论文检索

Deep Learning · Large Language Models

Hyunseo Kim, Sangam Lee, Kwangwook Seo, Dongha Lee

Search-augmented large language models (LLMs) remain insufficient for fully addressing diverse user needs, which requires recognizing how the same query can reflect different intents across users and delivering information in preferred forms. While recent systems such as ChatGPT and Gemini attempt personalization by leveraging user histories, systematic evaluation of such personalization is under-explored. To address this gap, we propose BESPOKE, the realistic benchmark for evaluating personalization in search-augmented LLMs. BESPOKE is designed to be both realistic, by collecting authentic chat and search histories directly from humans, and diagnostic, by pairing responses with fine-grained preference scores and feedback. The benchmark is constructed through long-term, deeply engaged human annotation, where human annotators contributed their own histories, authored queries with detailed information needs, and evaluated responses with scores and diagnostic feedback. Leveraging BESPOKE, we conduct systematic analyses that reveal key requirements for effective personalization in information-seeking tasks, providing a foundation for fine-grained evaluation of personalized search-augmented LLMs. Our code and data are available at https://anonymous.4open.science/r/bespoke-E82B.

Deep Learning · Generative Models and Autoencoders

Satoshi Hayakawa, Yuhta Takida, Masaaki Imaizumi, Hiromi Wakaki, Yuki Mitsufuji

Masked diffusion models have shown promising performance in generating high-quality samples in a wide range of domains, but accelerating their sampling process remains relatively underexplored. To investigate efficient samplers for masked diffusion, this paper theoretically analyzes the MaskGIT sampler for image modeling, revealing its implicit temperature sampling mechanism. Through this analysis, we show that MaskGIT is asymptotically equivalent to a choose-then-sample (CTS) formulation, instantiated as the “moment sampler,” which explicitly separates index selection from token sampling. This CTS reformulation is essential: it yields unbiased token sampling and exposes an algorithmic design space for index selection, both of which are inaccessible in MaskGIT’s original formulation. Regarding token sampling, we reveal that MaskGIT implicitly adopts a low-temperature sampler, which explains why MaskGIT often degrades with more sampling steps. The CTS reformulation of MaskGIT allows to fix the temperature sampling to ensure unbiasedness. We also improve the index selection in CTS through two key innovations: a partial caching technique for transformers that approximates longer sampling trajectories without proportional computational cost, and a hybrid approach formalizing the exploration-exploitation trade-off in adaptive unmasking. Experiments in image and text domains demonstrate our theory as well as the efficiency of our proposed methods, advancing both theoretical understanding and practical implementation of masked diffusion samplers.

Reinforcement Learning · Deep RL

Ayoub Belouadah, Sylvain Kubler, YVES LE TRAON

Safe reinforcement learning (Safe RL) aims to maximize expected return while satisfying safety constraints, typically modeled as constrained Markov decision processes. While primal-dual methods scale well to deep RL, they often suffer from delayed constraint correction, leading to oscillatory behavior and prolonged safety violations. In this paper, we propose *Constraint-Sensitive Policy Optimization (CSPO)*, a first-order primal-dual method that incorporates local constraint sensitivity into policy updates. CSPO augments the primal objective with a constraint-sensitive correction derived from the shortest signed distance to the safety boundary, enabling smarter recovery steps back to safety, compensating for delayed Lagrange multiplier updates, and reducing oscillations near the boundary, while preserving the KKT solutions of the original constrained problem. Extensive experiments on navigation and locomotion benchmarks demonstrate that CSPO achieves faster safety recovery and high reward preservation, resulting in higher constrained returns (+15.6\% average improvement) compared to state-of-the-art primal-dual and penalty-based methods.

Deep Learning · Sequential Models, Time series

Egor Serov, Ilya Kuleshov, Alexey Zaytsev

Neural Controlled Differential Equations (Neural CDEs) provide a powerful continuous-time framework for sequence modeling, yet the roughness of the driving control path often restricts their efficiency. Standard splines introduce high-frequency variations that force adaptive solvers to take excessively small steps, driving up the Number of Function Evaluations (NFE). We propose a novel approach to Neural CDE path construction that replaces exact interpolation with Kernel and Gaussian Process (GP) smoothing, enabling explicit control over trajectory regularity. To recover details lost during smoothing, we propose an attention-based Multi-View CDE (MV-CDE) and its convolutional extension (MVC-CDE), which employ learnable queries to inform path reconstruction. This framework allows the model to distribute representational capacity across multiple trajectories, each capturing distinct temporal patterns. Empirical results demonstrate that our method, MVC-CDE with GP, achieves state-of-the-art accuracy while significantly reducing NFEs and total inference time compared to spline-based baselines.

Reinforcement Learning · Deep RL

Zhouyang Yu, Guojian Zhan, Yang Guan, Jingliang Duan, Letian Tao, Shengbo Li

Off-policy reinforcement learning is vulnerable to overestimation bias, which is rooted in the total value uncertainty. However, existing methods typically misaddress this by targeting the epistemic component, neglecting the aleatoric component. We identify for the first time that this oversight fails to contain a massive bias surge, termed the **Aleatoric Impulse**. Although transient, this impulse fundamentally derails the learning trajectory, permanently locking the agent into suboptimal policies. To counteract this, we propose **A**leatoric **I**mpulse **D**amping **(AID)**, the first mechanism that models total value uncertainty by disentangling the return variance into epistemic and aleatoric components, followed by their adaptive weighted recombination. Leveraging this derived uncertainty, the critic constructs a pessimistic lower confidence bound to surgically suppress the impulse. Complementing this, the actor utilizes a symmetrical upper confidence bound to drive optimistic exploration, ensuring that the necessary pessimism does not compromise exploration efficiency. We integrate this mechanism into the Distributional Soft Actor-Critic algorithm to establish **DSAC-AID**. Extensive experiments on the high-dimensional Gym-MuJoCo and DeepMind Control Suite benchmarks demonstrate that it achieves state-of-the-art results in final performance.

Theory · Online Learning and Bandits

Yuming Shao, Zhixuan Fang

Linear bandits traditionally prioritize regret minimization, often overlooking statistical inference of the underlying parameter as a critical objective. In high-stakes settings such as healthcare, precise parameter estimation is indispensable, as it provides fundamental insights into system mechanisms and ensures robust decision-making under covariate shift. We investigate the tripartite balance between regret, inference, and safety, deriving a fundamental minimax lower bound that characterizes the Pareto-optimal frontier of these competing goals. We then propose SERMiSC, a novel algorithm that achieves the optimal trade-off by matching this lower bound while maintaining a near-constant $\tilde{O}(1)$ safety risk. Empirical results demonstrate that SERMiSC effectively navigates the Pareto frontier and outperforms various baselines, thereby validating our theoretical analysis.

Reinforcement Learning · Deep RL

Bernardo Ávila Pires, Mark Rowland, Diana Borsa, Zhaohan Guo, Khimya Khetarpal, Andre Barreto, David Abel, R{{\'e}}mi Munos, Will Dabney

We introduce distributional dynamic programming (DP) methods for optimizing statistical functionals of the return distribution, with standard reinforcement learning as a special case. Previous distributional DP methods could optimize the same class of expected utilities as classic DP. To go beyond, we combine distributional DP with stock augmentation, a technique previously introduced for classic DP in the context of risk-sensitive RL, where the MDP state is augmented with a statistic of the rewards obtained since the first time step. We find that a number of recently studied problems can be formulated as stock-augmented return distribution optimization, and we show that we can use distributional DP to solve them. We analyze distributional value and policy iteration, with bounds and a study of what objectives these distributional DP methods can or cannot optimize. We describe a number of applications outlining how to use distributional DP to solve different stock-augmented return distribution optimization problems, for example maximizing conditional value-at-risk, and homeostatic regulation. To highlight the practical potential of stock-augmented return distribution optimization and distributional DP, we introduce an agent that combines DQN and the core ideas of distributional DP, and empirically evaluate it for solving instances of the applications discussed.

Social Aspects · Accountability, Transparency, and Interpretability

Praneet Suresh, Jack Stanley, Sonia Joseph, Luca Scimeca, Danilo Bzdok

Pre-trained transformers have demonstrated remarkable generalization abilities, at times extending beyond the scope of their training data. Yet, real-world deployments often face unexpected or adversarial data that diverges from training data distributions. Without explicit mechanisms for handling such shifts, model reliability and safety degrade, urging more disciplined study of out-of-distribution (OOD) settings for transformers. By systematic experiments, we present a mechanistic framework for delineating the precise contours of transformer model robustness. We find that OOD inputs, including subtle typos and jailbreak prompts, drive language models to operate on an increased number of fallacious concepts in their internals. We leverage this device to quantify and understand the degree of distributional shift in prompts, enabling a mechanistically grounded fine-tuning strategy to robustify LLMs. Expanding the very notion of OOD from input data to a model’s private computational processes—a new transformer diagnostic at inference time—is a critical step toward making AI systems safe for deployment across science, business, and government.

Sonia Joseph, Quentin Garrido, Randall Balestriero, Matthew Kowal, Thomas Fel, Shahab Bakhtiari, Blake Richards, Michael Rabbat

A long-standing question in physical reasoning is whether video-based models need to rely on factorized representations of physical variables in order to make physically accurate predictions, or whether they can implicitly represent such variables in a distributed manner. While modern video world models achieve strong performance on intuitive physics benchmarks, it remains unclear which of these representational regimes they implement internally. Here, we present the first interpretability study to directly examine physical representations inside large-scale video encoders. Using layerwise probing, subspace geometry, patch-level decoding, and targeted attention ablations, we characterize where physical information becomes accessible and how it is organized within encoder-based video transformers. Across architectures, we identify a sharp intermediate-depth transition— which we call the \emph{Physics Emergence Zone}—at which physical variables become accessible. Physics-related representations peak shortly after this transition and degrade toward the output layers. Decomposing motion into explicit variables, we find that scalar quantities such as speed and acceleration are available from early layers onwards, whereas motion direction becomes accessible only at the Physics Emergence Zone. Notably, we find that direction is encoded through a high-dimensional population structure with circular geometry, requiring coordinated multi-feature intervention to control. These findings suggest that modern video models do not use factorized representations of physical variables like a classical physics engine. Instead, they use a distributed representation that is nonetheless sufficient for making physical predictions.

Optimization · Stochastic

Stefano Bruno, Youngsik Hwang, JaeHyeon An, Sotirios Sabanis, Dongyoung Lim

Flatness of the loss landscape has been widely studied as an important perspective for understanding the behavior and generalization of deep learning algorithms. Motivated by this view, we propose Flatness-Aware Stochastic Gradient Langevin Dynamics (fSGLD), a first-order optimization method that biases learning its dynamics toward flat basins while retaining the computational and memory efficiency of SGD and SGLD. We provide a non-asymptotic theoretical analysis showing that fSGLD converges to a flatness-biased Gibbs distribution under a theoretically prescribed coupling between the noise scale $\sigma$ and the inverse temperature $\beta$, together with explicit excess risk guarantees. We empirically evaluate fSGLD across standard optimizer benchmarks, Bayesian image classification, uncertainty quantification, and out-of-distribution detection, demonstrating consistently strong performance and reliable uncertainty estimates. Additional experiments confirm the effectiveness of the theoretically prescribed $\beta$–$\sigma$ coupling compared to decoupled choices.

Deep Learning · Algorithms

Naili Xing, Shaofeng Cai, Lingze Zeng, Jiaqi Zhu, Peng LU, Jian Pei, Beng Chin Ooi

Recent advances have shifted the paradigm of tabular learning toward tabular foundation models, yet their accuracy relies on a heavy inference cost that scales poorly with context size. Deep neural networks remain a highly competitive and more efficient modeling paradigm when equipped with well-designed architectures; however, identifying such architectures in a data-adaptive and budget-aware manner remains challenging. We propose pTNAS, the first progressive neural architecture search (NAS) approach tailored for tabular data, to enable fast identification of a viable architecture and continuously improve its search performance as more budget becomes available. pTNAS adopts a filter-and-refine optimization strategy that combines both efficient training-free and effective training-based architecture evaluation. At the filtering phase, we introduce pTProxy, a novel zero-cost proxy specifically designed for tabular networks that jointly captures architectural trainability and expressivity, to facilitate fast filtering of large architecture search spaces. At the refinement phase, pTNAS employs a fixed-budget scheduling algorithm to accurately identify the best-performing architecture from a small set of promising candidates. We further propose a budget-aware coordinator to optimize budget allocation holistically. Experiments show that pTNAS reduces the time to reach the globally best architecture by up to 82.75 X compared with other NAS approaches, and improves average predictive accuracy and end-to-end efficiency by up to 4.95 X compared with TabPFN.

Theory · Deep Learning

Naoki Nishikawa, Taiji Suzuki

Inference-time alignment, the approach of adapting pre-trained models to reward feedback during inference, has proven empirically effective at improving language-model performance. Despite its success, theoretical foundations remain underdeveloped, especially in practical settings where neural networks are employed as reward models. In this paper, we explore the advantages of neural networks and how to effectively train them for inference-time alignment. Assuming that the true reward function lies in Besov spaces to capture the non-uniform smoothness, we compare neural networks to linear estimators and show that feature learning capability of neural networks is crucial for improving performance. We further analyze algorithms for training neural-network reward estimators. Specifically, we consider a multi-step algorithm that alternates between sampling from the current policy and refitting the reward estimator, and prove that it improves the regret, especially when the true reward exhibits local structure.

Reinforcement Learning · Batch/Offline

Mathieu Petitbois, Rémy Portelas, sylvain lamprier

We study offline reinforcement learning of style-conditioned policies using explicit style supervision via subtrajectory labeling functions. In this setting, aligning style with high task performance is particularly challenging due to distribution shift and inherent conflicts between style and reward. Existing methods, despite introducing numerous definitions of style, often fail to reconcile these objectives effectively. To address these challenges, we propose a unified definition of behavior style and instantiate it into a practical framework. Building on this, we introduce Style-Conditioned Implicit Q-Learning (SCIQL), which leverages offline goal-conditioned reinforcement learning techniques, such as hindsight relabeling and value learning, and combine it with a new Gated Advantage Weighted Regression mechanism to efficiently optimize task performance while preserving style alignment. Experiments demonstrate that SCIQL achieves superior performance on both objectives compared to prior offline methods.

Deep Learning · Large Language Models

Jinwoo Kim, Taylor Berg-Kirkpatrick, Loris D'Antoni

Diffusion language models offer a promising alternative to autoregressive models due to their global, non-causal generation process, but their continuous latent dynamics make discrete constraints---e.g., the output should be a JSON file that matches a given schema---difficult to impose. We introduce a training-free guidance method for steering continuous diffusion language models to satisfy formal syntactic constraints expressed using regular expressions. Our approach constructs an analytic score estimating the probability that a latent state decodes to a valid string accepted by a given regular expression, and uses its gradient to guide sampling, _without_ training auxiliary classifiers. The denoising process targets the base model conditioned on syntactic validity. We implement our method in Diffinity on top of the PLAID diffusion model and evaluate it on 180 regular-expression constraints over JSON and natural-language benchmarks. Diffinity achieves 68-96\% constraint satisfaction while incurring only a small perplexity cost relative to unconstrained sampling, outperforming autoregressive constrained decoding in both constraint satisfaction and output quality.

Reinforcement Learning · Batch/Offline

Meng XU, Zhongying Chen, Weiwei Fu, Yan Li, Shuguang Wang, Jianping Wang

Offline Federated Deep Reinforcement Learning (FDRL) methods aggregate multiple client-side offline Deep Reinforcement Learning (DRL) models, each trained locally, to facilitate knowledge sharing while preserving privacy. Existing offline FDRL methods assign client weights during global aggregation using either simple averaging or Q-values, but they neglect the combined consideration of Q-values and policy inconsistency, the latter of which reflects the distributional discrepancy between the learned policy and the policy from offline data. This causes clients with no significant advantages in one aspect but obvious disadvantages in the other to disproportionately affect the global model, thereby degrading its capabilities in that aspect. During local training, clients in existing methods are compelled to fully adopt the global model, which negatively impacts clients when the global model is weak. To address these limitations, we propose a novel Federated Learning (FL) framework that can be seamlessly integrated into current offline FDRL approaches to improve their performance. Our method considers both policy inconsistency and Q-values to determine the weights of client models, with the latter adjusted by a scaling factor to avoid significant numerical discrepancies with the former. The aggregated global model is then distributed to clients to facilitate their learning from the global model. The impact of the global model on the local models is reduced when a client's model performance exceeds that of the global model, thereby mitigating the influence of a weaker global model. Experiments on the Datasets for Deep Data-Driven Reinforcement Learning (D4RL) demonstrate that our method enhances six state-of-the-art (SOTA) offline FDRL methods in terms of return and D4RL score.

Deep Learning · Large Language Models

Qi Liu, Xinhao Zheng, Renqiu Xia, Xingzhi Qi, Qinxiang Cao, Junchi Yan

Large language models (LLMs) have achieved remarkable progress in mathematical reasoning, yet persistently suffer from hallucinations and erroneous logic. While formal theorem proving (FTP) shows promise in process-level reliability, it is limited to _verification_ (checking known propositions). This leaves constructive problem-solving (finding unknown terms that satisfy specific conditions) underexplored and disconnected from process-level verifiability. To bridge this gap, we introduce **FPS** (_**F**ormal **P**roblem-**S**olving_), a principled framework to encompass the end-to-end problem-solving process in Lean 4. In FPS, the answer is an unknown metavariable coupled with a proof obligation, forcing it to be mathematically derived and verified. We further present **D-FPS** (_**D**eductive **FPS**_), which enforces a rigorous chain-of-thought structure, aligning formal derivation with human reasoning steps. To support this direction, we construct three benchmarks via the manual refactoring of over 1,000 problems: **FormalMath500**, **MiniF2F-Solving**, and **PutnamBench-Solving**. We further propose **RPE** (_**R**estricted **P**ropositional **E**quivalence_), a symbolic metric that evaluates semantic correctness beyond brittle string matching. Extensive experiments with state-of-the-art provers reveal that solving is significantly harder than proving, highlighting the ``alignment tax'' required to transition from loose validity checking to constructive, human-aligned reasoning.

Deep Learning · Large Language Models

Kongcheng Zhang, QI YAO, Shunyu Liu, Wenjian Zhang, Cen, Yang Zhou, Wenkai Fang, Yiru Zhao, Baisheng Lai, Mingli Song

Reinforcement Learning (RL) has shown promise for aligning Large Language Models (LLMs) to follow instructions with various constraints. Despite the encouraging results, RL improvement inevitably relies on sampling successful, high-quality responses; however, the initial model often struggles to generate responses that satisfy all constraints due to its limited capabilities, yielding sparse or indistinguishable rewards that impede learning. In this work, we propose ***H**indsight **i**nstruction **R**eplay* (HiR), a novel sample-efficient RL framework for complex instruction following tasks, which employs a *select*-then-*rewrite* strategy to *replay failed attempts as successes* based on the constraints that have been satisfied in hindsight. We perform RL on these replayed samples as well as the original ones, theoretically framing the objective as dual-preference learning at both the instruction- and response-level to enable efficient optimization using only a binary reward signal. Extensive experiments demonstrate that the proposed HiR yields promising results across different instruction following tasks, while requiring less computational budget. Our code and dataset are available at anonymous url.

Reinforcement Learning · Batch/Offline

Nazim Bendib, Nicolas Perrin-Gilbert, Olivier Sigaud

Offline zero-shot reinforcement learning (RL) aims to learn agents that optimize unseen reward functions without additional environment interaction. The standard approach to this problem trains task-conditioned policies by sampling task vectors that define linear reward functions over learned state representations. In most existing algorithms, these task vectors are randomly sampled, implicitly assuming this adequately captures the structure of the task space. We argue that doing so leads to suboptimal zero-shot generalization. To address this limitation, we propose extracting task vectors directly from the offline dataset and using them to define the task distribution used for policy training. We introduce a simple and general reward function extraction procedure that integrates into existing offline zero-shot RL algorithms. Across multiple benchmark environments and baselines, our approach improves zero-shot performance by an average of 20%, highlighting the importance of principled task sampling in offline zero-shot RL.

General Machine Learning · Evaluation

Archana Warrier, Dat Nguyen, Michelangelo Naim, Moksh Jain, Yichao Liang, Karen Schroeder, Cambridge Yang, Josh Tenenbaum, Sebastian Vollmer, Kevin Ellis 等

World models are central to building AI agents capable of flexible reasoning and planning. Yet current evaluations (i) test only properties measurable from observed interactions within an environment, such as next-frame prediction or task return, and (ii) do not test whether a learned model supports diverse queries about the same environment. In contrast, humans build \textit{general-purpose} models that can answer many different questions about an environment---including questions that require understanding global structure and counterfactual consequences. We propose *WorldTest*: a protocol for evaluating agents' ability to learn general-purpose models. A *WorldTest* benchmark pairs environments with multiple *environment-level queries*---properties of the full environment---rather than objectives defined only on observed trajectories. Individually, these queries can target global and counterfactual properties (e.g., reachability or the effects of interventions) that are not determined by any single rollout distribution. Collectively, they assess model generality across query types. We instantiate WorldTest as *AutumnBench*, a minimal yet expressive benchmark of 43 interactive grid-world environments and 129 tasks across three query families for both humans and learning agents. AutumnBench supports diverse environments and evaluations, including queries to evaluate prediction, counterfactual reasoning, and long-horizon planning. Experiments with 517 human participants and five frontier models show that humans substantially outperform these models, a gap we attribute to differences in exploration and belief updating. *WorldTest* and *AutumnBench* provide a rigorous framework for evaluating world-model learning and expose critical limitations in current approaches.

Theory · Reinforcement Learning and Planning

Safwan Labbi, Paul Mangold, Daniil Tiapkin, Eric Moulines

In this paper, we study the role of the critic in actor-critic for entropy-regularized, finite, discounted environments. We establish that, when the critic is exact, using the latter as a baseline is an actual variance-reduction method. In this case, actor-critic with stochastic gradients matches the sample complexity of deterministic policy gradient, reaching an $\epsilon$-optimal regularized value with $\tilde{O}(\log(1/\epsilon))$ samples. In practice, the critic is learned alongside the actor: the variance of the actor update is then influenced by the critic's variance and bias. Specifically, when the critic has a sufficiently small error, the variance reduction and rapid convergence are preserved. This suggests to learn the critic first, keeping it up to date after each actor update, underscoring the pivotal role of accurate critic estimation in actor-critic methods.