ICML 2026 Papers

Tokenizers provide the fundamental basis through which text is represented and processed by language models (LMs). Despite the importance of tokenization, its role in LM performance and behavior is poorly understood due to the challenge of measuring the impact of tokenization in isolation. To address this need, we present TokSuite, a collection of models and a benchmark that supports research into tokenization's influence on LMs. Specifically, we release fourteen pre-trained models that use different off-the-shelf tokenizers but are otherwise identical, using the same architecture, dataset, training budget, and initialization. We also release a multilingual robustness benchmark that measures model performance under real-world perturbations in English, Chinese, Farsi, Italian, and Turkish, curated by native annotators. Together, TokSuite allows robust decoupling of the influence of a model's tokenizer, supporting a series of novel findings that elucidate the respective benefits and shortcomings of a wide range of popular tokenizers.

We present a fine-grained theoretical analysis of the performance gap between two-stage reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO). Our study decomposes this gap into two sources: the explicit representation gap under exact optimization and the implicit representation gap under finite samples. In the exact optimization setting, we characterize how the relative capacities of the reward and policy model classes influence the final policy qualities. We show that RLHF, DPO, or online DPO can outperform one another depending on type of model mis-specifications. Notably, online DPO can outperform both RLHF and standard DPO when the reward and policy model classes are isomorphic and both mis-specified. In the approximate optimization setting, we provide a concrete construction where the ground-truth reward is sparse and show that RLHF requires significantly fewer samples than DPO to recover an effective reward model, highlighting a statistical advantage of two-stage learning. Together, these results provide a comprehensive understanding of the performance gap between RLHF and DPO under various settings, and offer practical insights into when each method is preferred.

While language models achieve impressive results, their *learning dynamics* are far from understood. Many domains of interest – such as natural language syntax, coding languages, arithmetic – are captured by context-free grammars (CFGs). In this work, we extend prior work on neural language modeling of CFGs in a novel direction: how language modeling behaves with respect to CFG *substructure*, namely sub*grammars*. We define subgrammars, and prove a set of fundamental theorems connecting language modeling and subgrammars. We show that language modeling loss recurses linearly over its top-level subgrammars; applied recursively, the loss decomposes into losses for "irreducible" subgrammars. Under additional assumptions, and empirically, parametrized models learn subgrammars in parallel, unlike children who first master simple substructures. We find that subgrammar pretraining can improve final performance, but only for tiny models relative to the grammar, while alignment analyses show that pretraining consistently leads to internal representations that better reflect the grammar’s substructure.

Modeling decision-making outside of controlled environments requires accounting for asynchronous, exogenous signals, such as notifications or algorithmic feeds, that dynamically alter user response times. Standard Drift-Diffusion Models (DDM) become analytically intractable when drift rates vary continuously with time. In this paper, we derive a closed-form analytical approximation for the first-passage time distribution of a single-boundary DDM with time-dependent drift, valid in the high-threshold regime. The main result allows us to analytically study the optimal timing of external signals to maximize the probability of a user response within our approximation framework. To evaluate our response time model, we conduct an extensive empirical comparison with state-of-the-art methods for user watch-time prediction and evaluation in simulated environments.

Large-scale, multi-subject widefield calcium imaging provides unprecedented access to brain-wide cortical dynamics. However, the high dimensionality, complex spatiotemporal structure, and substantial task-irrelevant activity in widefield recordings have largely restricted modeling efforts to single-session analyses, limiting scalability and generalization. While multi-subject pretrained models have been explored for some neural modalities, multi-subject models for widefield calcium imaging have not yet been demonstrated; further, subject-invariant zero-shot behavior decoding remains elusive for multi-subject models across neural modalities more broadly. As a first step toward foundation modeling of widefield data, we introduce WiCAT, a multi-subject model that leverages self-supervised pretraining to both outperform single-session models and enable zero-shot behavior decoding on unseen subjects. WiCAT introduces an atlas-grounded tokenization scheme without session-specific components and learns globally shared spatiotemporal representations. Across multiple widefield datasets, the pretrained model supports lightweight downstream decoding, transfers across subjects, tasks, and datasets, and outperforms baseline models. Notably, the model also achieves robust zero-shot continuous behavior decoding and left-out brain region reconstruction on unseen subjects.

Behavior cloning (BC) is a practical offline imitation learning method, but it often fails when expert demonstrations are limited. Recent works have introduced a class of architectures named predictive inverse dynamics models (PIDM) that combine a future state predictor with an inverse dynamics model. While PIDM often outperforms BC, the reasons behind its benefits remain unclear. In this paper, we provide a theoretical explanation: PIDM introduces a bias-variance tradeoff. While predicting the future state introduces bias, conditioning the IDM on the prediction can significantly reduce variance. We establish conditions for PIDM to achieve lower prediction error and higher sample efficiency than BC, with the gap widening when additional data sources are available. We validate the theoretical insights empirically in 2D navigation tasks, where BC requires up to five times (three times on average) more demonstrations than PIDM to reach comparable performance. Results are also illustrated in a complex 3D environment in a modern video game with high-dimensional visual inputs and stochastic transitions, where BC requires over 66\% more samples than PIDM.

A sustainable diet represents a multi-dimensional synergy among four essential pillars: nutrition adequacy, economic affordability, cultural acceptability, and environmental respect. Despite the prevalence of population-level sustainability modeling, practical implementation relies on effective individual-level adoption. This transition is often hindered by inter-individual heterogeneity, posing a formidable challenge in aligning sustainable diet requirements with individual preferences. To address this issue, we propose a personalized sustainable diet recommendation model based on a constraint-aware decision-making mechanism, where sustainability is incorporated through learnable constraints rather than modeled as user preferences. To systematically evaluate the proposed approach, we construct a sustainable diet dataset named SusDiet with about 150k recipes, characterized by broad coverage of sustainability indicators across four dimensions. Experimental results on this dataset show that our method promotes more sustainable choices without compromising individual preference. This work establishes a framework for aligning individual dietary choices with planetary health, offering quantitative evidence to guide future sustainable diet interventions and policy-making for sustainable development.

Reward models play a fundamental role in aligning large language models with human preferences. Existing methods predominantly follow two paradigms: scalar discriminative preference models, which are efficient but lack interpretability, and generative judging models, which offer richer reasoning at the cost of higher computational overhead. We observe that the log-probability margin between verdict tokens strongly correlates with prediction correctness, providing a reliable proxy for instance difficulty without additional inference cost. Building on this insight, we propose CAMEL, a confidence-gated reflection framework that performs a lightweight single-token preference decision first and selectively invokes reflection only for low-confidence instances. To induce effective self-correction, we train the model via reinforcement learning with counterfactual prefix augmentation, which exposes the model to diverse initial verdicts and encourages genuine revision. Empirically, CAMEL achieves state-of-the-art performance on three widely used reward-model benchmarks with 82.9% average accuracy, surpassing the best prior model by 3.2% and outperforming 70B-parameter models using only 14B parameters, while establishing a strictly better accuracy-efficiency Pareto frontier.

Reinforcement learning with verifiable rewards (RLVR) has enabled progress on reasoning-intensive tasks by relying on task-specific verifiers that provide automated correctness signals. However, many realistic language tasks are difficult to equip with reliable verifiers, motivating a growing reliance on reinforcement learning from human feedback (RLHF). In this setting, we argue that a closer examination of how human feedback should be interpreted is essential. We introduce Regret-based Preference Optimization (RePO), which reframes RLHF through *regret minimization* rather than reward maximization. Human preferences are often shaped by *prospective* anticipation of outcomes and *counterfactual* comparisons to alternative behaviors, rather than by immediate, outcome-independent utility. RePO captures this structure by modeling preferences as behavior-conditioned assessments of relative suboptimality. Within a KL-regularized reinforcement learning framework, RePO admits a closed-form policy update compatible with direct preference optimization. Experiments on mathematical reasoning benchmarks and human-annotated preference datasets demonstrate consistent performance gains, indicating that regret-based preference learning is an effective and human-aligned approach for training large language models.

Wireless channel modeling is essential for the design, analysis, and optimization of modern wireless sensing and communication systems. However, accurately modeling wireless channels in electrically large and complex environments remains a long-standing challenge, owing to the intricate interactions between radio-frequency (RF) signals and surrounding objects (e.g., reflection, diffraction, and scattering). Unlike conventional ray-tracing pipelines that rely on hand-engineer interaction rules, or black-box neural surrogates that do not explicitly model physical structure, we introduce SNRFT, a novel framework that integrates neural representations with physics-based RF propagation modeling. Our key idea is to view RF transport as a stochastic propagation process, from which a material-dependent attenuation coefficient emerges naturally as the rate parameter governing transport dynamics. This formulation inherently satisfies key physical constraints such as reciprocity and reversibility. Building on this foundation, we employ implicit neural representations to capture complex RF-object interactions while preserving the composability of traditional ray tracing. Extensive evaluations on real-world wireless communication and sensing testbeds demonstrate that SNRFT consistently outperforms existing methods, while requiring significantly fewer training samples.

Over the years, the unsupervised and supervised learning research directions of tensor networks (TNs) have mainly developed in parallel. In this paper, we provide a view for their cooperative advancement through a novel mixed tensor network learning (MTNL) framework that unifies the two fields. Specifically, inspired by supervised TN learning tasks, multiple TNs are fused in a deep-network style in MTNL, enhancing the expressive power for the unsupervised TN learning tasks. We then develop a more flexible TN structure search prior with theoretical guarantees for learning multiple TN structures, aligning with trends in many supervised learning setups. More interestingly, by combining these components within a Bayesian framework, we show that MTNL induces a lightweight uncertainty quantification mechanism that is theoretically guaranteed by its connection to the dropout-based counterpart problem, making the mechanism a potential alternative for large-scale learning problems. Finally, we demonstrate the effectiveness of the MTNL framework on tensor recovery, parameter-efficient fine-tuning, and tensor regression experiments.

Federated Learning (FL) offers a privacy-preserving pathway for aligning Large Language Models (LLMs); however, existing frameworks typically enforce a monolithic reward model, inevitably averaging out inherently conflicting user preferences (e.g., helpfulness vs. harmlessness). While Variational Preference Learning (VPL) offers a pathway to personalization, adapting it to decentralized settings presents a fundamental challenge: \textit{posterior collapse} driven by severe local data scarcity and heterogeneity. In this paper, we propose Federated Variational Preference Alignment with Gumbel-Softmax Prior (FedVPA-GP), a framework designed to disentangle diverse preferences without compromising privacy. To stabilize variational inference, we introduce a Federated Mixture Prior that enables clients to leverage the aggregate population distribution as a dynamic prior. Furthermore, we incorporate an Orthogonal Loss that explicitly enforces the separation of preference prototypes in the latent space. Experiments on the HH-RLHF dataset demonstrate that FedVPA-GP significantly outperforms monolithic baselines, successfully disentangling conflicting user intents and enabling dynamic preference switching.

LLMs cannot reliably recognize their parametric knowledge boundaries and often hallucinate answers to outside-of-boundary questions. In this paper, we introduce MASH (Modeling Abstention via Selective Help-seeking), a training framework that readily extracts abstentions from LLMs. Our key idea is that any external help-seeking by an LLM, i.e. search tool use, can serve as a proxy for abstention if the external help (search) is appropriately penalized while also rewarding answer accuracy. MASH operationalizes this idea using reinforcement learning with a pay-per-search reward. We run experiments on three knowledge-intensive QA datasets. Our results show that MASH substantially improves upon the selective help-seeking performance of prior efficient search approaches; on multi-hop datasets, it improves answer accuracy by 7.6%. Furthermore, MASH demonstrates strong off-the-shelf abstention performance, showcasing behavior competitive with prior abstention methods that additionally require predetermining model knowledge boundaries to construct training data. Overall, we show that MASH training effectively aligns search tool use with parametric knowledge, which can be successfully leveraged for making abstention decisions and efficient search tool use.

Existing online benchmarks for mobile GUI agents remain largely app-centric and task-homogeneous, failing to reflect the diversity and instability of real-world mobile usage. To this end, we introduce VenusBench-Mobile, a challenging online benchmark for evaluating general-purpose mobile GUI agents under realistic, user-centric conditions. VenusBench-Mobile builds two core evaluation pillars: defining what to evaluate via user-intent-driven task design that reflects real mobile usage, and how to evaluate through a capability-oriented annotation scheme for fine-grained agent behavior analysis. Extensive evaluation of state-of-the-art mobile GUI agents reveals large performance gaps relative to prior benchmarks, indicating that VenusBench-Mobile poses substantially more challenging and realistic tasks and that current agents remain far from reliable real-world deployment. Diagnostic analysis further shows that failures are dominated by deficiencies in perception and memory, which are largely obscured by coarse-grained evaluations. Moreover, even the strongest agents exhibit near-zero success under environment variations, highlighting their brittleness in realistic settings. Based on these insights, we believe VenusBench-Mobile provides an important stepping stone toward robust real-world deployment of mobile GUI agents. Code and data are available at https://github.com/inclusionAI/UI-Venus/tree/VenusBench-Mobile.

With large language models (LLMs) now performing strongly across diverse tasks, there is growing demand for them to personalize outputs for individual users. Personalization is typically framed as an additional layer on top of a base NLP task, requiring model responses to meet user-specific needs while still accomplishing the underlying task. From a token-level perspective, different tokens in a response contribute to personalization to varying degrees. Tokens with higher personalization relevance should therefore receive greater emphasis when developing personalized LLMs. However, accurately estimating such personalization degrees remains challenging. To address this challenge, we propose PerContrast, a self-contrast method that estimates each output token’s dependence on user-specific information through causal intervention. Building on this mechanism, we develop the PerCE loss, which adaptively upweights tokens with higher estimated personalization degrees during training via a bootstrap procedure, enabling the model to alternate between estimating and optimizing these tokens. Experiments on multiple LLMs demonstrate that PerCE substantially improves personalization performance with minimal additional cost, achieving average gains of over 10\% and up to 68.04\% on the LongLaMP dataset, along with strong cross-task and cross-scenario transferability. These results highlight the importance of token-level personalization modeling and establish token-aware training as a simple yet effective paradigm for advancing personalized LLMs.

Pairwise preference learning is central to machine learning, with recent applications in aligning language models with human preferences. A typical dataset consists of triplets $(x, y^+, y^-)$, where response $y^+$ is preferred over response $y^-$ for context $x$. The Bradley--Terry (BT) model is the predominant approach, modeling preference probabilities as a function of latent score differences. Standard practice assumes data follows this model and learns the latent scores accordingly. However, real data may violate this assumption, and it remains unclear what BT learning recovers in such cases. Starting from triplet comparison data, we formalize the preference information it encodes through the conditional preference distribution (CPRD). We give precise conditions for when BT is appropriate for modeling the CPRD, and identify factors governing sample efficiency---namely, margin and connectivity. Together, these results offer a data-centric foundation for understanding what preference learning actually recovers.

Recently, machine learning Hamiltonian (MLH) models have gained traction as fast approximations of electronic structures such as orbitals and electron densities, while also enabling direct evaluation of energies and forces from their predictions. However, despite their physical grounding, existing Hamiltonian models are evaluated mainly by reconstruction metrics, leaving it unclear how well they perform as energy–force predictors. We address this gap with a benchmark that computes energies and forces directly from predicted Hamiltonians. Within this framework, we propose QHFlow2, a state-of-the-art Hamiltonian model with an SO(2)-equivariant backbone and a two-stage edge update. QHFlow2 achieves $40\%$ lower Hamiltonian error than the previous best model with fewer parameters. Under direct evaluation on MD17/rMD17, it is the first Hamiltonian model to reach NequIP-level force accuracy while achieving up to $20\times$ lower energy MAE. On QH9, QHFlow2 reduces energy error by up to $20\times$ compared to MACE. Finally, we demonstrate that QHFlow2 exhibits consistent scaling behavior with respect to model capacity and data, and that improvements in Hamiltonian accuracy effectively translate into more accurate energy and force computations.

Traditional large language model (LLM) alignment methods are based on Reinforcement Learning From Human Feedback (RLHF), which learns a single reward model (implicitly or explicitly) from pairwise comparison data. This approach implicitly assumes homogeneous preferences across human labelers---an assumption that is violated in practice. As a result, the learned reward model in RLHF is generally misspecified: Prior work shows that it is inconsistent with the population-average utility, incurring large distortion, and that recovering this utilitarian objective is provably impossible in the worst case. In this work, we show that the average utility is recoverable under a mild assumption. Our method, the Sign Estimator, simply replaces the standard cross-entropy loss function with a notion of binary classification loss and yields a reward model that is ordinally consistent with the population-average utility. We further establish a fast finite-sample convergence rate of $O({n^{-{1}/{3}}})$ , which provides, to our knowledge, the first consistent estimator for heterogeneous preferences that does not suffer from the curse of dimensionality.

Emotional support plays an important role in dialogue systems, and its success depends on adapting to a user’s evolving and implicit needs across multi-turn interactions while leveraging the strong reasoning capacity of large language models. However, since signals about user needs are often weak, indirect, and can only be disambiguated through multi-turn interaction, existing emotional support methods often struggle to acquire and generalize relevant conversational knowledge efficiently. To bridge this gap, we introduce User-Aware Active Knowledge Acquisition (UKA), a gradient-free active dialogue learning framework that explicitly represents uncertainty about user needs and incorporates active learning into both knowledge acquisition and response selection. We propose a Theory-of-Mind uncertainty estimation mechanism that allows the model to prioritize responses, thereby eliciting more informative user feedback. UKA is capable of efficiently exploring user-aligned conversational knowledge during training while maintaining robustness at test time. Experiments across multiple dialogue benchmarks and model architectures demonstrate that our approach consistently outperforms strong baselines in dialogue quality and user alignment.