Large Language Models (LLMs) often show reduced performance, cultural alignment, and safety robustness in non-English languages, partly because English dominates both pre-training data and human preference alignment datasets. Training methods like Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO) require human preference data, which remains scarce and largely non-public for many languages beyond English. To address this gap, we introduce compar:IA, an open-source digital public service developed inside the French government and designed to collect large-scale human preference data from a predominantly French-speaking general audience. The platform uses a blind pairwise comparison interface to capture unconstrained, real-world prompts and user judgments across a diverse set of language models, while maintaining low participation friction and privacy-preserving automated filtering. As of 2026-02-07, compar:IA has collected over 600,000 free-form prompts and 250,000 preference votes, with approximately 89% of the data in French. We release three complementary datasets -- conversations, votes, and reactions -- under open licenses, and present initial analyses, including a French-language model leaderboard and user interaction patterns. Beyond the French context, compar:IA is evolving toward an international digital public good, offering reusable infrastructure for multilingual model training, evaluation, and the study of human-AI interaction.

Post-training improves instruction-following and helpfulness of large language models (LLMs) but often reduces generation diversity, which leads to repetitive outputs in open-ended settings, a phenomenon known as mode collapse. Motivated by evidence that LLM layers play distinct functional roles, we hypothesize that mode collapse can be localized to specific layers and that restoring a carefully chosen range of layers to their pre-trained weights can recover diversity while maintaining high output quality. To validate this hypothesis and decide which layers to restore, we design a proxy task -- Constrained Random Character(CRC) -- with an explicit validity set and a natural diversity objective. Results on CRC reveal a clear diversity-validity trade-off across restoration ranges and identify configurations that increase diversity with minimal quality loss. Based on these findings, we propose Selective Layer Restoration (SLR), a training-free method that restores selected layers in a post-trained model to their pre-trained weights, yielding a hybrid model with the same architecture and parameter count, incurring no additional inference cost. Across three different tasks (creative writing, open-ended question answering, and multi-step reasoning) and three different model families (Llama, Qwen, and Gemma), we find SLR can consistently and substantially improve output diversity while maintaining high output quality.

Developing robotic manipulation policies is iterative and hypothesis-driven: researchers test tactile sensing, gripper geometries, and sensor placements through real-world data collection and training. Yet even minor end-effector changes often require mechanical refitting and system re-integration, slowing iteration. We present RAPID, a full-stack reconfigurable platform designed to reduce this friction. RAPID is built around a tool-free, modular hardware architecture that unifies handheld data collection and robot deployment, and a matching software stack that maintains real-time awareness of the underlying hardware configuration through a driver-level Physical Mask derived from USB events. This modular hardware architecture reduces reconfiguration to seconds and makes systematic multi-modal ablation studies practical, allowing researchers to sweep diverse gripper and sensing configurations without repeated system bring-up. The Physical Mask exposes modality presence as an explicit runtime signal, enabling auto-configuration and graceful degradation under sensor hot-plug events, so policies can continue executing when sensors are physically added or removed. System-centric experiments show that RAPID reduces the setup time for multi-modal configurations by two orders of magnitude compared to traditional workflows and preserves policy execution under runtime sensor hot-unplug events. The hardware designs, drivers, and software stack are open-sourced at https://rapid-kit.github.io/ .

The robustness of Vision Language Models (VLMs) is commonly assessed through output-level invariance, implicitly assuming that stable predictions reflect stable multimodal processing. In this work, we argue that this assumption is insufficient. We introduce a representation-aware and frequency-aware evaluation framework that measures internal embedding drift, spectral sensitivity, and structural smoothness (spatial consistency of vision tokens), alongside standard label-based metrics. Applying this framework to modern VLMs across the SEEDBench, MMMU, and POPE datasets reveals three distinct failure modes. First, models frequently preserve predicted answers while undergoing substantial internal representation drift; for perturbations such as text overlays, this drift approaches the magnitude of inter-image variability, indicating that representations move to regions typically occupied by unrelated inputs despite unchanged outputs. Second, robustness does not improve with scale; larger models achieve higher accuracy but exhibit equal or greater sensitivity, consistent with sharper yet more fragile decision boundaries. Third, we find that perturbations affect tasks differently: they harm reasoning when they disrupt how models combine coarse and fine visual cues, but on the hallucination benchmarks, they can reduce false positives by making models generate more conservative answers.

Large Language Models (LLMs) face a fundamental safety-helpfulness trade-off due to static, one-size-fits-all safety policies that lack runtime controllabilityxf, making it difficult to tailor responses to diverse application needs. %As a result, models may over-refuse benign requests or under-constrain harmful ones. We present \textbf{PACT} (Prompt-configured Action via Chain-of-Thought), a framework for dynamic safety control through explicit, risk-aware reasoning. PACT operates under a hierarchical policy architecture: a non-overridable global safety policy establishes immutable boundaries for critical risks (e.g., child safety, violent extremism), while user-defined policies can introduce domain-specific (non-global) risk categories and specify label-to-action behaviors to improve utility in real-world deployment settings. The framework decomposes safety decisions into structured Classify$\rightarrow$Act paths that route queries to the appropriate action (comply, guide, or reject) and render the decision-making process transparent. Extensive experiments demonstrate that PACT achieves near state-of-the-art safety performance under global policy evaluation while attaining the best controllability under user-specific policy evaluation, effectively mitigating the safety-helpfulness trade-off. We will release the PACT model suite, training data, and evaluation protocols to facilitate reproducible research in controllable safety alignment.

Spoken content, such as online videos and podcasts, often spans multiple topics, which makes automatic topic segmentation essential for user navigation and downstream applications. However, current methods do not fully leverage acoustic features, leaving room for improvement. We propose a multi-modal approach that fine-tunes both a text encoder and a Siamese audio encoder, capturing acoustic cues around sentence boundaries. Experiments on a large-scale dataset of YouTube videos show substantial gains over text-only and multi-modal baselines. Our model also proves more resilient to ASR noise and outperforms a larger text-only baseline on three additional datasets in Portuguese, German, and English, underscoring the value of learned acoustic features for robust topic segmentation.

Standard trust-region methods constrain policy updates via Kullback-Leibler (KL) divergence. However, KL controls only an average divergence and does not directly prevent rare, large likelihood-ratio excursions that destabilize training--precisely the failure mode that motivates heuristics such as PPO's clipping. We propose overlap geometry as an alternative trust region, constraining distributional overlap via the Bhattacharyya coefficient (closely related to the Hellinger/Renyi-1/2 geometry). This objective penalizes separation in the ratio tails, yielding tighter control over likelihood-ratio excursions without relying on total variation bounds that can be loose in tail regimes. We derive Bhattacharyya-TRPO (BTRPO) and Bhattacharyya-PPO (BPPO), enforcing overlap constraints via square-root ratio updates: BPPO clips the square-root ratio q = sqrt(r), and BTRPO applies a quadratic Hellinger/Bhattacharyya penalty. Empirically, overlap-based updates improve robustness and aggregate performance as measured by RLiable under matched training budgets, suggesting overlap constraints as a practical, principled alternative to KL for stable policy optimization.

Existing LLM-as-a-Judge systems suffer from three fundamental limitations: limited adaptivity to task- and domain-specific evaluation criteria, systematic biases driven by non-semantic cues such as position, length, format, and model provenance, and evaluation inconsistency that leads to contradictory judgments across different evaluation modes (e.g., pointwise versus pairwise). To address these issues, we propose FairJudge, an adaptive, debiased, and consistent LLM-as-a-Judge. Unlike prior approaches that treat the judge as a static evaluator, FairJudge models judging behavior itself as a learnable and regularized policy. From a data-centric perspective, we construct a high-information-density judging dataset that explicitly injects supervision signals aligned with evaluation behavior. Building on this dataset, we adopt a curriculum-style SFT-DPO-GRPO training paradigm that progressively aligns rubric adherence, bias mitigation, and cross-mode consistency, while avoiding catastrophic forgetting. Experimental results on multiple internal and public benchmarks show that FairJudge consistently improves agreement and F1, reduces non-semantic biases, and outperforms substantially larger instruction-tuned LLMs. All resources will be publicly released after acceptance to facilitate future research.

Large Language Models (LLMs) are powerful text generators, yet they can produce toxic or harmful content even when given seemingly harmless prompts. This presents a serious safety challenge and can cause real-world harm. Toxicity is often subtle and context-dependent, making it difficult to detect at the token level or through coarse sentence-level signals. Moreover, efforts to mitigate toxicity often face a trade-off between safety and the coherence, or fluency of the generated text. In this work, we present a targeted subspace intervention strategy for identifying and suppressing hidden toxic patterns from underlying model representations, while preserving overall ability to generate safe fluent content. On the RealToxicityPrompts, our method achieves strong mitigation performance compared to existing baselines, with minimal impact on inference complexity. Across multiple LLMs, our approach reduces toxicity of state-of-the-art detoxification systems by 8-20%, while maintaining comparable fluency. Through extensive quantitative and qualitative analyses, we show that our approach achieves effective toxicity reduction without impairing generative performance, consistently outperforming existing baselines.

Surgical phase recognition is a critical component for context-aware decision support in intelligent operating rooms, yet training robust models is hindered by limited annotated clinical videos and large domain gaps between synthetic and real surgical data. To address this, we propose CauCLIP, a causality-inspired vision-language framework that leverages CLIP to learn domain-invariant representations for surgical phase recognition without access to target domain data. Our approach integrates a frequency-based augmentation strategy to perturb domain-specific attributes while preserving semantic structures, and a causal suppression loss that mitigates non-causal biases and reinforces causal surgical features. These components are combined in a unified training framework that enables the model to focus on stable causal factors underlying surgical workflows. Experiments on the SurgVisDom hard adaptation benchmark demonstrate that our method substantially outperforms all competing approaches, highlighting the effectiveness of causality-guided vision-language models for domain-generalizable surgical video understanding.

Vision Transformers (ViTs) have become a dominant architecture in computer vision, yet producing stable and high-resolution attribution maps for these models remains challenging. Architectural components such as patch embeddings and attention routing often introduce structured artifacts in pixel-level explanations, causing many existing methods to rely on coarse patch-level attributions. We introduce DAVE \textit{(\underline{D}istribution-aware \underline{A}ttribution via \underline{V}iT Gradient D\underline{E}composition)}, a mathematically grounded attribution method for ViTs based on a structured decomposition of the input gradient. By exploiting architectural properties of ViTs, DAVE isolates locally equivariant and stable components of the effective input--output mapping. It separates these from architecture-induced artifacts and other sources of instability.

Accurate prediction of outcomes is crucial for clinical decision-making and personalized patient care. Supervised machine learning algorithms, which are commonly used for outcome prediction in the medical domain, optimize for predictive accuracy, which can result in models latching onto spurious correlations instead of robust predictors. Causal structure learning methods on the other hand have the potential to provide robust predictors for the target, but can be too conservative because of algorithmic and data assumptions, resulting in loss of diagnostic precision. Therefore, we propose a novel model-agnostic regularization strategy, Adaptive-CaRe, for generalized outcome prediction in the medical domain. Adaptive-CaRe strikes a balance between both predictive value and causal robustness by incorporating a penalty that is proportional to the difference between the estimated statistical contribution and estimated causal contribution of the input features for model predictions. Our experiments on synthetic data establish the efficacy of the proposed Adaptive-CaRe regularizer in finding robust predictors for the target while maintaining competitive predictive accuracy. With experiments on a standard causal benchmark, we provide a blueprint for navigating the trade-off between predictive accuracy and causal robustness by tweaking the regularization strength, $λ$. Validation using real-world dataset confirms that the results translate to practical, real-domain settings. Therefore, Adaptive-CaRe provides a simple yet effective solution to the long-standing trade-off between predictive accuracy and causal robustness in the medical domain. Future work would involve studying alternate causal structure learning frameworks and complex classification models to provide deeper insights at a larger scale.

Background: High-level system testing of applications that use data from e-Government services as input requires test data that is real-life-like but where the privacy of personal information is guaranteed. Applications with such strong requirement include information exchange between countries, medicine, banking, etc. This review aims to synthesize the current state-of-the-practice in this domain. Objectives: The objective of this Systematic Review is to identify existing approaches for creating and evolving synthetic test data without using real-life raw data. Methods: We followed well-known methodologies for conducting systematic literature reviews, including the ones from Kitchenham as well as guidelines for analysing the limitations of our review and its threats to validity. Results: A variety of methods and tools exist for creating privacy-preserving test data. Our search found 1,013 publications in IEEE Xplore, ACM Digital Library, and SCOPUS. We extracted data from 75 of those publications and identified 37 approaches that answer our research question partly. A common prerequisite for using these methods and tools is direct access to real-life data for data anonymization or synthetic test data generation. Nine existing synthetic test data generation approaches were identified that were closest to answering our research question. Nevertheless, further work would be needed to add the ability to evolve synthetic test data to the existing approaches. Conclusions: None of the publications really covered our requirements completely, only partially. Synthetic test data evolution is a field that has not received much attention from researchers but needs to be explored in Digital Government Solutions, especially since new legal regulations are being placed in force in many countries.

Speech tokenizers are foundational to speech language models, yet existing approaches face two major challenges: (1) balancing trade-offs between encoding semantics for understanding and acoustics for reconstruction, and (2) achieving low bit rates and low token rates. We propose Speech Diffusion Tokenizer (SiTok), a diffusion autoencoder that jointly learns semantic-rich representations through supervised learning and enables high-fidelity audio reconstruction with diffusion. We scale SiTok to 1.6B parameters and train it on 2 million hours of speech. Experiments show that SiTok outperforms strong baselines on understanding, reconstruction and generation tasks, at an extremely low token rate of $12.5$ Hz and a bit-rate of 200 bits-per-second.

Test-time compute allocation in large reasoning models (LRMs) is widely used and has applications in mathematical problem solving, code synthesis, and planning. Recent work has addressed this problem by scaling self-consistency and parallel thinking, adding generic ``thinking tokens'' and prompting models to re-read the question before answering. Unfortunately, these approaches either inject task-agnostic tokens or mandate heuristics that do not explain -- and often ignore -- the \emph{spontaneous} repetition that many LRMs exhibit at the head of their internal chains. In contrast, we analyze and harness the model's tendency to restate the question, which we term the \emph{Echo of Prompt (EOP)}, as a front-loaded, compute-shaping mechanism. We formalize its probabilistic cost by casting echo removal as rejection-based conditioning and defining the \emph{Echo Likelihood Gap} $Δ\mathcal{L}$ as a computable proxy. This provides the missing theoretical link that links early repetition to likelihood gains and downstream accuracy. However, it does not by itself specify how to exploit EOP. Consequently, we develop \emph{Echo-Distilled SFT (ED-SFT)} to instill an ``echo-then-reason'' pattern through supervised finetuning, and \emph{Echoic Prompting (EP)} to re-ground the model mid-trace without training. While promising, quantifying benefits beyond verbosity is non-trivial. Therefore, we conduct length and suffix-controlled likelihood analyses together with layer-wise attention studies, showing that EOP increases answer to answer-prefix attention in middle layers, consistent with an \emph{attention refocusing} mechanism. We evaluate on GSM8K, MathQA, Hendrycks-MATH, AIME24, and MATH-500 under identical decoding settings and budgets, and find consistent gains over baselines. Code is available at https://github.com/hhh2210/echoes-as-anchors.

While generative modeling on time series facilitates more capable and flexible probabilistic forecasting, existing generative time series models do not address the multi-dimensional properties of time series data well. The prevalent architecture of Diffusion Transformers (DiT), which relies on simplistic conditioning controls and a single-stream Transformer backbone, tends to underutilize cross-variate dependencies in covariate-aware forecasting. Inspired by Multimodal Diffusion Transformers that integrate textual guidance into video generation, we propose Diffusion Transformers for Time Series (DiTS), a general-purpose architecture that frames endogenous and exogenous variates as distinct modalities. To better capture both inter-variate and intra-variate dependencies, we design a dual-stream Transformer block tailored for time-series data, comprising a Time Attention module for autoregressive modeling along the temporal dimension and a Variate Attention module for cross-variate modeling. Unlike the common approach for images, which flattens 2D token grids into 1D sequences, our design leverages the low-rank property inherent in multivariate dependencies, thereby reducing computational costs. Experiments show that DiTS achieves state-of-the-art performance across benchmarks, regardless of the presence of future exogenous variate observations, demonstrating unique generative forecasting strengths over traditional deterministic deep forecasting models.

Prototypical parts-based models offer a "this looks like that" paradigm for intrinsic interpretability, yet they typically struggle with ImageNet-scale generalization and often require computationally expensive backbone finetuning. Furthermore, existing methods frequently suffer from "prototype drift," where learned prototypes lack tangible grounding in the training distribution and change their activation under small perturbations. We present ProtoQuant, a novel architecture that achieves prototype stability and grounded interpretability through latent vector quantization. By constraining prototypes to a discrete learned codebook within the latent space, we ensure they remain faithful representations of the training data without the need to update the backbone. This design allows ProtoQuant to function as an efficient, interpretable head that scales to large-scale datasets. We evaluate ProtoQuant on ImageNet and several fine-grained benchmarks (CUB-200, Cars-196). Our results demonstrate that ProtoQuant achieves competitive classification accuracy while generalizing to ImageNet and comparable interpretability metrics to other prototypical-parts-based methods.

The task of establishing correspondences between two 3D shapes is a long-standing challenge in computer vision. While numerous studies address full-full and partial-full 3D shape matching, only a limited number of works have explored the partial-partial setting, very likely due to its unique challenges: we must compute accurate correspondences while at the same time find the unknown overlapping region. Nevertheless, partial-partial 3D shape matching reflects the most realistic setting, as in many real-world cases, such as 3D scanning, shapes are only partially observable. In this work, we introduce the first integer linear programming approach specifically designed to address the distinctive challenges of partial-partial shape matching. Our method leverages geometric consistency as a strong prior, enabling both robust estimation of the overlapping region and computation of neighbourhood-preserving correspondences. We empirically demonstrate that our approach achieves high-quality matching results both in terms of matching error and smoothness. Moreover, we show that our method is more scalable than previous formalisms.

Centralized training is the standard paradigm in deep learning, enabling models to learn from a unified dataset in a single location. In such setup, isotropic feature distributions naturally arise as a mean to support well-structured and generalizable representations. In contrast, continual learning operates on streaming and non-stationary data, and trains models incrementally, inherently facing the well-known plasticity-stability dilemma. In such settings, learning dynamics tends to yield increasingly anisotropic feature space. This arises a fundamental question: should isotropy be enforced to achieve a better balance between stability and plasticity, and thereby mitigate catastrophic forgetting? In this paper, we investigate whether promoting feature-space isotropy can enhance representation quality in continual learning. Through experiments using contrastive continual learning techniques on CIFAR-10 and CIFAR-100 data, we find that isotropic regularization fails to improve, and can in fact degrade, model accuracy in continual settings. Our results highlight essential differences in feature geometry between centralized and continual learning, suggesting that isotropy, while beneficial in centralized setups, may not constitute an appropriate inductive bias for non-stationary learning scenarios.

Weight initialization plays a crucial role in the optimization behavior and convergence efficiency of neural networks. Most existing initialization methods, such as Xavier and Kaiming initializations, rely on random sampling and do not exploit information from the optimization process itself. We propose a simple, yet effective, initialization strategy based on self-supervised pre-training using random noise as the target. Instead of directly training the network from random weights, we first pre-train it to fit random noise, which leads to a structured and non-random parameter configuration. We show that this noise-driven pre-training significantly improves convergence speed in subsequent tasks, without requiring additional data or changes to the network architecture. The proposed method is particularly effective for implicit neural representations (INRs) and Deep Image Prior (DIP)-style networks, which are known to exhibit a strong low-frequency bias during optimization. After noise-based pre-training, the network is able to capture high-frequency components much earlier in training, leading to faster and more stable convergence. Although random noise contains no semantic information, it serves as an effective self-supervised signal (considering its white spectrum nature) for shaping the initialization of neural networks. Overall, this work demonstrates that noise-based pre-training offers a lightweight and general alternative to traditional random initialization, enabling more efficient optimization of deep neural networks.

Standard chain-of-thought reasoning generates a solution in a single forward pass, committing irrevocably to each token and lacking a mechanism to recover from early errors. We introduce Inference-Time Rethinking, a generative framework that enables iterative self-correction by decoupling declarative latent thought vectors from procedural generation. We factorize reasoning into a continuous latent thought vector (what to reason about) and a decoder that verbalizes the trace conditioned on this vector (how to reason). Beyond serving as a declarative buffer, latent thought vectors compress the reasoning structure into a continuous representation that abstracts away surface-level token variability, making gradient-based optimization over reasoning strategies well-posed. Our prior model maps unstructured noise to a learned manifold of valid reasoning patterns, and at test time we employ a Gibbs-style procedure that alternates between generating a candidate trace and optimizing the latent vector to better explain that trace, effectively navigating the latent manifold to refine the reasoning strategy. Training a 0.2B-parameter model from scratch on GSM8K, our method with 30 rethinking iterations surpasses baselines with 10 to 15 times more parameters, including a 3B counterpart. This result demonstrates that effective mathematical reasoning can emerge from sophisticated inference-time computation rather than solely from massive parameter counts.

Adversarial attacks against deep neural networks are commonly constructed under $\ell_p$ norm constraints, most often using $p=1$, $p=2$ or $p=\infty$, and potentially regularized for specific demands such as sparsity or smoothness. These choices are typically made without a systematic investigation of how the norm parameter \( p \) influences the structural and perceptual properties of adversarial perturbations. In this work, we study how the choice of \( p \) affects sparsity and smoothness of adversarial attacks generated under \( \ell_p \) norm constraints for values of $p \in [1,2]$. To enable a quantitative analysis, we adopt two established sparsity measures from the literature and introduce three smoothness measures. In particular, we propose a general framework for deriving smoothness measures based on smoothing operations and additionally introduce a smoothness measure based on first-order Taylor approximations. Using these measures, we conduct a comprehensive empirical evaluation across multiple real-world image datasets and a diverse set of model architectures, including both convolutional and transformer-based networks. We show that the choice of $\ell_1$ or $\ell_2$ is suboptimal in most cases and the optimal $p$ value is dependent on the specific task. In our experiments, using $\ell_p$ norms with $p\in [1.3, 1.5]$ yields the best trade-off between sparse and smooth attacks. These findings highlight the importance of principled norm selection when designing and evaluating adversarial attacks.

Complex-valued neural networks (CVNNs) are rising in popularity for all kinds of applications. To safely use CVNNs in practice, analyzing their robustness against outliers is crucial. One well known technique to understand the behavior of deep neural networks is to investigate their behavior under adversarial attacks, which can be seen as worst case minimal perturbations. We design Phase Attacks, a kind of attack specifically targeting the phase information of complex-valued inputs. Additionally, we derive complex-valued versions of commonly used adversarial attacks. We show that in some scenarios CVNNs are more robust than RVNNs and that both are very susceptible to phase changes with the Phase Attacks decreasing the model performance more, than equally strong regular attacks, which can attack both phase and magnitude.

Vision-language-action (VLA) models typically inject proprioception only as a late conditioning signal, which prevents robot state from shaping instruction understanding and from influencing which visual tokens are attended throughout the policy. We introduce ThinkProprio, which converts proprioception into a sequence of text tokens in the VLM embedding space and fuses them with the task instruction at the input. This early fusion lets embodied state participate in subsequent visual reasoning and token selection, biasing computation toward action-critical evidence while suppressing redundant visual tokens. In a systematic ablation over proprioception encoding, state entry point, and action-head conditioning, we find that text tokenization is more effective than learned projectors, and that retaining roughly 15% of visual tokens can match the performance of using the full token set. Across CALVIN, LIBERO, and real-world manipulation, ThinkProprio matches or improves over strong baselines while reducing end-to-end inference latency over 50%.

Chemical exchange saturation transfer (CEST) MRI is a non-invasive imaging modality for detecting metabolites. It offers higher resolution and sensitivity compared to conventional magnetic resonance spectroscopy (MRS). However, quantification of CEST data is challenging because the measured signal results from a complex interplay of many physiological variables. Here, we introduce a transformer-based neural network to fit parameters such as metabolite concentrations, exchange and relaxation rates of a physical model derived from Bloch-McConnell equations to in-vitro CEST spectra. We show that our self-supervised trained neural network clearly outperforms the solution of classical gradient-based solver.

Autonomous agents that perform everyday manipulation actions need to ensure that their body motions are semantically correct with respect to a task request, causally effective within their environment, and feasible for their embodiment. In order to enable robots to verify these properties, we introduce the Law of Task-Achieving Body Motion as an axiomatic correctness specification for body motions. To that end we introduce scoped Task-Environment-Embodiment (TEE) classes that represent world states as Semantic Digital Twins (SDTs) and define applicable physics models to decompose task achievement into three predicates: SatisfiesRequest for semantic request satisfaction over SDT state evolution; Causes for causal sufficiency under the scoped physics model; and CanPerform for safety and feasibility verification at the embodiment level. This decomposition yields a reusable, implementation-independent interface that supports motion synthesis and the verification of given body motions. It also supports typed failure diagnosis (semantic, causal, embodiment and out-of-scope), feasibility across robots and environments, and counterfactual reasoning about robot body motions. We demonstrate the usability of the law in practice by instantiating it for articulated container manipulation in kitchen environments on three contrasting mobile manipulation platforms

We introduce Baichuan-M3, a medical-enhanced large language model engineered to shift the paradigm from passive question-answering to active, clinical-grade decision support. Addressing the limitations of existing systems in open-ended consultations, Baichuan-M3 utilizes a specialized training pipeline to model the systematic workflow of a physician. Key capabilities include: (i) proactive information acquisition to resolve ambiguity; (ii) long-horizon reasoning that unifies scattered evidence into coherent diagnoses; and (iii) adaptive hallucination suppression to ensure factual reliability. Empirical evaluations demonstrate that Baichuan-M3 achieves state-of-the-art results on HealthBench, the newly introduced HealthBench-Hallu and ScanBench, significantly outperforming GPT-5.2 in clinical inquiry, advisory and safety. The models are publicly available at https://huggingface.co/collections/baichuan-inc/baichuan-m3.

We study the online resource allocation problem in which at each round, a budget $B$ must be allocated across $K$ arms under censored feedback. An arm yields a reward if and only if two conditions are satisfied: (i) the arm is activated according to an arm-specific Bernoulli random variable with unknown parameter, and (ii) the allocated budget exceeds a random threshold drawn from a parametric distribution with unknown parameter. Over $T$ rounds, the learner must jointly estimate the unknown parameters and allocate the budget so as to maximize cumulative reward facing the exploration--exploitation trade-off. We prove an information-theoretic regret lower bound $Ω(T^{1/3})$, demonstrating the intrinsic difficulty of the problem. We then propose RA-UCB, an optimistic algorithm that leverages non-trivial parameter estimation and confidence bounds. When the budget $B$ is known at the beginning of each round, RA-UCB achieves a regret of order $\widetilde{\mathcal{O}}(\sqrt{T})$, and even $\mathcal{O}(\mathrm{poly}\text{-}\log T)$ under stronger assumptions. As for unknown, round dependent budget, we introduce MG-UCB, which allows within-round switching and infinitesimal allocations, and matches the regret guarantees of RA-UCB. We then validate our theoretical results through experiments on real-world datasets.

Graph Neural Networks (GNNs) have established themselves as the leading models for learning on graph-structured data, generally categorized into spatial and spectral approaches. Central to these architectures is the Graph Shift Operator (GSO), a matrix representation of the graph structure used to filter node signals. However, selecting the optimal GSO, whether fixed or learnable, remains largely empirical. In this paper, we introduce a novel alignment gain metric that quantifies the geometric distortion between the input signal and label subspaces. Crucially, our theoretical analysis connects this alignment directly to generalization bounds via a spectral proxy for the Lipschitz constant. This yields a principled, computation-efficient criterion to rank and select the optimal GSO for any prediction task prior to training, eliminating the need for extensive search.

Reliable benchmarking is critical for advancing Vision-Language-Action (VLA) models, as it reveals their generalization, robustness, and alignment of perception with language-driven manipulation tasks. However, existing benchmarks often provide limited or misleading assessments due to insufficient evaluation protocols that inadequately capture real-world distribution shifts. This work systematically rethinks VLA benchmarking from both evaluation and data perspectives, introducing LIBERO-X, a benchmark featuring: 1) A hierarchical evaluation protocol with progressive difficulty levels targeting three core capabilities: spatial generalization, object recognition, and task instruction understanding. This design enables fine-grained analysis of performance degradation under increasing environmental and task complexity; 2) A high-diversity training dataset collected via human teleoperation, where each scene supports multiple fine-grained manipulation objectives to bridge the train-evaluation distribution gap. Experiments with representative VLA models reveal significant performance drops under cumulative perturbations, exposing persistent limitations in scene comprehension and instruction grounding. By integrating hierarchical evaluation with diverse training data, LIBERO-X offers a more reliable foundation for assessing and advancing VLA development.