| Total: 143
Modern neural networks can be accurate yet poorly calibrated. We present Random Geometric Calibration (RGC), a post hoc method that augments confidence estimation with a geometric signal derived from distances to training data in an aggregated semantic representation space. Prior geometric calibration methods in feature space often rely on architecture-specific layer choices, which can be unstable across models and datasets. To avoid this dependence, RGC samples a fixed number of intermediate blocks uniformly at random and reuses this set at both calibration and test time. We instantiate RGC with RGCL, which applies spatial pyramid pooling to each sampled activation, projects features to a fixed dimension, and computes a nearest-neighbor separation score that is mapped to probabilities via isotonic regression. We also present RGCC, which replaces pooling and projection with uniform coordinate sampling for improved computational efficiency. We evaluate RGCL and RGCC across multiple datasets, including CIFAR-10, CIFAR-100, and Tiny-ImageNet, and across a diverse set of architectures such as ResNet, DenseNet, and DINOv2. Our results show that RGCL achieves state-of-the-art or near state-of-the-art calibration performance, reducing expected calibration error by approximately 78% on average and by up to 97% for some models. Finally, we provide representation analysis that explains why RGCL and RGCC succeed without layer selection and characterize their computational overheads. We observe that layers exhibiting favorable geometric structure tend to emerge in later network blocks, and that randomized aggregation implicitly emphasizes these layers, leading to robust and reliable calibration performance.
Graph Neural Networks (GNNs) have become increasingly central to several analysis tasks in various domains, ranging, e.g., from social networks to molecular analysis. Understanding the decision-making process of GNNs is a critical challenge in machine learning, particularly when post-hoc explanations, attempting to answer why specific inputs are classified in a certain way by a given model, are required for specific input predictions. Existing post-hoc explainability methods for GNNs often rely on counterfactual reasoning, grounded in the `if this had not occurred' thinking, that typically identifies minimal subgraphs perturbations that would change a prediction. Notably, less attention has been posed on the equally important semifactual reasoning, grounded in the `even if this has not occurred' thinking, that results in identifying maximal subgraphs perturbations that would not change a prediction. In this paper, we introduce a novel technique for post-hoc explainability queries in GNNs by focusing on the semifactual reasoning. We first thoroughly investigate the computational complexity of several reasoning and optimization problems related to the computation of semifactuals, and then propose a novel learning architecture for addressing their computation. Finally, we experimentally evaluate the proposed approach in a variety of settings, showing its effectiveness in generating high-quality semifactual explanations.
We investigate the computational complexity of neural network verification in quantised settings. We distinguish three classes of Feedforward Neural Networks (FNNs): rational FNNs with exact rational weights, quantised FNNs whose weights come from a finite-width arithmetic, and dynamically quantised FNNs in which rational networks are evaluated with respect to a given finite-width arithmetic. We consider two types of specifications used in the literature. Linear programming (LP) specifications are conjunctions of linear constraints, while bit-vector (BV) specifications allow reasoning at the bit level and can express non-linear constraints. Our results give a complexity landscape of these verification problems. For quantised FNNs with fixed arithmetic precision, we show that verification under both LP and BV specifications remains NP-complete, matching the complexity of the rational case. For dynamically quantised FNNs with BV specifications, we establish upper bounds, complementing a previously known PSPACE-hardness result.
Coverage-based prime implicant explanations are formal explanations offering a number of valuable assets, especially in terms of faithfulness and generality. Unfortunately, deriving a coverage-based prime implicant explanation for an instance is computationally hard in the general case (the problem of identifying such an explanation being at the second level of the polynomial hierarchy). In this paper, we focus on the computation of a coverage-based prime implicant explanation for an instance given a tree-based model. We show that the specific nature of the domain theory linking the Boolean conditions in such models makes the problem computationally easier. We present a greedy algorithm to derive coverage-based prime implicant explanations when dealing with a tree-based model. We also present an empirical evaluation showing that this algorithm is efficient enough to be used in many practical cases.
Accurate spatio-temporal forecasting requires modeling interactions across multiple spatial and temporal scales. Existing Graph Neural Network (GNN) forecasters primarily operate at a single local scale, limiting their ability to capture global processes that govern system dynamics. Hierarchical GNNs (HGNNs) aim to address this limitation by learning multiscale representations, but in practice, they often fail to preserve global information during coarse-to-fine propagation. We identify this limitation as downward memory loss, where global trends learned at coarse resolutions diminish before influencing fine-grained predictions, leading to misaligned local dynamics. We propose HiGFlow, an HGNN forecaster that explicitly preserves multiscale trends through a self-updating memory buffer that integrates into the coarse-to-fine information flow. This design maintains global contextual signals as an inductive bias throughout hierarchical decoding. We provide a theoretical analysis indicating that when decoding, HiGFlow preserves multiscale trends where residual-based architectures may fail to do so. Empirically, HiGFlow achieves substantially lower MAE and RMSE than state-of-the-art forecasting models across multiple benchmark datasets, demonstrating the importance of explicit global memory in multiscale spatio-temporal forecasting. Our implementation is available at https://github.com/TB862/HiGFlow.
Attributed graph clustering has achieved remarkable success by synergistically integrating topological structures and node attributes. While subspace learning has emerged as a dominant paradigm for node partitioning, most existing methods rely on implicit low-rank constraints, which often fail to capture complex nonlinear manifolds and suffer from prohibitive computational overhead on large-scale graphs. In this paper, we propose ELSS (Explicit Low-rank Structured Subspace learning), a scalable and robust framework that transcends implicit formulations. Specifically, ELSS learns an explicit and nonlinear low-rank subspace within a graph-structured embedding space, effectively uncovering latent cluster structures. To effectively mitigate the pervasive oversmoothing issue, we introduce a homophily-aware adaptive graph filter, which dynamically calibrates smoothing intensity to preserve discriminative ego-information. Furthermore, to ensure linear scalability, we develop a PageRank-guided structural sampling strategy for anchor-based approximation, which identifies pivotal landmarks based on their global topological prestige. Theoretical analysis guarantees that ELSS effectively mitigates spectral collapse while maintaining a linear complexity. Extensive experiments on diverse benchmarks demonstrate that ELSS consistently delivers superior clustering accuracy over state-of-the-art methods.
Large language models (LLMs) are attractive for context-aware time series forecasting because they can integrate heterogeneous textual signals, yet their discrete, language-oriented tokenization and embedding interfaces are misaligned with continuous numerical values, often harming numerical ordering and forecasting reliability. We propose TempoWave, a plug-and-play temporal wavelet digit interface that maps each scalar observation into digit-wise embeddings constructed from multi-wavelet, multi-scale coefficients. By directly overriding standard token representations, TempoWave seamlessly exposes both fine-grained local fluctuations and macro global structures in a transformer-compatible form, ensuring that precise numerical formatting, distinct digit identity, and robustness to common normalization operations are maintained throughout the LLM pipeline. Experiments across five context-enriched forecasting benchmarks demonstrate that TempoWave consistently improves LLM-based forecasters over standard numeric tokenization and alternative embedding interfaces, achieving a new state-of-the-art. These results highlight the numeric interface as a key bottleneck and suggest that principled multi-resolution embeddings can better couple LLMs' contextual reasoning with precise forecasting. Our code is available at https://github.com/DC-research/TempoWAVE.
Multimodal semantic communication systems face a critical challenge in extracting and aligning semantic features across heterogeneous modalities within a unified representation space, particularly under noisy transmission conditions. To address this, we propose Bridge, a cross-modal learning framework that integrates video, audio, and text into a unified semantic space through feature disentanglement and contrastive alignment. Bridge separates modality-specific and modality-invariant representations, enhancing both intra-modal precision and inter-modal generalization. During decoding, it leverages heterogeneous foundation models to preserve semantic fidelity under channel noise. Extensive experiments on multiple multimodal benchmarks under varying SNR levels demonstrate that Bridge maintains high semantic consistency and reconstruction quality, improving image reconstruction by approximately 16.3% in low-SNR regimes. Our work provides a practical and theoretically grounded framework for next-generation multimodal semantic communication systems.
Retrieval-Augmented Generation (RAG) is sensitive to the vast hyperparameters of the retriever and generator, yet optimizing them using given queries remains a challenging task due to the complex hyperparameter interactions and expensive evaluation costs. Existing algorithms are ineffective and slow in convergence, since they often treat RAG as a monolithic black box or only optimize partial hyperparameters. In this paper, we propose CDS4RAG, a framework that optimizes the full hyperparameters of RAG using given queries via a new formulation of cyclic dual-sequential problem. CDS4RAG is special in the sense that it distinguishes the hyperparameters of the retriever and generator, optimizing them in turn and in a cyclic manner. Such a paradigm allows us to design fine-grained within-cycle budget provision and expedite the optimization via cross-cycle seeding when optimizing the generator. Importantly, CDS4RAG is an algorithm-agnostic framework that can be paired with diverse general algorithms. Through experiments on four common benchmarks and two backbone LLMs, we reveal that CDS4RAG considerably boosts the vanilla algorithms in 21/24 cases while significantly outperforming state-of-the-art algorithms in all cases with up to 1.54x improvements of generation quality and better speedup.
Detecting hidden confounding is crucial for reliable causal analysis from observational data, directly determining which downstream causal inference method to be deployed. Inspired by the theory of higher-order regression, recent sample-efficient hypothesis testing strategies overcome the restrictive requirement of multiple heterogeneous data environment. Despite their progress on single-environment confounder detection, such methods suffer from intrinsic flaws that the structural functions of the causal models should be specified in prior (linear or specific kernel functions). By contrast, real-world data acquisition exhibits diverse, unknown forms of structural functions, imposing an important but challenging gap between theories of higher-order regressions to practical confounder detection. In this paper, we contribute a Bi-level Kernel Confounder Detection (BiKCD) framework by learning adaptive kernelized structural space of structural functions. Subsequently, our BiKCD constructs hypothesis testing by comparing coefficients from the higher-order regression and the classical ordinary least squares in learned kernelized space. Finally, the hypothesis is calibrated to ensure valid inference under adaptivity. Theoretically, we establish an oracle-type risk bound for the selected structural space over a candidate kernel family, with the Type-I error control for the downstream test. Extensive experiments on synthetic and real-world datasets demonstrate the effectiveness of the proposed BiKCD.
Multi-view anchor graph clustering has emerged as a high-efficiency paradigm of multi-view learning. However, how to design an effective anchor alignment mechanism within this framework remains an open challenge, which is formally termed the Anchor-Unaligned Problem (AUP). Current research fails to adequately address two pivotal aspects of this challenge: first, the construction of anchor graphs and the alignment of anchors are two independent stages, overlooking their potential synergistic reinforcement; second, selecting anchors from different views as an alignment baseline often renders the clustering performance highly sensitive to the baseline choice. To address these issues, we propose a unified framework termed One-Step Self-Aligned Anchor Learning for Multi-View Clustering (OSAA-MVC). Departing from conventional two-stage strategies, we integrate anchor alignment and anchor graph construction into a joint optimization process, thereby enabling their mutual reinforcement to improve clustering performance. To avoid the baseline selection issue, we introduce a novel mixed anchor strategy, which effectively bypasses the necessity of manual baseline selection while simultaneously capturing both view-specific and cross-view information. This strategy can stabilize clustering performance at a consistently high level. Extensive experiments demonstrate the superior efficiency and effectiveness of our proposed method compared to state-of-the-art competitors. The code is available at https://github.com/Jiamiao2024/OSAA-MVC.
Unsupervised learning of neural operators is constrained by numerical instability, causing predictions to diverge in long-horizon rollouts. To address this, we present a physics-constrained unsupervised neural operator for long-horizon PDE learning on generalized geometries (UNOP). This framework replaces differential constraints with integral consistency for stable, label-free learning. Unlike prior works, UNOP is built upon Latent Integral Physics Embedding (LIPE), which enforces physical consistency through integral constraints. To extend integral formulations to generalized geometries, the Geometry-Agnostic Latent Adapter (GALA) projects them onto a unified latent grid of PDE inputs, providing a regularized domain for spatial integral evaluation. Based on this shared embedding, the Gated Spectral Evolution Operator (GSEO) performs stable temporal integration while retaining spatial regions with sharp gradients and fine-scale structures, with the evolution constrained by the LIPE objective. Experiments on 1D, 2D, and 3D benchmarks show UNOP outperforms state of the art methods, reducing error accumulation by up to 60% in 20-step rollouts. Code is available at https://github.com/chengxinrui/UNOP.
Offline safe reinforcement learning learns high-return policies that satisfy hard safety constraints using only a pre-collected dataset. This setting is challenging due to the inability to explore, and the risk of propagating value errors through unsafe state-space regions. To address this, first, we characterize the safe state region by developing a framework for learning control barrier functions (CBFs) using a novel generalized Bellman operator, yielding a persistent safety set, from which the agent can remain safe indefinitely. Second, we show that several existing safety set estimation methods (e.g., reachability-constrained RL) can be formulated within our CBF learning framework, highlighting its generality. We further propose a new CBF that ensures safety under environment dynamics uncertainty, unlike standard CBFs designed for deterministic settings. Third, we propose a new reward maximization algorithm that effectively exploits our learned persistent safety set for reward critic estimation. Empirical results on standard benchmarks show that our approach achieves state-of-the-art safety with fewer constraint violations while maintaining competitive returns.
Recent advances in text-to-audio (TTA) and audio-to-audio (ATA) generation models have enabled the creation of highly realistic environmental sounds, raising growing concerns about malicious audio manipulation in real-world scenarios. To address this emerging threat, the ESDD 2026 Challenge was introduced as the first large-scale benchmark for Environmental Sound Deepfake Detection (ESDD), featuring two tracks that evaluate generalization to unseen generators and robustness under black-box, low-resource conditions. In this paper, we present BEAT2AASIST, an enhanced deepfake detection framework built upon the BEATs-AASIST baseline. Motivated by the observation that token-based audio representations may weaken explicit preservation of structured acoustic cues, the proposed method introduces a dual-branch AASIST architecture that explicitly splits BEATs-derived representations along frequency or channel dimensions. This design enables specialized modeling of complementary spoofing artifacts that may be attenuated in unified representations. To further enrich acoustic features, we incorporate multi-layer fusion strategies that aggregate information from multiple transformer layers using concatenation, CNN-gated, and SE-gated mechanisms. In addition, vocoder-based data augmentation with multiple high-fidelity neural vocoders is employed to enhance robustness against unseen and black-box spoofing attacks. Experimental results on the EnvSDD dataset demonstrate that BEAT2AASIST achieves strong and consistent performance across both challenge tracks. In particular, the proposed approach attains 3rd place in Track 2 and 4th place in Track 1 in the ESDD 2026 Challenge, despite using fewer ensemble components than top-ranked systems. These results suggest that explicit modeling of heterogeneous acoustic subspaces, combined with targeted representation fusion and data augmentation, provides an effective and efficient design strategy for real-world environmental sound deepfake detection. The code is available at https://github.com/ikwak2/BEAT2AASIST.
Partial multi-label learning (PML) addresses weakly-supervised scenarios where each instance is associated with a candidate label set containing both ground-truth and noisy labels. Existing PML methods primarily focus on instance-level features or pairwise label correlations for disambiguation. Building on recent insights that exploiting pairwise label correlations improves disambiguation, we observe that high-order label correlations provide even stronger disambiguation evidence in partial settings because ground-truth labels with high-order correlations frequently co-occur, while noise labels produce inconsistent combinations that rarely repeat. To exploit this property, we propose REHC-PML (Label Confidence REcovery with High-order Label Correlations in Partial Multi-label Learning), which mines frequent high-order co-occurrences from candidate sets to identify global high-order label correlations, selects instance-relevant correlations via Gumbel-Softmax pruning, and propagates their evidence to constituent labels for confidence recovery. Self-training iteratively refines pseudo-labels and trains the classifier. Extensive experiments on both real-world and UCI datasets demonstrate the effectiveness of REHC-PML.
Sleep staging is pivotal for assessing sleep quality and clinical diagnostics. While contactless radio-frequency (RF) sensing offers an unobtrusive alternative to Polysomnography (PSG), existing methods often struggle with generalization across diverse devices and environments due to the scarcity of annotated RF data. To overcome this limitation, we propose XSensorSleep, a cross-sensor domain generalization framework. Uniquely, XSensorSleep is trained exclusively on large-scale respiratory datasets from chest and abdominal belts and is directly transferable to RF-derived respiratory signals. To bridge the domain gap, we partition the training data into multiple pseudo-domains based on data sources and sensor types, forcing the model to extract domain-invariant features from heterogeneous respiratory patterns. Furthermore, we introduce a hierarchical alignment strategy: Epoch-Level Feature Alignment (ELFA) suppresses sensor-specific morphological artifacts, while Spectral Temporal Alignment (STA) captures invariant global sleep architectures. Specifically, STA leverages the rotation-invariance of eigenvalue spectra to align whole-night transition dynamics, ensuring the model prioritizes universal physiological rhythms over device-dependent signal fluctuations. Extensive evaluations across various home and clinical RF datasets demonstrate that XSensorSleep achieves superior generalization, enabling practical, automated, and label-free non-contact sleep staging.
Partial-label learning (PLL) addresses challenging scenarios where each instance is associated with a set of candidate labels and only one is the truth. Most existing PLL methods rely on static disambiguation heuristics, which are prone to error propagation when the ambiguity labels are high. To address this issue, we propose a novel mutual information guided reinforcement learning framework for partial label disambiguation (MR-PLL). In this framework, label disambiguation is formulated as a sequential Markov decision process, where an agent dynamically discriminates whether to retain, modify, or abstain from correcting ambiguous labels. To ensure reliable decision making, we introduce a mutual information guided gating mechanism that adaptively adjusts the confidence of soft label propagation according to the dependency between feature representations and labels. The abstention mechanism allows the model to postpone uncertain decisions, resulting in more reliable disambiguation. Furthermore, we design an information weighted reward term during the actor-critic process to gradually improve the label disambiguation ability of the policy, and provide theoretical analysis on convergence and bias reduction. Experiments on benchmark and real datasets verify the effectiveness of the proposed algorithm.
Compression techniques such as quantization and low-rank approximation enable large language models (LLMs) to run on current edge hardware with limited computing power, but the key challenge lies in balancing the allocation of precision and low rank within a fixed memory limit. We propose a training-free framework, BALANCER, which achieves global budget allocation for mixed-precision quantization and low-rank correction through information-guided subspace matrices. BALANCER collects a small set of calibration signals—activation statistics, reference gradient magnitudes, and a diagonal curvature proxy—and converts them into information-guided subspace matrices. Instead of relying on coarse-grained layer-wise importance allocation, BALANCER transforms these signals into gain curves for each weight in the matrix singular subspace, allowing a principled greedy allocator to distribute compression bits and ranks across the entire model. On the LLaMA-3.1-8B model, with an average of 3.6 bits per parameter, BALANCER only increases the perplexity from 6.63 to 6.78 (+0.15), while reducing memory from 16 GB to 3.9 GB, and enabling the model to be deployed on edge GPUs such as the Jetson Orin Nano (8 GB).
Aggregating temporal signals from historic interactions is a key step in future link prediction on dynamic graphs. However, incorporating long histories is resource-intensive. Hence, temporal graph neural networks (TGNNs) often rely on historical neighbors sampling heuristics such as uniform sampling or recent neighbors selection. These heuristics are static and fail to adapt to the underlying graph structure. We introduce FLASH, a learnable and graph-adaptive neighborhood selection mechanism that generalizes existing heuristics. FLASH integrates seamlessly into TGNNs and is trained end-to-end using a self-supervised ranking loss. We provide theoretical evidence that commonly used heuristics hinder TGNNs performance, motivating our design. Extensive experiments across multiple benchmarks demonstrate consistent and significant performance improvements for TGNNs equipped with FLASH
Unlike test-time adaptation (TTA), open-set TTA (OSTTA) aims to robustly adapt to in-distribution (ID) domain shifts while suppressing the detrimental impact of out-of-distribution (OOD) samples encountered at test time. To this end, we propose a novel OSTTA approach named Noise-Immune Self-Purification (NISP). By exploiting the zero-shot priors of pre-trained vision-language models (VLMs), NISP advances from coarse pseudo-labeling to fine-grained, noise-resilient adaptation. Technically, we first introduce a dual-cluster Bayesian Gaussian mixture model to fit VLMs-derived scores, achieving coarse pseudo-ID/OOD separation via posterior-risk thresholding. Subsequently, NISP constructs a noise-immune fine-grained adaptation where the adapter enforces consensus-discrepancy constraints to refine coarse pseudo-labels and suppress noise propagation. We then devise the Jaccard Consistency Score for OOD discrimination. Overall the coarse-to-fine pipeline enables rigorous self-purification and robust online adaptation. Theoretically we show that consensus-discrepancy losses mitigate the deleterious effects of noise. Empirically, NISP achieves state-of-the-art results across multiple OSTTA benchmarks, validating its efficacy. The code is available at https://github.com/njustkmg/IJCAI26-NISP.
Deep multi-view clustering draws plentiful attention in various domains, owing to remarkable performance in learning patterns from complementary information of multi-view data. However, previous methods encounter two challenges. They utilize a single pre-defined clustering strategy to perceive diverse structures from data in multiple views subject to heterogeneous distributions, failing to fully capture intricate complementary structures. They leverage either the feature fusion or the result fusion in clustering, which cannot fully integrate complementary information. Therefore, a decoupled multi-agent actor-critic (PRIME) is proposed via defining multi-view clustering as a partially observable Markov game, which establishes a dynamic trial-and-error assignment process of samples in a decentralized perception with centralized learning framework to adaptively learn the optimal clustering strategy. In PRIME, an actor leverages the policy gradient paradigm to independently implement a Markov decision process of data partition in a view, which fully explores data structures to tailor a clustering sub-policy for a view. Meanwhile, a critic utilizes the value function paradigm to centrally guide the Markov game among actors in different views, which constructs both feature and result fusion to progressively enhance complementary knowledge integration for robust clustering results. Extensive experiments on 6 benchmark datasets verify the superiority of PRIME against 10 methods.
The three-dimensional bin packing problem (3D-BPP) is widely applied in logistics and warehousing. However, existing learning-based approaches often neglect stability constraints and struggle to generalize across diverse bin dimensions. To address this, we propose a novel deep reinforcement learning framework, One4Many-StablePacker (O4M-SP). The primary advantage of O4M-SP lies in its ability to handle variable bin dimensions in a single training process while explicitly enforcing two types of practical stability constraints: support constraints, which ensure an item's bottom center lies within the convex hull of the underlying contact area, and weight constraints, which restrict the total vertical load on an item to its bearing capacity. Our training method introduces two innovative mechanisms. First, a weighted reward function integrates the loading rate with a novel height difference metric for packing layouts, promoting improved bin utilization via flatter packing configurations. Second, clipped policy gradient optimization with tailored policy drifting mitigates entropy collapse, encouraging exploration at critical decision nodes during packing to prevent premature convergence. Extensive experiments demonstrate that O4M-SP generalizes effectively across diverse bin dimensions and significantly outperforms baseline methods. Furthermore, the framework exhibits strong practical applicability by solving complex scenarios under strict stability constraints.
Time series forecasting is vital in diverse sectors such as energy and transportation, where non-stationary dynamics are deeply intertwined with external events in other modalities such as texts. However, incorporating natural language-based external events to improve non-stationary forecasting remains largely unexplored, as most approaches still rely on a single modality, resulting in limited contextual knowledge and model underperformance. Enabling fine-grained multimodal interactions between temporal and textual data is challenged by two fundamental issues: (1) the gap in modeling interactions among discrete external events and continuous time series in a unified framework; (2) classical uniform diffusion timestep ignores event-induced non-stationary variability, leading to imbalanced denoising difficulty across diffusion stages. In this work, we propose event-aware non-stationary time series forecasting EventTSF, an autoregressive diffusion framework that integrates historical time series and textual events via step-wise diffusion. To mitigate the imbalanced denoising difficulty of uniform timestep sampling, EventTSF uses an event-aware flow-matching timestep conditioned on event semantics. Extensive experiments on 7 synthetic and real-world datasets show that EventTSF outperforms 12 non-stationary time series forecasting baselines, achieving average gains of 41.3% in probabilistic forecasting and 27.5% in deterministic forecasting across all evaluation metrics.
Inverse reinforcement learning (IRL) has made significant progress in recovering reward functions from expert demonstrations. However, a key challenge remains: how to extract reward functions that generalize across related but distinct tasks. In this paper, we address this by focusing on transferable IRL, learning intrinsic rewards that can drive effective behavior in unseen but structurally aligned environments. Our method leverages a variational autoencoder to learn an abstract representation of the state space shared across multiple source tasks. This abstracted space captures high-level features that are invariant across tasks, enabling the learning of a unified abstract reward function. The learned reward is then used to train policies in a separate, previously unseen target task without requiring new demonstrations in the target task. We evaluate our approach on multiple environments from Gymnasium and AssistiveGym, demonstrating that the learned abstract rewards consistently support successful policy learning in novel task settings.
Federated knowledge distillation (Fed-KD) exchanges distilled predictions on a shared proxy dataset, often reducing communication and accommodating heterogeneous client architectures. However, in multimodal federated learning (MFL), modality dominance can bias local optimization and contaminate the aggregated global teacher, degrading both multimodal accuracy and robustness to missing modalities, especially under non-IID heterogeneity. We propose Federated Adaptive-Modality Balancing (FedAMB). At the client side, Selective-Modality Regulation (SMR) models dominance as a state-dependent phenomenon and intervenes only when it destabilizes training, strengthening weak modalities without over-regularization. At the server side, Component-wise Modality Distillation (CMD) regulates how aggregated knowledge is transferred to each modality branch, preventing the propagation of fusion-biased teachers while directly improving unimodal representations. Experiments on CREMA-D, AVE, and UR-FUNNY show that FedAMB consistently improves multimodal accuracy and missing-modality robustness.