Agents in cyber-physical systems are increasingly entrusted with safety-critical tasks. Ensuring safety of these agents often requires localizing the pose for subsequent actions. Pose estimates can, e.g., be obtained from various combinations of lidar sensors, cameras, and external services such as GPS. Crucially, in safety-critical domains, a rough estimate is insufficient to formally determine safety, i.e., guaranteeing safety even in the worst-case scenario, and external services might additionally not be trustworthy. We address this problem by presenting a certified pose estimation in 3D solely from a camera image and a well-known target geometry. This is realized by formally bounding the pose, which is computed by leveraging recent results from reachability analysis and formal neural network verification. Our experiments demonstrate that our approach efficiently and accurately localizes agents in both synthetic and real-world experiments.

Real-time whole-body teleoperation is a critical method for humanoid robots to perform complex tasks in unstructured environments. However, developing a unified controller that robustly supports diverse human motions remains a significant challenge. Existing methods typically distill multiple expert policies into a single general policy, which often inevitably leads to performance degradation, particularly on highly dynamic motions. This paper presents TeleGate, a unified whole-body teleoperation framework for humanoid robots that achieves high-precision tracking across various motions while avoiding the performance loss inherent in knowledge distillation. Our key idea is to preserve the full capability of domain-specific expert policies by training a lightweight gating network, which dynamically activates experts in real-time based on proprioceptive states and reference trajectories. Furthermore, to compensate for the absence of future reference trajectories in real-time teleoperation, we introduce a VAE-based motion prior module that extracts implicit future motion intent from historical observations, enabling anticipatory control for motions requiring prediction such as jumping and standing up. We conducted empirical evaluations in simulation and also deployed our technique on the Unitree G1 humanoid robot. Using only 2.5 hours of motion capture data for training, our TeleGate achieves high-precision real-time teleoperation across diverse dynamic motions (e.g., running, fall recovery, and jumping), significantly outperforming the baseline methods in both tracking accuracy and success rate.

Uncertainty calibration in pre-trained transformers is critical for their reliable deployment in risk-sensitive applications. Yet, most existing pre-trained transformers do not have a principled mechanism for uncertainty propagation through their feature transformation stack. In this work, we propose a diffusion-inspired reconfiguration of transformers in which each feature transformation block is modeled as a probabilistic mapping. Composing these probabilistic mappings reveals a probability path that mimics the structure of a diffusion process, transporting data mass from the input distribution to the pre-trained feature distribution. This probability path can then be recompiled on a diffusion process with a unified transition model to enable principled propagation of representation uncertainty throughout the pre-trained model's architecture while maintaining its original predictive performance. Empirical results across a variety of vision and language benchmarks demonstrate that our method achieves superior calibration and predictive accuracy compared to existing uncertainty-aware transformers.

Large language models (LLMs) excel at complex tasks with advances in reasoning capabilities. However, existing reward mechanisms remain tightly coupled to final correctness and pay little attention to the underlying reasoning process: trajectories with sound reasoning but wrong answers receive low credit, while lucky guesses with flawed logic may be highly rewarded, affecting reasoning generalization. From a causal perspective, we interpret multi-candidate reasoning for a fixed question as a family of counterfactual experiments with theoretical supports. Building on this, we propose Group Causal Counterfactual Policy Optimization to explicitly train LLMs to learn generalizable reasoning patterns. It proposes an episodic causal counterfactual reward that jointly captures (i) robustness, encouraging the answer distribution induced by a reasoning step to remain stable under counterfactual perturbations; and (ii) effectiveness, enforcing sufficient variability so that the learned reasoning strategy can transfer across questions. We then construct token-level advantages from this reward and optimize the policy, encouraging LLMs to favor reasoning patterns that are process-valid and counterfactually robust. Extensive experiments on diverse benchmarks demonstrate its advantages.

Generative policies based on diffusion models and flow matching have shown strong promise for offline reinforcement learning (RL), but their applicability remains largely confined to continuous action spaces. To address a broader range of offline RL settings, we extend flow matching to a general framework that supports discrete action spaces with multiple objectives. Specifically, we replace continuous flows with continuous-time Markov chains, trained using a Q-weighted flow matching objective. We then extend our design to multi-agent settings, mitigating the exponential growth of joint action spaces via a factorized conditional path. We theoretically show that, under idealized conditions, optimizing this objective recovers the optimal policy. Extensive experiments further demonstrate that our method performs robustly across diverse settings and benchmarks, including high-dimensional control, multi-agent games, and dynamically changing preferences over multiple objectives, while outperforming traditional offline RL methods in practical multi-modal decision-making scenarios. Our discrete framework can also be applied to continuous-control problems through action quantization, providing a flexible trade-off between representational complexity and performance.

Many engineering and scientific workflows rely on expensive black-box evaluations, requiring sequential decisions that must both improve task performance and reduce uncertainty. Bayesian optimization (BO) and Bayesian experimental design (BED) provide powerful but largely separate treatments of goal-directed optimization and information-seeking experimentation, leaving limited guidance for hybrid settings in which learning and optimization are intrinsically coupled. We propose Pragmatic Curiosity (PraC), a unified framework for hybrid learning and optimization via active inference. PraC evaluates candidate queries by trading information gain about a task-relevant latent symbol against an expected regret-based potential over outcomes. This formulation exposes three operational design choices: which latent quantity should be clarified, how task value is encoded as regret, and how strongly information gain should be exchanged against pragmatic value. We instantiate PraC across three regimes of increasing complexity: decision-oriented plume monitoring with fixed global symbols and known downstream losses, targeted active search with induced local symbols and evolving coverage goals, and composite Bayesian optimization with hierarchical regret learning under unknown preferences. Across these regimes, PraC reduces downstream decision risk, improves coverage of critical outcome regions, and jointly learns predictive and preference structures without relying on task-specific staging rules.

Reward guidance, also known as posterior sampling, is a popular method for test-time adaptation and post-training in continuous diffusion models. In this paper, we study reward guidance for discrete diffusion language models; now, one cannot differentiate through the natural outputs of the model because they are discrete tokens. We introduce a novel mechanism called EntRGi (Entropy aware Reward Guidance) to address this issue. EntRGi dynamically interpolates between continuous token relaxations and sampled hard tokens, on a token-by-token basis, using the diffusion model's predictive entropy. We demonstrate that EntRGi maintains both reward model reliability and optimization accuracy, while existing approaches sacrifice one for the other. We empirically validate our approach on 7B-parameter diffusion language models across two settings: (1) test-time adaptation, and (2) RGRL (Reward Guided Reinforcement Learning), our recipe for post-training on reward-guided data, showing consistent improvements over state-of-the-art methods. Our code is available at https://atutej.github.io/entrgi-rgrl

Neural operators have emerged as powerful surrogates for the solution of partial differential equations (PDEs), yet their ability to handle general, highly variable boundary conditions (BCs) remains limited. Existing approaches often fail when the solution operator exhibits strong sensitivity to boundary forcings. We propose a general framework for conditioning neural operators on complex non-homogeneous BCs through function extensions. Our key idea is to map boundary data to latent pseudo-extensions defined over the entire spatial domain, enabling any standard operator learning architecture to consume boundary information. The resulting operator, coupled with an arbitrary domain-to-domain neural operator, can learn rich dependencies on complex BCs and input domain functions at the same time. To benchmark this setting, we construct 18 challenging datasets spanning Poisson, linear elasticity, and hyperelasticity problems, with highly variable, mixed-type, component-wise, and multi-segment BCs on diverse geometries. Our approach achieves state-of-the-art accuracy, outperforming baselines by large margins, while requiring no hyperparameter tuning across datasets. Overall, our results demonstrate that learning boundary-to-domain extensions is an effective and practical strategy for imposing complex BCs in existing neural operator frameworks, enabling accurate and robust scientific machine learning models for a broader range of PDE-governed problems.

The choice of activation function fundamentally shapes the representational capacity and parameter efficiency of deep neural networks, yet most widely used activations lack rigorous theoretical guarantees on these properties. We provide a theoretical analysis of DeepBern-Nets (DBNs) -- networks employing learnable Bernstein polynomial activations -- showing that their approximation error decays with the network depth $L$ and the polynomial order $n$ with a rate of $\mathcal{O}(n^{-L})$, exponentially faster than the polynomial rate of ReLU architectures while remaining fully differentiable. We validate these predictions through $1{,}344$ experiments on large scientific datasets (HIGGS and SUSY), comparing DBNs against ReLU, Leaky ReLU, SELU, and GeLU. DBNs achieve over $70\%$ parameter reduction across the majority of architectures -- reaching $99.9\%$ at scale -- converge to ReLU's final loss in as few as $26\%$ of the training epochs, and attain up to $45\%$ lower final loss. These advantages hold over all tested activations, confirming that DBN's gains stem from the learnable polynomial structure rather than mere smoothness.

Vision-language models such as CLIP often struggle to faithfully understand long, detail-rich captions, relying on dominant scene cues while overlooking fine-grained visual evidence. We propose a hierarchical vision-language learning principle for understanding scenes as part-to-whole compositions: before forming a whole-scene representation, a model should uncover what semantic parts appear where in the image. To this end, we propose CAFT (Cross-domain Alignment of Forests and Trees), a vision-language model that jointly learns local text-region alignment at intermediate representations and global image-text alignment at the final representation. Exploiting the organization of long captions, where local descriptions often correspond to scene parts, CAFT employs a fine-to-coarse image encoder and a part-whole text encoder to discover localized part semantics and progressively compose them into a global image-text representation. Trained on 30M image-text pairs, CAFT achieves state-of-the-art performance on six long-text retrieval benchmarks and exhibits strong scaling behavior. Experiments show that CAFT learns fine-grained representations that localize textual semantics in image regions without explicit region-level supervision.

Multi-Agent Discussion (MAD) has garnered increasing attention very recently, where multiple LLM instances collaboratively solve problems via structured discussion. However, we find that current MAD methods easily suffer from discussion inconsistency, LLMs fail to reach a coherent solution, due to the misalignment between their individual contexts.In this paper, we introduce a multi-LLM context learning method (M2CL) that learns a context generator for each agent, capable of dynamically generating context instructions per discussion round via automatic information organization and refinement. Specifically, inspired by our theoretical insights on the context instruction, M2CL train the generators to control context coherence and output discrepancies via a carefully crafted self-adaptive mechanism.It enables LLMs to avoid premature convergence on majority noise and progressively reach the correct consensus. We evaluate M2CL on challenging tasks, including academic reasoning, embodied tasks, and mobile control. The results show that the performance of M2CL significantly surpasses existing methods by 20%--50%, while enjoying favorable transferability and computational efficiency.

Quantization Error Reconstruction (QER) reduces accuracy loss in Post-Training Quantization (PTQ) by approximating weights as $\mathbf{W} \approx \mathbf{Q} + \mathbf{L}\mathbf{R}$, using a rank-$r$ correction to reconstruct quantization error. Prior methods devote the full rank budget to error reconstruction, which is suboptimal when $\mathbf{W}$ has intrinsic low-rank structure and quantization corrupts dominant directions. We propose Structured Residual Reconstruction (SRR), a rank-allocation framework that preserves the top-$k$ singular subspace of the activation-scaled weight before quantization, quantizes only the residual, and uses the remaining rank $r-k$ for error reconstruction. We derive a theory-guided criterion for selecting $k$ by balancing quantization-exposed energy and unrecoverable error under rank constraints. We further show that resulting $\mathbf{Q} + \mathbf{L}\mathbf{R}$ parameterization naturally supports Quantized Parameter-Efficient Fine-Tuning (QPEFT), and stabilizes fine-tuning via gradient scaling along preserved directions. Experiments demonstrate consistent perplexity reductions across diverse models and quantization settings in PTQ, along with a 5.9 percentage-point average gain on GLUE under 2-bit QPEFT. The project page is available at https://ai-isl.github.io/srr.

We study cooperative multi-agent reinforcement learning in the setting of reward-free exploration, where multiple agents jointly explore an unknown MDP in order to learn its dynamics (without observing rewards). We focus on a tabular finite-horizon MDP and adopt a phased learning framework. In each learning phase, multiple agents independently interact with the environment. More specifically, in each learning phase, each agent is assigned a policy, executes it, and observes the resulting trajectory. Our primary goal is to characterize the tradeoff between the number of learning phases and the number of agents, especially when the number of learning phases is small. Our results identify a regime change governed by the horizon $H$. When the number of learning phases equals $H$, we present a computationally efficient algorithm that uses only $\tilde{O}(S^6 H^6 A / ε^2)$ agents to obtain an $ε$ approximation of the dynamics (i.e., yields an $ε$-optimal policy for any reward function). We complement our algorithm with a lower bound showing that any algorithm restricted to $ρ< H$ phases requires at least $A^{H/ρ}$ agents to achieve constant accuracy. Thus, we show that having $Θ(H)$ learning phases is both necessary and sufficient when restricting the number of agents to be polynomial.

Text-to-Image (T2I) diffusion models enable high quality open ended synthesis, but practical use requires suppressing unsafe generations while preserving behavior on benign prompts. We study this tension relative to the frozen generator, using its prompt conditioned distribution as the preservation reference. Since T2I safety is commonly evaluated by bounded risk scores on generated images, total variation (TV) bounds how much expected risk can change from this reference. We call this fixed reference constraint the Safety-Prompt Alignment Tradeoff (SPAT): reducing expected unsafety requires prompt conditioned distributional deviation. To make this deviation selective and adjustable, we define the tau safe set as prompts whose reference risk is at most tau, and cast intervention as projection toward nearby prompts in this set. We propose Selective Prompt prOjecTion (SPOT), an inference time framework that approximates this projection without retraining the generator or learning a category specific rewriter. SPOT uses an LLM to rank candidate rewrites and a safeguard VLM to accept generated images under the same tau. Across four datasets and three diffusion backbones, SPOT achieves relative inappropriate (IP) score reductions from 14.2% to 44.4% over strong safety alignment baselines while keeping benign prompt behavior close to the fixed reference.

Large reasoning models (LRMs) achieve remarkable performance by leveraging reinforcement learning (RL) on reasoning tasks to generate long chain-of-thought (CoT) reasoning. However, this over-optimization often prioritizes compliance, making models vulnerable to harmful prompts. To mitigate this safety degradation, recent approaches rely on external teacher distillation, yet this introduces a distributional discrepancy that degrades native reasoning. We formalize safety realignment as a KL projection onto the safe simplex and prove that the student's own safety-filtered distribution is the unique KL-optimal target, while any external teacher incurs an irreducible excess KL penalty. Guided by this analysis, we propose ThinkSafe, a self-generated alignment framework that restores safety without external teachers. Our key insight is that while compliance suppresses safety mechanisms, models often retain latent knowledge to identify harm. ThinkSafe unlocks this via lightweight refusal steering, which preserves the KL-optimal target while increasing the acceptance rate. Experiments on DeepSeek-R1-Distill and Qwen3 show ThinkSafe significantly improves safety while preserving reasoning proficiency, and achieves superior safety and comparable reasoning to GRPO with roughly an order of magnitude less compute. Code, models, and datasets are available at https://github.com/seanie12/ThinkSafe and https://huggingface.co/Seanie-lee/collections.

Multimodal deep learning (MDL) has achieved remarkable success across various domains, yet its practical deployment is often hindered by incomplete multimodal data. Existing incomplete MDL methods either discard missing modalities, risking the loss of valuable task-relevant information, or recover them, potentially introducing irrelevant noise, leading to the discarding-imputation dilemma. To address this dilemma, in this paper, we propose DyMo, a new inference-time dynamic modality selection framework that adaptively identifies and fuses reliable recovered modalities, fully exploring task-relevant information beyond the conventional discard-or-impute paradigm. Central to DyMo is a novel selection algorithm that maximizes multimodal task-relevant information for each test sample. Since direct estimation of such information at test time is intractable due to the unknown data distribution, we theoretically establish a connection between information and the task loss, which we compute at inference time as a tractable proxy. Building on this, a novel principled reward function is proposed to guide modality selection. In addition, we design a flexible multimodal network architecture compatible with arbitrary modality combinations, alongside a tailored training strategy for robust representation learning. Extensive experiments on diverse natural and medical image datasets show that DyMo significantly outperforms state-of-the-art incomplete/dynamic MDL methods across various missing-data scenarios. Our code is available at https://github.com//siyi-wind/DyMo.

Advances in generative modeling have recently been adapted to tabular data containing discrete and continuous features. However, generating mixed-type features that combine discrete states with an otherwise continuous distribution in a single feature remains challenging. We advance the state-of-the-art in diffusion models for tabular data with a cascaded approach. We first generate a low-resolution version of a tabular data row, that is, the collection of the purely categorical features and a coarse categorical representation of numerical features. Next, this information is leveraged in the high-resolution flow matching model via a novel guided conditional probability path and data-dependent coupling. The low-resolution representation of numerical features explicitly accounts for discrete outcomes, such as missing or inflated values, and therewith enables a more faithful generation of mixed-type features. We formally prove that this cascade tightens the transport cost bound. The results indicate that our model generates significantly more realistic samples and captures distributional details more accurately, for example, the detection score improves by 51.9\%. Code is available at https://github.com/muellermarkus/tabcascade.

Kolmogorov--Arnold Networks (KANs) have recently emerged as a structured alternative to standard MLPs, yet a principled theory for their training dynamics, generalization, and privacy properties remains limited. In this paper, we analyze gradient descent (GD) for training two-layer KANs and derive general bounds that characterize their training dynamics, generalization, and utility under differential privacy (DP). As a concrete instantiation, we specialize our analysis to logistic loss under an NTK-separable assumption, where we show that polylogarithmic network width suffices for GD to achieve an optimization rate of order $1/T$ and a generalization rate of order $1/n$, with $T$ denoting the number of GD iterations and $n$ the sample size. In the private setting, we characterize the noise required for $(ε,δ)$-DP and obtain a utility bound of order $\sqrt{d}/(nε)$ (with $d$ the input dimension), matching the classical lower bound for general convex Lipschitz problems. Our results imply that polylogarithmic width is not only sufficient but also necessary under differential privacy, revealing a qualitative gap between non-private (sufficiency only) and private (necessity also emerges) training regimes. Experiments further illustrate how these theoretical insights can guide practical choices, including network width selection and early stopping.

Classifier-free guidance (CFG) is a widely used technique for controllable generation in diffusion and flow-based models. Despite its empirical success, CFG relies on a heuristic linear extrapolation that is often sensitive to the guidance scale. In this work, we provide a principled interpretation of CFG through the lens of optimization. We demonstrate that the velocity field in flow matching corresponds to the gradient of a sequence of smoothed distance functions, which guides latent variables toward the scaled target image set. This perspective reveals that the standard CFG formulation is an approximation of this gradient, where the prediction gap, the discrepancy between conditional and unconditional outputs, governs guidance sensitivity. Leveraging this insight, we reformulate the CFG sampling as a homotopy optimization with a manifold constraint. This formulation necessitates a manifold projection step, which we implement via an incremental gradient descent scheme during sampling. To improve computational efficiency and stability, we further enhance this iterative process with Anderson Acceleration without requiring additional model evaluations. Our proposed methods are training-free and consistently refine generation fidelity, prompt alignment, and robustness to the guidance scale. We validate their effectiveness across diverse benchmarks, demonstrating significant improvements on large-scale models such as DiT-XL-2-256, Flux, and Stable Diffusion 3.5.

Prior-Data Fitted Networks (PFNs) enable amortized Bayesian inference in a single forward pass, yet their internal representations remain opaque. It is unknown whether PFNs encode identifiable Bayesian structure or merely memorize input-output mappings. We provide mechanistic evidence that PFNs learn structured spectral representations and that these can be extracted as explicit kernels. First, probing experiments across three architectures, including the publicly released TabPFN, show that spectral information is linearly decodable from the latent attention score and organized along a dominant principal axis. Activation patching and targeted subspace interventions establish that this information is causally used for prediction and concentrated in a low-dimensional subspace, with spectral directions an order of magnitude more effective than random ones. Crucially, these properties hold on TabPFN with both synthetic out-of-distribution inputs and real-world time series (Airline Passengers, Milk Production), indicating they are emergent features of PFN-style amortization over continuous regression tasks rather than artifacts of training prior. Second, we introduce a Filter Bank Decoder that maps frozen PFN latents to explicit spectral densities, reconstructing stationary kernels via Bochner's theorem. The resulting kernels support GP regression competitive with iterative baselines while requiring only a single forward pass, demonstrating that PFN priors are not merely implicit but are explicitly recoverable as portable Bayesian objects.

Catastrophic forgetting during knowledge injection impairs the ability of large language models to acquire new knowledge without overwriting previously mastered knowledge. Recent studies analyze forgetting from a gradient similarity perspective and mitigate forgetting through vector projection. However, these methods primarily characterize gradient similarity at the aggregate direction level, leaving the parameter wise contributions to forgetting underexplored. In this paper, we decompose gradient similarity into parameter wise contributions and identify two types of parameters during forgetting: Conflicting Parameters, whose updates contribute to forgetting and typically account for 50 percent to 75 percent of parameters, and Collaborative Parameters, whose updates mitigate forgetting and account for 25 percent to 50 percent. Based on this analysis, we propose Collaborative Parameter Learning, CPL, a parameter wise training rule that freezes Conflicting Parameters and updates only Collaborative Parameters. Experiments comparing CPL with seven baseline methods show that CPL learns 20.2% to 48.2% more questions with negligible forgetting, while reducing peak VRAM by approximately 3 GB per billion model parameters and computation time by 16.5 percent. Extensive evaluations on parameter consumption, out of set generalization, cross prompt generalization, multimodal tasks, open ended question answering, and multilingual settings demonstrate that CPL effectively mitigates forgetting across diverse scenarios.

Across self-consistency samples from an LLM, vote agreement tracks instance difficulty: on SemEval-2026 Task 4 (Narrative Story Similarity), supermajority cases (>= 7/8 votes) resolve at 85 percent accuracy, split votes at 67 percent, and perfect ties at 61 percent, a monotone gradient that holds across the development set. We exploit this in CascadeMind, which routes eight Gemini 2.5 Flash votes by consensus, escalates split votes to additional sampling rounds, and falls through to a symbolic ensemble of theory-inspired narrative signals only on perfect ties (5 percent of cases). The system reached 72.75 percent on Track A test, placing 10th of 44 teams. Ablations show that the symbolic component contributes negligibly end-to-end and that nearly all gains come from confidence-aware routing. The takeaway is methodological: for narrative similarity, calibrating when to spend more compute on a hard instance matters more than adding auxiliary representations to reason about it.

Semantic associations such as the link between "bird" and "flew" are foundational for language modeling as they enable models to go beyond memorization and instead generalize and generate coherent text. Understanding how these associations are learned and represented in language models is essential for connecting deep learning with linguistic theory and developing a mechanistic foundation for large language models. In this work, we analyze how these associations emerge from natural language data in attention-based language models through the lens of training dynamics. By leveraging a leading-term approximation of the gradients, we develop closed-form expressions for the weights at early stages of training that explain how semantic associations first take shape. Through our analysis, we reveal that each set of weights of the transformer has closed-form expressions as simple compositions of three basis functions (bigram, token-interchangeability, and context mappings), reflecting the statistics of the text corpus and uncovering how each component of the transformer captures semantic associations based on these compositions. Experiments on real-world LLMs demonstrate that our theoretical weight characterizations closely match the learned weights, and qualitative analyses further show how our theorem shines light on interpreting the learned associations in transformers.

Retinal prostheses restore limited visual perception, but low spatial resolution and temporal persistence make reading difficult. In sequential letter presentation, the afterimage of one symbol can interfere with perception of the next, leading to systematic recognition errors. Rather than relying on future hardware improvements, we investigate whether optimizing the visual symbols themselves can mitigate this temporal interference. We present SymbolSight, a computational framework that selects symbol-to-letter mappings to minimize confusion among frequently adjacent letters. Using simulated prosthetic vision (SPV) and a neural proxy observer, we estimate pairwise symbol confusability and optimize assignments using language-specific bigram statistics. Across simulations in Arabic, Bulgarian, and English, the resulting heterogeneous symbol sets reduced predicted confusion by a median factor of 22 relative to native alphabets. These results suggest that standard typography is poorly matched to serial, low-bandwidth prosthetic vision and demonstrate how computational modeling can narrow the design space of visual encodings, identifying high-potential candidates for future psychophysical and clinical evaluation rather than predicting present-day clinical reading performance directly.

In this work we develop a novel insect-inspired model for visual point-goal navigation. This combines abstracted models of two insect brain structures that have been implicated, respectively, in associative learning and path integration. We draw an analogy between the formal benchmark of the Habitat point-goal navigation task and the ability of insects to discover, learn, and refine visually guided paths around obstacles between a discovered food location and their nest. We demonstrate that the simple insect-inspired model exhibits performance comparable to recent state-of-the-art models at many orders of magnitude less computational cost. Testing in a more realistic simulated environment shows the approach is robust to perturbations.

Large Language Models (LLMs) are increasingly used for clinical decision support, where hallucinations and unsafe suggestions may pose direct risks to patient safety. These risks are hard to assess: subtle clinical errors are often missed by generic metrics and LLM judges using general criteria, while expert-authored fine-grained rubrics are expensive and difficult to scale. In this paper, we propose a retrieval-augmented multi-agent framework designed to automate the generation of instance-specific evaluation rubrics. Our approach grounds evaluation in authoritative medical evidence by decomposing retrieved content into atomic facts and synthesizing them with user interaction constraints to form verifiable, fine-grained evaluation criteria. Evaluated on HealthBench and LLMEval-Med datasets, our framework achieves Clinical Intent Alignment (CIA) scores of 50.20% and 31.90%, significantly outperforming the GPT-4o baseline and demonstrating robust cross-lingual generalization. In discriminative tests on HealthBench, our rubrics yield a 7.8% higher win rate than GPT-4o baseline with nearly double score $Δ$, while ablation studies confirm its structural necessity. Beyond evaluation, our rubrics effectively guide response refinement, improving quality by 9.2%. This provides a scalable, cross-lingual foundation for both evaluating and improving medical LLMs. The code is available at https://github.com/AmbeChen/Automated-Rubric-Generation.

Backdoor attacks can cause reinforcement learning (RL) policies to behave normally under clean inputs while executing malicious behaviors when triggers are present. Existing RL backdoor attacks are primarily studied in simulation and often assume that attackers can reliably manipulate the observations driving policy decisions. This assumption becomes fragile in real-world deployment, where RL policies commonly rely on multimodal observations. Attackers can manipulate visual inputs through physical triggers, but auxiliary states such as LiDAR and odometry signals remain uncontrollable and vary across trajectories. We study this overlooked challenge and propose a diffusion-guided backdoor attack framework (DGBA) for real-world RL. DGBA uses small printable visual patches as triggers and learns a stochastic trigger distribution via conditional diffusion to maintain consistent attack activation under varying uncontrollable states. We further introduce an advantage-based poisoning strategy that injects triggers only at decision-critical training states. Experiments on a physical TurtleBot3 platform show that DGBA consistently outperforms prior RL backdoor attacks while preserving normal task performance. Demo videos and code are available in the supplementary material.

Prefill attacks are an effective and low-cost jailbreaking method, as they directly insert an acceptance sequence (e.g., "Sure, here is how to...") at the start of an LLM's output and lead the model to continue the response. We make two contributions to this prior work. First, we show that an unsophisticated adversary can improve the well-known prefill attacks by ensembling a small number of prefill variants. Running three easy-to-generate prefills yields a combined attack success rate (ASR) of 22%, 90%, and 99% on Gemma-7B, Llama-3.1-8B, and Qwen3-8B respectively, an up to 38% improvement over the standard "Sure, here's..." prefill and up to 82% over our reproduction of GCG (Zou et al., 2023). Second, we introduce "sockpuppetting", a hybrid attack that optimizes an adversarial suffix placed inside the "assistant" message block of the chat template, rather than within the user prompt. The rolling variant of this attack, RollingSockpuppetGCG, increases prompt-agnostic ASR by up to 64% over our universal GCG baseline on Llama-3.1-8B. Both findings highlight the need for defences against output-prefix injection in open-weight models. Code: https://gitlab.com/asendotsinski/sockpuppetting

Variational quantum circuits are increasingly studied as continuous-function approximators, but quantum regression remains difficult to train when global losses, finite-shot stochasticity, and circuit-depth growth combine to produce weak or ill-conditioned gradient signals. We study this trainability problem in a controlled hybrid quantum--classical regression design. The central ingredient is a capacity-controlled classical embedding that acts as a learnable geometric preconditioner: it reshapes the input distribution seen by a data-reuploading variational circuit while preserving a low-dimensional quantum bottleneck. We pair this representation design with a curriculum protocol that grows circuit depth progressively and switches from SPSA-based stochastic exploration to Adam-based analytic-gradient fine-tuning. We formalize the mechanism through a local quantum-tangent contraction statement: in the linearized quantum-parameter dynamics, the embedding changes the empirical Gram matrix that controls residual contraction and one-step loss decrease. Across finite-size statevector audits on PDE-informed regression benchmarks and small-data tabular tasks, the Hybrid QNN lowers error relative to Pure QNN baselines under matched quantum-model budgets. Strong classical references remain competitive, and in several cases are better in absolute error; the evidence therefore supports a trainability claim for the hybrid QNN design rather than a claim of classical or hardware quantum advantage.

While GUI agents have shown strong performance under explicit and completion instructions, real-world deployment requires aligning with users' more complex implicit intents. In this work, we highlight Hierarchical Implicit Intent Alignment for Personalized GUI Agent (PersonalAlign), a new agent task that requires agents to leverage long-term user records as persistent context to resolve omitted preferences in vague instructions and anticipate latent routines by user state for proactive assistance. To facilitate this study, we introduce AndroidIntent, a benchmark designed to evaluate agents' ability in resolving vague instructions and providing proactive suggestions through reasoning over long-term user records. We annotated 775 user-specific preferences and 215 routines from 20k long-term records across different users for evaluation. Furthermore, we introduce Hierarchical Intent Memory Agent (HIM-Agent), which maintains a continuously updating personal memory and hierarchically organizes user preferences and routines for personalization. Finally, we evaluate a range of GUI agents on AndroidIntent, including GPT-5, Qwen3-VL, and UI-TARS, further results show that HIM-Agent significantly improves both execution and proactive performance by 15.7% and 7.3%.