Decomposing complex data into factorized representations can reveal reusable components and enable synthesizing new samples via component recombination. We investigate this in the context of diffusion-based models that learn factorized latent spaces without factor-level supervision. In images, factors can capture background, illumination, and object attributes; in robotic videos, they can capture reusable motion components. To improve both latent factor discovery and quality of compositional generation, we introduce an adversarial training signal via a discriminator trained to distinguish between single-source samples and those generated by recombining factors across sources. By optimizing the generator to fool this discriminator, we encourage physical and semantic consistency in the resulting recombinations. Our method outperforms implementations of prior baselines on CelebA-HQ, Virtual KITTI, CLEVR, and Falcor3D, achieving lower FID scores and better disentanglement as measured by MIG and MCC. Furthermore, we demonstrate a novel application to robotic video trajectories: by recombining learned action components, we generate diverse sequences that significantly increase state-space coverage for exploration on the LIBERO benchmark.

Patients are increasingly using large language models (LLMs) to seek answers to their healthcare-related questions. However, benchmarking efforts in LLMs for question answering often focus on medical exam questions, which differ significantly in style and content from the questions patients actually raise in real life. To bridge this gap, we sourced data from Google's People Also Ask feature by querying the top 200 prescribed medications in the United States, curating a dataset of medical questions people commonly ask. A considerable portion of the collected questions contains incorrect assumptions and dangerous intentions. We demonstrate that the emergence of these corrupted questions is not uniformly random and depends heavily on the degree of incorrectness in the history of questions that led to their appearance. Current LLMs that perform strongly on other benchmarks struggle to identify incorrect assumptions in everyday questions.

Low-resource machine translation (MT) has gained increasing attention as parallel data from low-resource language communities is collected, but many approaches for improving low-resource MT remain underexplored. We investigate a self-supervised reinforcement learning fine-tuning for translation in low-resource settings using round-trip bootstrapping with the No Language Left Behind (NLLB) family of models. Our approach translates English into a target low-resource language and then back into English, using a combination of chrF++ and BLEU as the reward function on the reconstructed English sentences. Using the NLLB-MD dataset, we evaluate both the 600M and 1.3B parameter NLLB models and observe consistent improvements for the following languages: Central Aymara, Friulian, Wolof, Dyula, Bhojpuri and Russian. Qualitative inspection of translation outputs indicates increased fluency and semantic fidelity. We argue that our method can further benefit from scale, enabling models to increasingly leverage their pretrained knowledge and continue self-improving. Code available at: https://github.com/Copticoder/MT-via-Round-Trip-RL

Early identification of stroke symptoms is essential for enabling timely intervention and improving patient outcomes, particularly in prehospital settings. This study presents a fast, non-invasive multimodal deep learning framework for automatic binary stroke screening based on data collected during the F.A.S.T. assessment. The proposed approach integrates complementary information from facial expressions, speech signals, and upper-body movements to enhance diagnostic robustness. Facial dynamics are represented using landmark based features and modeled with a Transformer architecture to capture temporal dependencies. Speech signals are converted into mel spectrograms and processed using an Audio Spectrogram Transformer, while upper-body pose sequences are analyzed with an MLP-Mixer network to model spatiotemporal motion patterns. The extracted modality specific representations are combined through an attention-based fusion mechanism to effectively learn cross modal interactions. Experiments conducted on a self-collected dataset of 222 videos from 37 subjects demonstrate that the proposed multimodal model consistently outperforms unimodal baselines, achieving 95.83% accuracy and a 96.00% F1-score. The model attains a strong balance between sensitivity and specificity and successfully detects all stroke cases in the test set. These results highlight the potential of multimodal learning and transfer learning for early stroke screening, while emphasizing the need for larger, clinically representative datasets to support reliable real-world deployment.

Speculative decoding (SD) has become a popular technique to accelerate Large Language Model (LLM) inference, yet its real-world effectiveness remains unclear as prior evaluations rely on research prototypes and unrealistically small batch sizes. We present, to our knowledge, the first systematic study of SD on a production-grade and widely deployed inference engine (vLLM), covering multiple SD variants ($n$-gram, EAGLE/EAGLE-3, Draft-Model, Multi-Token Prediction) across diverse workloads, model scales, and batch sizes. We analyze key factors governing SD performance, and quantify a theoretical upper bound on SD speedup. Our results show that verification by the target model dominates the execution, while acceptance length varies markedly across output token positions, requests, and datasets. Comparing measured performance with theoretical bounds reveals substantial gaps between observed and theoretical upper bounds, and we leverage this observation to highlight new research opportunities that our study opens up in improving SD.

Academic paper search is a fundamental task in scientific research, yet most existing approaches rely on rigid, predefined workflows that struggle with complex, conditional queries. To address this limitation, we propose PaperScout, an autonomous agent that reformulates paper search as a sequential decision-making process. Unlike static workflows, PaperScout dynamically decides whether, when, and how to invoke search and expand tools based on accumulated retrieval context. However, training such agents presents a fundamental challenge: standard reinforcement learning methods, typically designed for single-turn tasks, suffer from a granularity mismatch when applied to multi-turn agentic tasks-where token-level optimization diverges from the granularity of sequence-level interactions-leading to noisy credit assignment and unstable training dynamics. We introduce Proximal Sequence Policy Optimization (PSPO), a process-aware, sequence-level policy optimization method that aligns optimization with agent--environment interaction. Comprehensive experiments on both synthetic and real-world benchmarks demonstrate that PaperScout significantly outperforms strong workflow-driven and RL baselines in both recall and relevance, validating the effectiveness of our adaptive agentic framework and optimization strategy.

We propose a framework that amortizes the cost of inference-time reasoning by converting transient critiques into retrievable guidelines, through a file-based memory system and agent-controlled tool calls. We evaluate this method on the Rubric Feedback Bench, a novel dataset for rubric-based learning. Experiments demonstrate that our augmented LLMs rapidly match the performance of test-time refinement pipelines while drastically reducing inference cost.

Learning from Preferences in Reinforcement Learning (PbRL) has gained attention recently, as it serves as a natural fit for complicated tasks where the reward function is not easily available. However, preferences often come with uncertainty and noise if they are not from perfect teachers. Much prior literature aimed to detect noise, but with limited types of noise and most being uniformly distributed with no connection to observations. In this work, we formalize the notion of targeted feature-dependent noise and propose several variants like trajectory feature noise, trajectory similarity noise, margin dependent noise, and Language Model noise. We evaluate feature-dependent noise, where noise is correlated with certain features in complex continuous control tasks from DMControl and Meta-world. Our experiments show that in some feature-dependent noise settings, the state-of-the-art noise-robust PbRL method's learning performance is significantly deteriorated, while PbRL method with no explicit denoising can surprisingly outperform noise-robust PbRL in the majority of settings. We also find language models' noise exhibits similar characteristics to feature-dependent noise, thereby simulating realistic humans and call for further study in learning with feature-dependent noise robustly.

Multi-model routing has evolved from an engineering technique into essential infrastructure, yet existing work lacks a systematic, reproducible benchmark for evaluating vision-language models (VLMs). We present VL-RouterBench to assess the overall capability of VLM routing systems systematically. The benchmark is grounded in raw inference and scoring logs from VLMs and constructs quality and cost matrices over sample-model pairs. In scale, VL-RouterBench covers 14 datasets across 3 task groups, totaling 30,540 samples, and includes 15 open-source models and 2 API models, yielding 519,180 sample-model pairs and a total input-output token volume of 34,494,977. The evaluation protocol jointly measures average accuracy, average cost, and throughput, and builds a ranking score from the harmonic mean of normalized cost and accuracy to enable comparison across router configurations and cost budgets. On this benchmark, we evaluate 10 routing methods and baselines and observe a significant routability gain, while the best current routers still show a clear gap to the ideal Oracle, indicating considerable room for improvement in router architecture through finer visual cues and modeling of textual structure. We will open-source the complete data construction and evaluation toolchain to promote comparability, reproducibility, and practical deployment in multimodal routing research.

The reasoning performance of large language models (LLMs) can be substantially improved by training them with reinforcement learning (RL). The RL objective for LLM training involves a regularization term, which is the reverse Kullback-Leibler (KL) divergence between the trained policy and the reference policy. Since computing the KL divergence exactly is intractable, various estimators are used in practice to estimate it from on-policy samples. Despite its wide adoption, including in several open-source libraries, there is no systematic study analyzing the numerous ways of incorporating KL estimators in the objective and their effect on the downstream performance of RL-trained models. Recent works show that prevailing practices for incorporating KL regularization do not provide correct gradients for stated objectives, creating a discrepancy between the objective and its implementation. In this paper, we further analyze these practices and study the gradients of several estimators configurations, revealing how design choices shape gradient bias. We substantiate these findings with empirical observations by RL fine-tuning \texttt{Qwen2.5-7B}, \texttt{Llama-3.1-8B-Instruct} and \texttt{Qwen3-4B-Instruct-2507} with different configurations and evaluating their performance on both in- and out-of-distribution tasks. Through our analysis, we observe that, in on-policy settings: (1) estimator configurations with biased gradients can result in training instabilities; and (2) using estimator configurations resulting in unbiased gradients leads to better performance on in-domain as well as out-of-domain tasks. We also investigate the performance resulting from different KL configurations in off-policy settings and observe that KL regularization can help stabilize off-policy RL training resulting from asynchronous setups.

This paper presents a novel, cost-effective, and scalable approach to track numerous assets distributed in forested environments using commodity Radio Frequency Identification (RFID) targeting wildfire response applications. Commodity RFID systems suffer from poor tag localization when dispersed in forested environments due to signal attenuation, multi-path effects and environmental variability. Current methods to address this issue via fingerprinting rely on dispersing tags at known locations {\em a priori}. In this paper, we address the case when it is not possible to tag known locations and show that it is possible to localize tags to accuracies comparable to global positioning systems (GPS) without such a constraint. For this, we propose Gaussian Process to model various environments solely based on RF signal response signatures and without the aid of additional sensors such as global positioning GPS or cameras, and match an unknown RF to the closest match in a model dictionary. We utilize a new weighted log-likelihood method to associate an unknown environment with the closest environment in a dictionary of previously modeled environments, which is a crucial step in being able to use our approach. Our results show that it is possible to achieve localization accuracies of the order of GPS, but with passive commodity RFID, which will allow the tracking of dozens of wildfire assets within the vicinity of mobile readers at-a-time simultaneously, does not require known positions to be tagged {\em a priori}, and can achieve localization at a fraction of the cost compared to GPS.

Human beings solve complex problems through critical thinking, where reasoning and evaluation are intertwined to converge toward correct solutions. However, most existing large language models (LLMs) treat the reasoning and verification as separate processes: they either generate reasoning without explicit self-checking or rely on external verifiers to detect errors post hoc. The former lacks immediate feedback, while the latter increases system complexity and hinders synchronized learning. Motivated by human critical thinking, we propose Stepwise Think-Critique (STC), a unified and end-to-end trainable framework that interleaves reasoning and self-critique at every intermediate step within a single model. STC is trained with a hybrid reinforcement learning objective that integrates reasoning rewards and critique-consistency rewards, thereby jointly optimizing solution correctness and reliability of self-evaluation. Experiments on mathematical reasoning benchmarks show that STC demonstrates strong critical-thinking capabilities and produces more interpretable reasoning traces, representing a step toward LLMs with built-in critical thinking.

Embodied agents are expected to assist humans by actively exploring unknown environments and reasoning about spatial contexts. When deployed in real life, agents often face sequential tasks where each new task follows the completion of the previous one and may include infeasible objectives, such as searching for non-existent objects. However, most existing research focuses on isolated goals, overlooking the core challenge of sequential tasks: the ability to reuse spatial knowledge accumulated from previous explorations to guide subsequent reasoning and exploration. In this work, we investigate this underexplored yet practically significant embodied AI challenge. Specifically, we propose 3DSPMR, a 3D SPatial Memory Reasoning framework that utilizes Field-of-View (FoV) coverage as an explicit geometric prior. By integrating FoV-based constraints, 3DSPMR significantly enhances an agent's memory, reasoning, and exploration capabilities across sequential tasks. To facilitate research in this area, we further introduce SEER-Bench, a novel Sequential Embodied Exploration and Reasoning Benchmark that spans two foundational tasks: Embodied Question Answering (EQA) and Embodied Multi-modal Navigation (EMN). SEER-Bench uniquely incorporates both feasible and infeasible tasks to provide a rigorous and comprehensive evaluation of agent performance. Extensive experiments verify that 3DSPMR achieves substantial performance gains on both sequential EQA and EMN tasks.

3D vision foundation models have shown strong generalization in reconstructing key 3D attributes from uncalibrated images through a single feed-forward pass. However, when deployed in online settings such as driving scenarios, predictions are made over temporal windows, making it non-trivial to maintain consistency across time. Recent strategies align consecutive predictions by solving global transformation, yet our analysis reveals their fundamental limitations in assumption validity, local alignment scope, and robustness under noisy geometry. In this work, we propose a higher-DOF and long-term alignment framework based on Thin Plate Spline, leveraging globally propagated control points to correct spatially varying inconsistencies. In addition, we adopt a point-agnostic submap registration design that is inherently robust to noisy geometry predictions. The proposed framework is fully plug-and-play, compatible with diverse 3D foundation models and camera configurations (e.g., monocular or surround-view). Extensive experiments demonstrate that our method consistently yields more coherent geometry and lower trajectory errors across multiple datasets, backbone models, and camera setups, highlighting its robustness and generality. Code is available at https://github.com/Xian-Bei/TALO.

Large Language Models (LLMs) are increasingly deployed for clinical reasoning tasks, which inherently require eliciting calibrated probabilistic beliefs based on available evidence. However, real-world clinical data are frequently incomplete, with missingness patterns often informative of patient prognosis; for example, ordering a rare laboratory test reflects a clinician's latent suspicion. In this work, we investigate whether LLMs can be steered to leverage this informative missingness for prognostic inference. To evaluate how well LLMs align their verbalized probabilistic beliefs with an underlying target distribution, we analyze three common prompt-based interventions: explicit serialization, instruction steering, and in-context learning. We introduce a bias-variance decomposition of the log-loss to clarify the mechanisms driving gains in predictive performance. Using a real-world intensive care testbed, we find that while explicit structural steering and in-context learning can improve probabilistic alignment, the models do not natively leverage informative missingness without careful interventions.

Multimodal large language models (MLLMs) are increasingly deployed in real-world, agentic settings where outputs must not only be correct, but also conform to predefined data schemas. Despite recent progress in structured generation in textual domain, there is still no benchmark that systematically evaluates schema-grounded information extraction and reasoning over visual inputs. In this work, we conduct a comprehensive study of visual structural output capabilities for MLLMs with our carefully designed SO-Bench benchmark. Covering four visual domains, including UI screens, natural images, documents, and charts, SO-Bench is built from over 6.5K diverse JSON schemas and 1.8K curated image-schema pairs with human-verified quality. Benchmarking experiments on open-sourced and frontier proprietary models reveal persistent gaps in predicting accurate, schema compliant outputs, highlighting the need for better multimodal structured reasoning. Beyond benchmarking, we further conduct training experiments to largely improve the model's structured output capability. We make the benchmark and evaluation publicly available at https://github.com/apple/ml-sobench

Recent years have witnessed remarkable progress in world models, which primarily aim to capture the spatio-temporal correlations between an agent's actions and the evolving environment. However, existing approaches often suffer from tight runtime coupling or depend on offline reward signals, resulting in substantial inference overhead or hindering end-to-end optimization. To overcome these limitations, we introduce WPT, a World-to-Policy Transfer training paradigm that enables online distillation under the guidance of an end-to-end world model. Specifically, we develop a trainable reward model that infuses world knowledge into a teacher policy by aligning candidate trajectories with the future dynamics predicted by the world model. Subsequently, we propose policy distillation and world reward distillation to transfer the teacher's reasoning ability into a lightweight student policy, enhancing planning performance while preserving real-time deployability. Extensive experiments on both open-loop and closed-loop benchmarks show that our WPT achieves state-of-the-art performance with a simple policy architecture: it attains a 0.11 collision rate (open-loop) and achieves a 79.23 driving score (closed-loop) surpassing both world-model-based and imitation-learning methods in accuracy and safety. Moreover, the student sustains up to 4.9x faster inference, while retaining most of the gains.

Robotic laser systems offer the potential for sub-millimeter, non-contact, high-precision tissue resection, yet existing platforms lack volumetric planning and intraoperative feedback. We present RATS (Robot-Assisted Tissue Surgery), an intelligent opto-mechanical, optical coherence tomography (OCT)-guided robotic platform designed for autonomous volumetric soft tissue resection in surgical applications. RATS integrates macro-scale RGB-D imaging, micro-scale OCT, and a fiber-coupled surgical laser, calibrated through a novel multistage alignment pipeline that achieves OCT-to-laser calibration accuracy of 0.161+-0.031mm on tissue phantoms and ex vivo porcine tissue. A super-Gaussian laser-tissue interaction (LTI) model characterizes ablation crater morphology with an average RMSE of 0.231+-0.121mm, outperforming Gaussian baselines. A sampling-based model predictive control (MPC) framework operates directly on OCT voxel data to generate constraint-aware resection trajectories with closed-loop feedback, achieving 0.842mm RMSE and improving intersection-over-union agreement by 64.8% compared to feedforward execution. With OCT, RATS detects subsurface structures and modifies the planner's objective to preserve them, demonstrating clinical feasibility.

Recent advances in self-supervised visual representation learning have demonstrated the effectiveness of predictive latent-space objectives for learning transferable features. In particular, Image-based Joint-Embedding Predictive Architecture (I-JEPA) learns representations by predicting latent embeddings of masked target regions from visible context. However, it predicts target regions in parallel and all at once, lacking ability to order predictions meaningfully. Inspired by human visual perception, which attends selectively and progressively from primary to secondary cues, we propose DSeq-JEPA, a Discriminative Sequential Joint-Embedding Predictive Architecture that bridges latent predictive and autoregressive self-supervised learning. Specifically, DSeq-JEPA integrates a discriminatively ordered sequential process with JEPA-style learning objective. This is achieved by (i) identifying primary discriminative regions using an attention-derived saliency map that serves as a proxy for visual importance, and (ii) predicting subsequent regions in discriminative order, inducing a curriculum-like semantic progression from primary to secondary cues in pre-training. Extensive experiments across tasks -- image classification (ImageNet), fine-grained visual categorization (iNaturalist21, CUB, Stanford Cars), detection/segmentation (MS-COCO, ADE20K), and low-level reasoning (CLEVR) -- show that DSeq-JEPA consistently learns more discriminative and generalizable representations compared to I-JEPA variants. Project page: https://github.com/SkyShunsuke/DSeq-JEPA.

Group Relative Policy Optimization (GRPO) has shown promise in aligning image and video generative models with human preferences. However, applying it to modern flow matching models is challenging because of its deterministic sampling paradigm. Current methods address this issue by converting Ordinary Differential Equations (ODEs) to Stochastic Differential Equations (SDEs), which introduce stochasticity. However, this SDE-based GRPO suffers from issues of inefficient credit assignment and incompatibility with high-order solvers for fewer-step sampling. In this paper, we first reinterpret existing SDE-based GRPO methods from a distance optimization perspective, revealing their underlying mechanism as a form of contrastive learning. Based on this insight, we propose Neighbor GRPO, a novel alignment algorithm that completely bypasses the need for SDEs. Neighbor GRPO generates a diverse set of candidate trajectories by perturbing the initial noise conditions of the ODE and optimizes the model using a softmax distance-based surrogate leaping policy. We establish a theoretical connection between this distance-based objective and policy gradient optimization, rigorously integrating our approach into the GRPO framework. Our method fully preserves the advantages of deterministic ODE sampling, including efficiency and compatibility with high-order solvers. We further introduce symmetric anchor sampling for computational efficiency and group-wise quasi-norm reweighting to address reward flattening. Extensive experiments demonstrate that Neighbor GRPO significantly outperforms SDE-based counterparts in terms of training cost, convergence speed, and generation quality.

Recent studies show that text to image (T2I) diffusion models are vulnerable to backdoor attacks, where a trigger in the input prompt can steer generation toward harmful or unintended content. Beyond the trigger token itself, backdoor effects can spread to neighboring tokens in the text embedding space. To address this, we introduce PEPPER (PErcePtion Guided PERturbation), a backdoor defense that rewrites the caption into a semantically distant yet visually similar caption while adding unobstructive elements. With this rewriting strategy, PEPPER disrupt the trigger embedded in the input prompt, dilute the influence of trigger tokens and thereby achieve enhanced robustness. Experiments show that PEPPER is particularly effective against text encoder based attacks, substantially reducing attack success while preserving generation quality. Beyond this, PEPPER can be paired with any existing defenses yielding consistently stronger and generalizable robustness than any standalone method. Our code will be released on Github.

In-context learning (ICL) -- the capacity of a model to infer and apply abstract patterns from examples provided within its input -- has been extensively studied in large language models trained for next-token prediction on human text. In fact, prior work often attributes this emergent behavior to distinctive statistical properties in human language. This raises a fundamental question: can ICL arise organically in other sequence domains purely through large-scale predictive training? To explore this, we turn to genomic sequences, an alternative symbolic domain rich in statistical structure. Specifically, we study the Evo2 genomic model, trained predominantly on next-nucleotide (A/T/C/G) prediction, at a scale comparable to mid-sized LLMs. We develop a controlled experimental framework comprising symbolic reasoning tasks instantiated in both linguistic and genomic forms, enabling direct comparison of ICL across genomic and linguistic models. Our results show that genomic models, like their linguistic counterparts, exhibit log-linear gains in pattern induction as the number of in-context demonstrations increases. To the best of our knowledge, this is the first evidence of organically emergent ICL in genomic sequences, supporting the hypothesis that ICL arises as a consequence of large-scale predictive modeling over rich data. These findings extend emergent meta-learning beyond language, pointing toward a unified, modality-agnostic view of in-context learning.

While recent Large Vision-Language Models (LVLMs) exhibit strong multimodal reasoning abilities, they often produce ungrounded or hallucinated responses because they rely too heavily on linguistic priors instead of visual evidence. This limitation highlights the absence of a quantitative measure of how much these models actually use visual information during reasoning. We propose Draft and Refine (DnR), an agent framework driven by a question-conditioned utilization metric. The metric quantifies the model's reliance on visual evidence by first constructing a query-conditioned relevance map to localize question-specific cues and then measuring dependence through relevance-guided probabilistic masking. Guided by this metric, the DnR agent refines its initial draft using targeted feedback from external visual experts. Each expert's output (such as boxes or masks) is rendered as visual cues on the image, and the model is re-queried to select the response that yields the largest improvement in utilization. This process strengthens visual grounding without retraining or architectural changes. Experiments across VQA and captioning benchmarks show consistent accuracy gains and reduced hallucination, demonstrating that measuring visual utilization provides a principled path toward more interpretable and evidence-driven multimodal agent systems. Code is available at https://github.com/EavnJeong/Draft-and-Refine-with-Visual-Experts.

Sustainability and efficiency have become essential considerations in the development and deployment of Artificial Intelligence systems, but existing regulatory practices for Green AI still lack standardized, model-agnostic evaluation protocols. Recently, sustainability auditing pipelines for ML and usual practices by researchers show three main pitfalls: 1) they disproportionally emphasize epoch/batch learning settings, 2) they do not formally model the long-term sustainability cost of adapting and re-training models, and 3) they effectively measure the sustainability of sterile experiments, instead of estimating the environmental impact of real-world, long-term AI lifecycles. In this work, we propose a novel evaluation protocol for assessing the long-term sustainability of ML models, based on concepts inspired by Online ML, which measures sustainability and performance through incremental/continual model retraining parallel to real-world data acquisition. Through experimentation on diverse ML tasks using a range of model types, we demonstrate that traditional static train-test evaluations do not reliably capture sustainability under evolving datasets, as they overestimate, underestimate and/or erratically estimate the actual cost of maintaining and updating ML models. Our proposed sustainability evaluation pipeline also draws initial evidence that, in real-world, long-term ML life-cycles, higher environmental costs occasionally yield little to no performance benefits.

Post-Training Quantization (PTQ) has become the de-facto standard for efficient LLM deployment, yet its optimization objective remains fundamentally incomplete. Standard PTQ methods minimize reconstruction error (e.g., MSE or KL divergence) without accounting for behavioral alignment--a critical property instilled through safety fine-tuning. We demonstrate that this objective mismatch introduces a systematic vulnerability: models can maintain low perplexity yet exhibit significant degradation in safety alignment, revealing that perplexity alone is an insufficient and often misleading proxy for deployment readiness. To address this, we propose Contrastive Alignment Quantization (CAQ), which extends the PTQ objective design space by integrating a Contrastive Alignment Loss (CAL). CAL introduces a principled push-pull mechanism that jointly optimizes distributional fidelity and behavioral alignment: it steers the quantized model toward its safe, instruction-tuned counterpart while diverging from the unaligned, pre-trained reference. CAQ requires no specialized safety datasets, relying solely on standard calibration data, and introduces negligible computational overhead over existing transformation-based PTQ pipelines. We show that CAQ enables robust 4-bit (W4A4) quantization across diverse model families--including LLaMA, Qwen, and Mistral--achieving superior safety alignment where state-of-the-art PTQ methods fail, without sacrificing general capabilities. Anonymized code is available in the supplementary material.

Safety evaluation of large language models (LLMs) increasingly relies on LLM-as-a-judge pipelines, but strong judges can still be expensive to use at scale. We study whether structured multi-agent debate can improve judge reliability while keeping backbone size and cost modest. To do so, we introduce HAJailBench, a human-annotated jailbreak benchmark with 11,100 labeled interactions spanning diverse attack methods and target models, and we pair it with a Multi-Agent Judge framework in which critic, defender, and judge agents debate under a shared safety rubric. On HAJailBench, the framework improves over matched small-model prompt baselines and prior multi-agent judges, while remaining more economical than GPT-4o under the evaluated pricing snapshot. Ablation results further show that a small number of debate rounds is sufficient to capture most of the gain. Together, these results support structured, value-aligned debate as a practical design for scalable LLM safety evaluation.

Precision agriculture demands continuous and accurate monitoring of soil moisture (M) and key macronutrients, including nitrogen (N), phosphorus (P), and potassium (K), to optimize yields and conserve resources. Wireless soil sensing has been explored to measure these four components; however, current solutions require recalibration (i.e., retraining the data processing model) to handle variations in soil texture, characterized by aluminosilicates (Al) and organic carbon (C), limiting their practicality. To address this, we introduce SoilX, a calibration-free soil sensing system that jointly measures six key components: {M, N, P, K, C, Al}. By explicitly modeling C and Al, SoilX eliminates texture- and carbon-dependent recalibration. SoilX incorporates Contrastive Cross-Component Learning (3CL), with two customized terms: the Orthogonality Regularizer and the Separation Loss, to effectively disentangle cross-component interference. Additionally, we design a novel tetrahedral antenna array with an antenna-switching mechanism, which can robustly measure soil dielectric permittivity independent of device placement. Extensive experiments demonstrate that SoilX reduces estimation errors by 23.8% to 31.5% over baselines and generalizes well to unseen fields.

Safety-aligned large language models (LLMs) are becoming increasingly widespread, especially in sensitive applications where fairness is essential and biased outputs can cause significant harm. However, evaluating the fairness of models is a complex challenge, and approaches that do so typically utilize standard question-answer (QA) styled schemes. Such methods often overlook deeper issues by interpreting the model's refusal responses as positive fairness measurements, which creates a false sense of fairness. In this work, we introduce the concept of silenced biases, which are unfair preferences encoded within models' latent space and are effectively concealed by safety-alignment. Previous approaches that considered similar indirect biases often relied on prompt manipulation or handcrafted implicit queries, which present limited scalability and risk contaminating the evaluation process with additional biases. We propose the Silenced Bias Benchmark (SBB), which aims to uncover these biases by employing activation steering to reduce model refusals during QA. SBB supports easy expansion to new demographic groups and subjects, presenting a fairness evaluation framework that encourages the future development of fair models and tools beyond the masking effects of alignment training. We demonstrate our approach over multiple LLMs, where our findings expose an alarming distinction between models' direct responses and their underlying fairness issues.

Data augmentation is widely used in vision to introduce variation and mitigate overfitting, by enabling models to learn invariant properties. However, augmentation only indirectly captures these properties and does not explicitly constrain the learned function to satisfy them beyond the empirical training set. We propose generative hints, a training methodology that directly enforces known functional invariances over the input distribution. Our approach leverages a generative model trained on the training data to approximate the input distribution and to produce unlabeled synthetic images, which we refer to as virtual examples. On these virtual examples, we impose hint objectives that explicitly constrain the model's predictions to satisfy known invariance properties, such as spatial invariance. Although the original training dataset is fully labeled, generative hints train the model in a semi-supervised manner by combining the standard classification objective on real data with an auxiliary hint objectives applied to unlabeled virtual examples. Across multiple datasets, architectures, invariance types, and loss functions, generative hints consistently outperform standard data augmentation, achieving accuracy improvements of up to 2.10% on fine-grained visual classification benchmarks and an average gain of 1.29% on the CheXpert medical imaging dataset.

Recent hybrid video generation models combine autoregressive temporal dynamics with diffusion-based spatial denoising, but their sequential, iterative nature leads to error accumulation and long inference times. In this work, we propose a distillation-based framework for efficient causal video generation that enables high-quality synthesis with extremely limited denoising steps. Our approach builds upon the Distribution Matching Distillation (DMD) framework and proposes a novel Adversarial Self-Distillation (ASD) strategy, which aligns the outputs of the student model's n-step denoising process with its (n+1)-step version at the distribution level. This design provides smoother supervision by bridging small intra-student gaps and more informative guidance by combining teacher knowledge with locally consistent student behavior, substantially improving training stability and generation quality in extremely few-step scenarios (e.g., 1-2 steps). In addition, we present a First-Frame Enhancement (FFE) strategy, which allocates more denoising steps to the initial frames to mitigate error propagation while applying larger skipping steps to later frames. Extensive experiments on VBench demonstrate that our method surpasses state-of-the-art approaches in both one-step and two-step video generation. Notably, our framework produces a single distilled model that flexibly supports multiple inference-step settings, eliminating the need for repeated re-distillation and enabling efficient, high-quality video synthesis.