic manipulation. Our main insight is that utilizing multiple critics in an actor-critic framework to gracefully combine multiple reward functions leads to a significant improvement in latent-variable skill discovery for robotic manipulation while overcoming possible interference occurring among rewards which hinders convergence to useful skills. Furthermore, in the context of tabletop manipulation, we demonstrate the applicability of our novel skill discovery approach to acquire safe and efficient motor primitives in a hierarchical reinforcement learning fashion and leverage them through planning, significantly surpassing baseline approaches for skill discovery.

David Emukpere, Romain Deffayet, Bingbing Wu et al.

Learning robotic manipulation skills from vision is a promising approach for developing robotics applications that can generalize broadly to real-world scenarios. As such, many approaches to enable this vision have been explored with fruitful results. Particularly, object-centric representation methods have been shown to provide better inductive biases for skill learning, leading to improved performance and generalization. Nonetheless, we show that object-centric methods can struggle to learn simple manipulation skills in multi-object environments. Thus, we propose DOCIR, an object-centric framework that introduces a disentangled representation for objects of interest, obstacles, and robot embodiment. We show that this approach leads to state-of-the-art performance for learning pick and place skills from visual inputs in multi-object environments and generalizes at test time to changing objects of interest and distractors in the scene. Furthermore, we show its efficacy both in simulation and zero-shot transfer to the real world.

Sami Jullien, Romain Deffayet, Jean-Michel Renders et al.

Distributional reinforcement learning (RL) has proven useful in multiple benchmarks as it enables approximating the full distribution of returns and extracts rich feedback from environment samples. The commonly used quantile regression approach to distributional RL -- based on asymmetric $L_1$ losses -- provides a flexible and effective way of learning arbitrary return distributions. In practice, it is often improved by using a more efficient, asymmetric hybrid $L_1$-$L_2$ Huber loss for quantile regression. However, by doing so, distributional estimation guarantees vanish, and we empirically observe that the estimated distribution rapidly collapses to its mean. Indeed, asymmetric $L_2$ losses, corresponding to expectile regression, cannot be readily used for distributional temporal difference. Motivated by the efficiency of $L_2$-based learning, we propose to jointly learn expectiles and quantiles of the return distribution in a way that allows efficient learning while keeping an estimate of the full distribution of returns. We prove that our proposed operator converges to the distributional Bellman operator in the limit of infinite estimated quantile and expectile fractions, and we benchmark a practical implementation on a toy example and at scale. On the Atari benchmark, our approach matches the performance of the Huber-based IQN-1 baseline after $200$M training frames but avoids distributional collapse and keeps estimates of the full distribution of returns.

Till Kletti, Jean-Michel Renders, Patrick Loiseau

We consider the problem of computing a sequence of rankings that maximizes consumer-side utility while minimizing producer-side individual unfairness of exposure. While prior work has addressed this problem using linear or quadratic programs on bistochastic matrices, such approaches, relying on Birkhoff-von Neumann (BvN) decompositions, are too slow to be implemented at large scale. In this paper we introduce a geometrical object, a polytope that we call expohedron, whose points represent all achievable exposures of items for a Position Based Model (PBM). We exhibit some of its properties and lay out a Carathéodory decomposition algorithm with complexity $O(n^2\log(n))$ able to express any point inside the expohedron as a convex sum of at most $n$ vertices, where $n$ is the number of items to rank. Such a decomposition makes it possible to express any feasible target exposure as a distribution over at most $n$ rankings. Furthermore we show that we can use this polytope to recover the whole Pareto frontier of the multi-objective fairness-utility optimization problem, using a simple geometrical procedure with complexity $O(n^2\log(n))$. Our approach compares favorably to linear or quadratic programming baselines in terms of algorithmic complexity and empirical runtime and is applicable to any merit that is a non-decreasing function of item relevance. Furthermore our solution can be expressed as a distribution over only $n$ permutations, instead of the $(n-1)^2 + 1$ achieved with BvN decompositions. We perform experiments on synthetic and real-world datasets, confirming our theoretical results.

Thibaut Thonet, Yagmur Gizem Cinar, Eric Gaussier et al.

Information retrieval (IR) systems traditionally aim to maximize metrics built on rankings, such as precision or NDCG. However, the non-differentiability of the ranking operation prevents direct optimization of such metrics in state-of-the-art neural IR models, which rely entirely on the ability to compute meaningful gradients. To address this shortcoming, we propose SmoothI, a smooth approximation of rank indicators that serves as a basic building block to devise differentiable approximations of IR metrics. We further provide theoretical guarantees on SmoothI and derived approximations, showing in particular that the approximation errors decrease exponentially with an inverse temperature-like hyperparameter that controls the quality of the approximations. Extensive experiments conducted on four standard learning-to-rank datasets validate the efficacy of the listwise losses based on SmoothI, in comparison to previously proposed ones. Additional experiments with a vanilla BERT ranking model on a text-based IR task also confirm the benefits of our listwise approach.

Pedro Chahuara, Nicolas Grislain, Grégoire Jauvion et al.

This paper describes an engine to optimize web publisher revenues from second-price auctions. These auctions are widely used to sell online ad spaces in a mechanism called real-time bidding (RTB). Optimization within these auctions is crucial for web publishers, because setting appropriate reserve prices can significantly increase revenue. We consider a practical real-world setting where the only available information before an auction occurs consists of a user identifier and an ad placement identifier. The real-world challenges we had to tackle consist mainly of tracking the dependencies on both the user and placement in an highly non-stationary environment and of dealing with censored bid observations. These challenges led us to make the following design choices: (i) we adopted a relatively simple non-parametric regression model of auction revenue based on an incremental time-weighted matrix factorization which implicitly builds adaptive users' and placements' profiles; (ii) we jointly used a non-parametric model to estimate the first and second bids' distribution when they are censored, based on an on-line extension of the Aalen's Additive model. Our engine is a component of a deployed system handling hundreds of web publishers across the world, serving billions of ads a day to hundreds of millions of visitors. The engine is able to predict, for each auction, an optimal reserve price in approximately one millisecond and yields a significant revenue increase for the web publishers.

Vaishali Pal, Fabien Guillot, Manish Shrivastava et al.

Spoken dialogue systems typically use a list of top-N ASR hypotheses for inferring the semantic meaning and tracking the state of the dialogue. However ASR graphs, such as confusion networks (confnets), provide a compact representation of a richer hypothesis space than a top-N ASR list. In this paper, we study the benefits of using confusion networks with a state-of-the-art neural dialogue state tracker (DST). We encode the 2-dimensional confnet into a 1-dimensional sequence of embeddings using an attentional confusion network encoder which can be used with any DST system. Our confnet encoder is plugged into the state-of-the-art 'Global-locally Self-Attentive Dialogue State Tacker' (GLAD) model for DST and obtains significant improvements in both accuracy and inference time compared to using top-N ASR hypotheses.

Jean-Michel Renders

We consider the problem of Active Search, where a maximum of relevant objects - ideally all relevant objects - should be retrieved with the minimum effort or minimum time. Typically, there are two main challenges to face when tackling this problem: first, the class of relevant objects has often low prevalence and, secondly, this class can be multi-faceted or multi-modal: objects could be relevant for completely different reasons. To solve this problem and its associated issues, we propose an approach based on a non-stationary (aka restless) extension of Thompson Sampling, a well-known strategy for Multi-Armed Bandits problems. The collection is first soft-clustered into a finite set of components and a posterior distribution of getting a relevant object inside each cluster is updated after receiving the user feedback about the proposed instances. The "next instance" selection strategy is a mixed, two-level decision process, where both the soft clusters and their instances are considered. This method can be considered as an insurance, where the cost of the insurance is an extra exploration effort in the short run, for achieving a nearly "total" recall with less efforts in the long run.

Yuxin Chen, Jean-Michel Renders, Morteza Haghir Chehreghani et al.

We consider the optimal value of information (VoI) problem, where the goal is to sequentially select a set of tests with a minimal cost, so that one can efficiently make the best decision based on the observed outcomes. Existing algorithms are either heuristics with no guarantees, or scale poorly (with exponential run time in terms of the number of available tests). Moreover, these methods assume a known distribution over the test outcomes, which is often not the case in practice. We propose an efficient sampling-based online learning framework to address the above issues. First, assuming the distribution over hypotheses is known, we propose a dynamic hypothesis enumeration strategy, which allows efficient information gathering with strong theoretical guarantees. We show that with sufficient amount of samples, one can identify a near-optimal decision with high probability. Second, when the parameters of the hypotheses distribution are unknown, we propose an algorithm which learns the parameters progressively via posterior sampling in an online fashion. We further establish a rigorous bound on the expected regret. We demonstrate the effectiveness of our approach on a real-world interactive troubleshooting application and show that one can efficiently make high-quality decisions with low cost.

Matthias Galle, Jean-Michel Renders, Guillaume Jacquet

Clustering web documents has numerous applications, such as aggregating news articles into meaningful events, detecting trends and hot topics on the Web, preserving diversity in search results, etc. At the same time, the importance of named entities and, in particular, the ability to recognize them and to solve the associated co-reference resolution problem are widely recognized as key enabling factors when mining, aggregating and comparing content on the Web. Instead of considering these two problems separately, we propose in this paper a method that tackles jointly the problem of clustering news articles into events and cross-document co-reference resolution of named entities. The co-occurrence of named entities in the same clusters is used as an additional signal to decide whether two referents should be merged into one entity. These refined entities can in turn be used as enhanced features to re-cluster the documents and then be refined again, entering into a virtuous cycle that improves simultaneously the performances of both tasks. We implemented a prototype system and report results using the TDT5 collection of news articles, demonstrating the potential of our approach.

Phong Le, Marc Dymetman, Jean-Michel Renders

We introduce an LSTM-based method for dynamically integrating several word-prediction experts to obtain a conditional language model which can be good simultaneously at several subtasks. We illustrate this general approach with an application to dialogue where we integrate a neural chat model, good at conversational aspects, with a neural question-answering model, good at retrieving precise information from a knowledge-base, and show how the integration combines the strengths of the independent components. We hope that this focused contribution will attract attention on the benefits of using such mixtures of experts in NLP.