ICML 2026 Papers

Multimodal Federated Learning (MMFL) addresses collaborative training across clients with heterogeneous modality configurations, where effective client selection becomes critical under the compounded challenges of modality, distribution, and quantity heterogeneity. Existing selection methods operate within a reactive paradigm, responding to current observations without anticipating how decisions influence future optimization trajectories. This myopic approach leads to suboptimal convergence when training dynamics shift rapidly under severe heterogeneity. We propose FedSSM, which reconceptualizes client selection as a proactive decision-making process by predicting training dynamics through decision-aware state space models. The prediction error yields a \emph{surprise} signal that quantifies uncertainty and governs adaptive participation budgets and exploration-exploitation trade-offs via counterfactual reasoning over candidate actions. For aggregation, we introduce trust-weighted fusion with modality-specific routing, where surprise calibrates sensitivity to client anomalies. Experiments on four multimodal benchmarks demonstrate that FedSSM achieves 2.5--4.5\% accuracy improvements over state-of-the-art methods while reducing communication rounds by over 30\%.

Dirac-Frenkel instantaneous residual minimization evolves nonlinear parametrizations of PDE solutions in time, but ill-conditioning can render the parameter dynamics non-unique. We interpret this non-uniqueness as a gauge freedom: nullspace directions that leave the time derivative unchanged can be used to select better-conditioned parameter velocities. Building on Onsager's minimum-dissipation principle, we introduce a history variable---interpretable as momentum---and inject it only along the nullspace directions. The resulting Dirac-Frenkel-Onsager dynamics preserve instantaneous residual minimization, in contrast to standard regularization that can introduce bias, while promoting temporally smooth parameter evolution. Examples demonstrate that the approach leads to increased robustness in singular and near-singular regimes.

Scientific computing has long relied on double precision (64-bit floating point) arithmetic to guarantee accuracy in simulations of real-world phenomena. However, the growing availability of hardware accelerators such as Graphics Processing Units (GPUs) has made low-precision formats attractive due to their superior performance, reduced memory footprint, and improved energy efficiency. In this work, we investigate the role of mixed-precision arithmetic in neural-network based Variational Monte Carlo (VMC), a widely used method for solving computationally otherwise intractable quantum many-body systems. We first derive general analytical bounds on the error introduced by reduced precision on Metropolis-Hastings MCMC, and then empirically validate these bounds on the use-case of VMC. We demonstrate that significant portions of the algorithm, in particular, sampling the quantum state, can be executed in half precision without loss of accuracy. More broadly, this work provides a theoretical framework to assess the applicability of mixed-precision arithmetic in machine-learning approaches that rely on MCMC sampling. In the context of VMC, we additionally demonstrate the practical effectiveness of mixed-precision strategies, enabling more scalable and energy-efficient simulations of quantum many-body systems.

Recent progress in multi-turn reinforcement learning (RL) has significantly improved reasoning LLMs' performances on complex interactive tasks. Despite advances in stabilization techniques such as fine-grained credit assignment and trajectory filtering, instability remains pervasive and often leads to training collapse. We argue that this instability stems from inefficient exploration in multi-turn settings, where policies continue to generate low-information actions that neither reduce uncertainty nor advance task progress. To address this issue, we propose Token- and Turn-level Policy Optimization (T$^2$PO), an uncertainty-aware framework that explicitly controls exploration at fine-grained levels. At the token level, T$^2$PO monitors uncertainty dynamics and triggers a thinking intervention once the marginal uncertainty change falls below a threshold. At the turn level, T$^2$PO identifies interactions with negligible exploration progress and dynamically resamples such turns to avoid wasted rollouts. We evaluate T$^2$PO in diverse environments, including WebShop, ALFWorld, and Search QA, demonstrating substantial gains in training stability and performance improvements with better exploration efficiency. Code is available at https://github.com/WillDreamer/T2PO.

Synthesizing authentic phonocardiograms (PCG) from ubiquitous electrocardiograms (ECG) is a critical task for accessible cardiac monitoring. Existing generative models, however, struggle to capture the heart's complex electromechanical coupling, failing to meet the dual requirements of temporal precision and physiological fidelity needed for clinically relevant waveform analysis. We introduce the Temporally-Aware VAE-Diffusion model, a synergistic hybrid architecture that resolves this trade-off. Our architecture enforces tight physiological coupling through an Enhanced Condition Fusion mechanism and explicitly models long-range cardiac dynamics via Temporal Attention Blocks. On the EPHNOGRAM benchmark, our model sets a new state-of-the-art, achieving a Pearson correlation of 0.910$\pm$0.008, 95.95\% S1 detection accuracy, and a precise 12.0 ms timing error, significantly outperforming leading diffusion and Transformer baselines. Crucially, our work provides a reproducible zero-shot transfer evaluation for ECG-to-PCG synthesis. Evaluated on the synchronized PhysioNet/CinC 2016 training-a/MITHSDB subset without target-domain training, our model preserves high waveform fidelity and clinically relevant timing structure under domain shift, including on pathological recordings. These results support cross-dataset robustness of the proposed synthesis framework, while downstream diagnostic validation remains an important direction for future work.

Bayesian additive spanning tree (BAST) is an useful tool for interpretable, non-parametric regression on complex constrained domains. It improves upon the performance of Bayesian additive regression trees (BART) by replacing axis-aligned splits through binary tree components by cuts on a spanning tree components, enabling the formation of contiguous splits that respect the underlying complex structure. While BAST is effective for learning on constrained spaces, it still relies on hard partitions, albeit on spanning trees, which limits its ability to represent smoothly varying functions on constrained domains. We propose Probabilistically-routed Bayesian additive spanning trees (PR-BAST), a principled relaxation that replaces hard cuts on spanning tree components with probabilistic routing along spanning tree components. PR-BAST represents the regression surface as an additive ensemble of such spanning tree-aligned smooth components. Conditional on a fixed spanning tree, each component in PR-BAST induces a Gaussian random field with a sparse, tree-structured precision matrix, enabling scalable posterior computation via sparse linear algebra. We theoretically establish that PR-BAST yields strictly faster posterior contraction rates compared to BAST under graph-smooth truth. Experiments on synthetic and real datasets demonstrate that PR-BAST consistently improves accuracy over BAST and other baselines, while retaining the interpretability of tree-based models.

Existing calibration methods for Large Language Models (LLMs) often overlook a critical dimension of trustworthiness: a model's {\em behavioral robustness} to irrelevant or misleading information. In this paper, we argue that a model's true confidence should reflect its stability under cognitive pressure. We introduce \textsc{CaliDist}, a novel post-hoc calibration approach that directly measures and penalizes a model's susceptibility to distraction. \textsc{CaliDist} quantifies how an LLM's predictions and uncertainty change when its input prompt is perturbed with semantic \textit{distractors}. This stability (or lack thereof) signal is then used to adaptively scale the model's initial confidence score. Our extensive experiments on seven Natural Language Understanding classification benchmarks using six distinct LLMs show that \textsc{CaliDist} consistently achieves lower Expected Calibration Error (ECE) and Brier Score compared with strong baselines. Remarkably, our method reduces the ECE from 23\% to 7\% on average—a relative improvement of 70\%—demonstrating that behavioral stability is a powerful signal for calibration. We make our code and datasets available at \url{github.com/m-anas-j/CaliDist}.

Anomaly detection (AD) methods often ignore causal dependencies and treat anomalies as outliers, which is brittle when anomalies are primarily mechanism violations rather than extreme values. We propose CausalAno, a causal-aware detector that trains a causal GAN on normal data and leverages its discriminator to learn mechanism-consistent representations. We score test samples by fitting a Gaussian model in this feature space and computing the Mahalanobis distance, measuring deviation from the normal causal manifold. We demonstrate the effectiveness of CausalAno with extensive experiments on 28 tabular datasets (18 continuous-only and 10 mixed-type), comparing against 16 SOTA baselines. Our results show consistent improvements across both mixed-type and numerical-only settings. Our ablation studies further confirm that the gains come from the causal factorization in the causal GAN rather than a generic GAN backbone. Our CausalAno offers a practical and effective solution for real-world applications where anomalies often arise from mechanism-violating behaviors.

Large Vision-Language Models (LVLMs) have achieved remarkable progress in multimodal reasoning tasks; however, their learning remains constrained by the limitations of human-annotated supervision. Recent self-rewarding approaches attempt to overcome this constraint by allowing models to act as their own critics or reward providers. Yet, purely text-based self-evaluation struggles to verify complex visual reasoning steps and often suffers from evaluation hallucinations. To address these challenges, inspired by recent advances in tool-integrated reasoning, we propose Agent0-VL, a self-evolving vision-language agent that achieves continual improvement with tool-integrated reasoning. Agent0-VL incorporates tool usage not only into reasoning but also into self-evaluation and self-repair, enabling the model to introspect, verify, and refine its reasoning through evidence-grounded analysis. It unifies two synergistic roles within a single LVLM: a Solver that performs multi-turn tool-integrated reasoning, and a Verifier that generates structured feedback and fine-grained self-rewards through tool-grounded critique. These roles interact through a Self-Evolving Reasoning Cycle, where tool-based verification and reinforcement learning jointly align the reasoning and evaluation distributions for stable self-improvement. Through this zero-external-reward evolution, Agent0-VL aligns its reasoning and verification behaviors without any human annotation or external reward models, achieving continual self-improvement. Experiments on chart reasoning, geometric problem solving, and visual scientific analysis show that Agent0-VL achieves an 12.5% improvement over the Qwen-VL base model.

Vector Symbolic Algebras (VSAs) enable robust neurosymbolic reasoning by encoding information into high-dimensional distributed representations. For continuous domains, Spatial Semantic Pointers (SSPs) extend this framework by mapping variables onto precise toroidal manifolds. While generative models offer a promising avenue for cleaning up (denoising) these representations, standard approaches like Flow Matching assume a flat Euclidean geometry. We demonstrate that this assumption fails for SSPs: Euclidean linear interpolants "cut through" the manifold's interior, destroying the phase and magnitude structure required for accurate decoding. To resolve this, we employ Geodesic Flow Matching, adapting Riemannian transport dynamics to strictly restrict the denoising flow to the SSP manifold. We validate this approach in a Spiking Neural SLAM system, showing that manifold-aware cleanup stabilizes path integration against drift. The method achieves a 72\% reduction in tracking error and enables a 40\% increase in neural efficiency compared to classical baselines.

This paper presents a theoretical investigation into the generalization capabilities of cross-domain knowledge distillation. Utilizing a high-dimensional asymptotic analysis of a linear teacher–student model, we characterize the excess risk while accounting for both model and covariate shifts. Our results provide a formal guarantee for the efficacy of distillation: even when the source and target domains differ substantially, there still may exist a regime where the student model achieves superior generalization ability over the student-only baseline. Moreover, we identify a \textit{crossed double descent} phenomenon: the excess risk can vary non-monotonically with the teacher’s and student’s dimension-to-sample-size ratios. These results provide rigorous insight into when and why distillation helps across domains.

As the demand for emotional intelligence in large language models (LLMs) grows, a key challenge lies in understanding the internal mechanisms that give rise to emotional expression and in controlling emotions in generated text. This study addresses three core questions: (1) Do LLMs contain context-agnostic mechanisms shaping emotional expression? (2) What form do these mechanisms take? (3) Can they be harnessed for universal emotion control? We first construct a controlled dataset, $\textit{SEV}$ (Scenario–Event with Valence), to elicit comparable internal states across emotions. Subsequently, we extract context-agnostic emotion directions that reveal consistent, cross-context encoding of emotion (Q1). We identify neurons and attention heads that locally implement emotional computation through analytical decomposition and causal analysis, and validate their causal roles via ablation and enhancement interventions. Next, we quantify each sublayer's causal influence on the model’s final emotion representation and integrate the identified local components into coherent global emotion circuits that drive emotional expression (Q2). Directly modulating these circuits achieves 99.65% emotion-expression accuracy on the test set, surpassing prompting- and steering-based methods (Q3). To our knowledge, this is the first systematic study to uncover and validate emotion circuits in LLMs, offering new insights into interpretability and controllable emotional intelligence.

While the emerging field of agentic software engineering has spurred extensive research into post-training, this paradigm alone does not fully address the distribution mismatch between traditional static pre-training and dynamic deployment environments. In this paper, we instead investigate agentic mid-training as a scalable complementary approach. Central to our approach is *agent-native data* comprising two complementary components: *contextually-native trajectories* that preserve the complete information flow an agent experiences, offering broad coverage and diversity; and *environmentally-native trajectories* whose observations stem from actual tool invocations and test executions, providing interaction authenticity. On `SWE-Bench Verified`, our recipe outperforms the previous open software engineering mid-training recipe `Kimi-Dev` under two post-training settings with the same base model and agentic scaffold, while using fewer than half mid-training tokens (73.1B). Furthermore, our 32B and 72B models achieve state-of-the-art resolution rates of **56.1\%** and **58.5\%** among open agentic recipes using agentic scaffolds, despite starting from non-coder `Qwen2.5` base models. We also observe performance gains on general code generation and scientific benchmarks. We open-source a significant portion of our datasets, recipes, and model checkpoints to facilitate further research.

Medical large vision-language models (Med-LVLMs) have recently achieved remarkable progress in vision–language comprehension and medical image segmentation. However, existing models still struggle to unify these two capabilities, which is essential for achieving clinically reasoning that connects visual findings with semantic interpretation. We present MedSIGHT, a unified framework that equips Med-LVLMs with structured, pixel-level understanding for grounded visual comprehension. MedSIGHT introduces a novel Region Perceiver module that produces region-centric tokens, encoding spatial information directly into representation space of the language model. We further propose a medical region codebook into the LLM vocabulary, allowing the model to generate discrete region codes as symbolic representations of anatomical and pathological regions. These codes are decoded through the Region Perceiver to reconstruct segmentation mask, achieving end-to-end spatial grounding. Lastly, MedSIGHT combines Region Perceiver, Codebook and LLM using our proposed progressive training strategy to gradually aligns these modules stably. Trained on only 72K multimodal instruction pairs, MedSIGHT achieves state-of-the-art performance across diverse imaging modalities on both medical comprehension and segmentation tasks. Code and model are publicly available at [GitHub](https://github.com/Aofei-Chang/MedSIGHT).

We are considering the problem of sampling from a posterior distribution related to Bayesian inverse problems arising in science, engineering, and imaging. Our method belongs to the family of independence Metropolis--Hastings (IMH) sampling algorithms. These are quite common in Bayesian inference. Relying on the existence of an approximate posterior distribution that is cheaper to sample from but can have significant bias, we introduce Proximal-IMH, a scheme that removes this bias: it corrects samples from the approximate posterior solving an auxiliary optimization problem, yielding a local adjustment that trades off adherence to the exact model against stability around the approximate reference point. For idealized settings, we prove that the proximal correction tightens the match between approximate and exact posteriors, and thereby improves acceptance rates and mixing. The new method works with both linear and nonlinear input-output operators and is especially suitable for inverse problems where exact posterior sampling is too expensive. We perform several numerical experiments that include multimodal and data-driven priors and nonlinear input-output operators. The results show that Proximal-IMH reliably outperforms existing IMH variants.

Conventional LLMs may suffer from corpus heterogeneity and subtle changes in conditions. While finetuning can create the catastrophe forgetting issue, applications of meta-learning on LLMs are also limited due to their complexity and scalability. In this paper, we activate the meta-signal of $\beta$ within the SwiGLU blocks, resulting in a meta-gating mechanism that adaptively adjusts the nonlinearity of FFN. A hypernetwork is employed to dynamically produce $\beta$ under textual conditions, providing meta-controllability over LLMs. By testing on different condition types such as task, domain, persona, and style, our method outperforms finetuning and meta-learning baselines, and can generalize reasonably on unseen tasks, condition types, or instructions. Our codes are in https://github.com/AaronJi/MeGan.

We present a principled framework for confidence estimation in computed tomography (CT) reconstruction. Based on the sequential likelihood mixing framework (Kirschner et al., 2025), we establish confidence regions with theoretical coverage guarantees for deep-learning-based CT reconstructions. We consider a realistic forward model following the Beer-Lambert law, i.e., a log-linear forward model with Poisson noise, closely reflecting clinical and scientific imaging conditions. The framework is general and applies to both classical reconstruction algorithms and deep learning methods alike, including U-Nets, U-Net ensembles, and generative Diffusion models. Empirically, we demonstrate that deep reconstruction methods yield substantially tighter confidence regions than classical reconstructions, without sacrificing theoretical coverage guarantees. Our approach allows the detection of hallucinations in reconstructed images and provides interpretable visualizations of confidence regions. This establishes deep models not only as powerful estimators, but also as reliable tools for uncertainty-aware medical imaging.

Learning rate warm-up -- increasing the learning rate at the beginning of training -- has become a ubiquitous heuristic in modern deep learning, yet its theoretical foundations remain poorly understood. In this work, we provide a principled explanation for why warm-up improves training. We rely on a generalization of the $(L_0, L_1)$-smoothness condition, which bounds local curvature as a linear function of the loss suboptimality and exhibits desirable closure properties. We show -- both theoretically and empirically -- that this condition is satisfied by common neural architectures and accurately captures the curvature of the optimization landscape early in training. Adapting the learning rate in response to this curvature condition naturally induces a warm-up–like schedule, and we show that this choice yields provably faster convergence guarantees than using a fixed learning rate. Experiments on language and vision models show that the resulting one-parameter warm-up schedule can match tuned linear warm-up and improve over no warm-up.

Large language models (LLMs) are increasingly deployed as autonomous agents for multi-turn decision-making tasks. However, current agents typically rely on fixed cognitive patterns: non-thinking models generate immediate responses, while thinking models engage in deep reasoning uniformly. This rigidity is inefficient for long-horizon tasks, where cognitive demands vary significantly from step to step, with some requiring strategic planning and others only routine execution. In this paper, we introduce CogRouter, a framework that trains agents to dynamically adapt cognitive depth at each step. Grounded in ACT-R theory, we design four hierarchical cognitive levels ranging from instinctive responses to strategic planning. Our two-stage training approach includes Cognition-aware Supervised Fine-tuning (CogSFT) to instill stable level-specific patterns, and Cognition-aware Policy Optimization (CoPO) for step-level credit assignment via confidence-aware advantage reweighting. The key insight is that appropriate cognitive depth should maximize the confidence of the resulting action. Experiments on ALFWorld and ScienceWorld demonstrate that CogRouter achieves state-of-the-art performance with superior efficiency.