While many vision-language models (VLMs) are developed to answer well-defined, straightforward questions with highly specified targets, as in most benchmarks, they often struggle in practice with complex open-ended tasks, which usually require multiple rounds of exploration and reasoning in the visual space. Such visual thinking paths not only provide step-by-step exploration and verification as an AI detective but also produce better interpretations of the final answers. However, these paths are challenging to evaluate due to the large exploration space of intermediate steps. To bridge the gap, we develop an evaluation suite, ``Visual Reasoning with multi-step EXploration (V-REX)'', which is composed of a benchmark of challenging visual reasoning tasks requiring native multi-step exploration and an evaluation protocol. V-REX covers rich application scenarios across diverse domains. V-REX casts the multi-step exploratory reasoning into a Chain-of-Questions (CoQ) and disentangles VLMs' capability to (1) Planning: breaking down an open-ended task by selecting a chain of exploratory questions; and (2) Following: answering curated CoQ sequentially to collect information for deriving the final answer. By curating finite options of questions and answers per step, V-REX achieves a reliable quantitative and fine-grained analysis of the intermediate steps. By assessing SOTA proprietary and open-sourced VLMs, we reveal consistent scaling trends, significant differences between planning and following abilities, and substantial room for improvement in multi-step exploratory reasoning.

High-quality geometric diagram generation presents both a challenge and an opportunity: it demands strict spatial accuracy while offering well-defined constraints to guide generation. Inspired by recent advances in geometry problem solving that employ formal languages and symbolic solvers for enhanced correctness and interpretability, we propose GeoLoom, a novel framework for text-to-diagram generation in geometric domains. GeoLoom comprises two core components: an autoformalization module that translates natural language into a specifically designed generation-oriented formal language GeoLingua, and a coordinate solver that maps formal constraints to precise coordinates using the efficient Monte Carlo optimization. To support this framework, we introduce GeoNF, a dataset aligning natural language geometric descriptions with formal GeoLingua descriptions. We further propose a constraint-based evaluation metric that quantifies structural deviation, offering mathematically grounded supervision for iterative refinement. Empirical results demonstrate that GeoLoom significantly outperforms state-of-the-art baselines in structural fidelity, providing a principled foundation for interpretable and scalable diagram generation.

Large language models (LLMs) tackle complex tasks by generating long chains of thought or "reasoning traces" that act as latent variables in the generation of an output given a query. A model's ability to generate such traces can be optimized with reinforcement learning (RL) to improve their utility in predicting an answer. This optimization comes at a high computational cost, especially for narrative-related tasks that involve retrieving and processing many tokens. To this end, we propose LiteReason, a latent reasoning method that can be interleaved with standard token sampling and easily combined with RL techniques. LiteReason employs a lightweight Reasoning Projector module, trained to produce continuous latent tokens that help the model 'skip' reasoning steps. During RL, the policy model decides when to activate the projector, switching between latent and discrete reasoning as needed. Experimental results on plot hole detection and book chapter generation show that our method outperforms latent reasoning baselines and comes close to matching non-latent RL training, while reducing final reasoning length by 77-92%. Overall, LiteReason guides RL training to a more efficient part of the performance-computation tradeoff curve.

Operator-learning systems are not governed solely by total parameter count; for one query, the relevant bottleneck can be the model that must be loaded and evaluated. We study this distinction for classical neural operators on compact Sobolev subsets through a constructive comparison between routed mixtures of neural operators (MoNOs) and a fixed single-neural-operator construction. The comparison concerns expert-active complexity relative to that baseline, with total stored size and routing search accounted separately. A MoNO routes each input function through a tree to one expert. Our main theorem shows that every scalar uniformly continuous nonlinear operator with bounded output Sobolev radius on the approximation set admits a MoNO approximation whose active expert has smaller depth, width, and rank scaling than the analyzed single-neural-operator construction; for Lipschitz targets these expert quantities are bounded by $\mathcal{O}(\varepsilon^{-1})$. The theorem turns localization into an operator-level accounting of active expert size, routing depth, and number of experts. We also prove a quantitative universal approximation theorem for the underlying neural-operator architecture, with explicit dependence on compact-set diameter and modulus of continuity.

While promising, graph reasoners based on Large Language Models (LLMs) lack built-in invariance to symmetries in graph representations. Operating on sequential graph serializations, LLMs can produce different outputs under node reindexing, edge reordering, or formatting changes, raising robustness concerns. We systematically analyze these effects, studying how fine-tuning impacts encoding sensitivity as well generalization on unseen tasks. We propose a principled decomposition of graph serializations into node labeling, edge encoding, and syntax, and evaluate LLM robustness to variations of each of these factors on a comprehensive benchmarking suite. We also contribute a novel set of spectral tasks to further assess generalization abilities of fine-tuned reasoners. Results show that larger (non-fine-tuned) models are more robust. Fine-tuning reduces sensitivity to node relabeling but may increase it to variations in structure and format, while it does not consistently improve performance on unseen tasks.

Passive acoustic monitoring offers the potential to enable long-term, spatially extensive assessments of coral reefs. To explore this approach, we deployed underwater acoustic recorders at ten coral reef sites around Singapore waters over two years. To mitigate the persistent anthropogenic and current-induced noise masking the low-frequency reef soundscape, we trained a convolutional neural network denoiser. Analysis of the acoustic data reveals distinct morning and evening choruses. Though the correlation with environmental variates was obscured in the low-frequency part of the noisy recordings, the denoised data showed correlations of acoustic activity indices such as sound pressure level and acoustic complexity index with diver-based assessments of reef health such as live coral richness and cover, and algal cover. Furthermore, the shrimp snap rate, computed from the high-frequency acoustic band, is robustly correlated with the reef parameters, both temporally and spatially. This study demonstrates that passive acoustics holds valuable information that can help with reef monitoring, provided the data is effectively denoised and interpreted. This methodology can be extended to other marine environments where acoustic monitoring is hindered by persistent noise.

Large language models (LLMs) are often queried multiple times at test time, with predictions aggregated by majority vote. While effective, this self-consistency (Wang et al., 2023) strategy requires a fixed number of calls and fails when the correct answer is infrequent. We introduce Confidence-Guided Early Stopping (CGES), a Bayesian framework that forms posteriors over candidate answers and adaptively halts sampling once one answer accumulates enough posterior mass. We prove guarantees in both an ideal calibrated regime and a realistic noisy-confidence regime under a directional drift condition. Averaged over five reasoning benchmarks, CGES reduces the average number of calls by 58% on average (from 16.0 to 6.7) while matching its accuracy within 0.4 percentage points of self-consistency.

While benchmarks measure the accuracy of LLM-powered agents, they mostly assume full automation, failing to represent the collaborative nature of real-world use cases. In this paper, we make two major steps towards the rigorous assessment of human-agent interactions. First, we propose PULSE, a framework for more efficient human-centric evaluation of agent designs, which comprises collecting user feedback, training an ML model to predict user satisfaction, and computing results by combining human satisfaction ratings with model-generated pseudo-labels. Second, we deploy PULSE in software engineering -- one of the highest-impact, real-world domains for LLM agents -- via a large-scale web platform built around the open-source agent OpenHands. Across 15k users, we evaluate how three agent design decisions impact developer satisfaction rates. We also show how PULSE can lead to more robust conclusions about agent design, reducing confidence intervals by 40\% compared to a standard A/B test. Finally, we find substantial discrepancies between in-the-wild results with benchmark performance (e.g., the anti-correlation between claude-sonnet-4 and gpt-5), underscoring the limitations of benchmark-driven evaluation. Our framework PULSE provides guidance for future evaluations, and our findings identify opportunities for better software agent designs.

Solving large-scale Generalized Eigenvalue Problems (GEPs) is a fundamental yet computationally prohibitive task in science and engineering. As a promising direction, contour integral (CI) methods offer an efficient and parallelizable framework. However, their performance is critically dependent on the selection of integration contours -- improper selection without reliable prior knowledge of eigenvalue distribution can incur significant computational overhead and compromise numerical accuracy. To address this challenge, we propose Deepcontour, a novel hybrid framework that integrates a deep learning-based spectral predictor with Kernel Density Estimation (KDE) for principled contour design. Specifically, Deepcontour utilizes its specialized Eigen-Neural-Operator (ENO) to provide rapid spectral distribution priors, driving a KDE module to automatically construct the optimized integration contours, which guide the CI solver to efficiently find the desired eigenvalues. Deepcontour achieves up to a 5.63x speedup across diverse scientific datasets while maintaining strict numerical rigor. By merging the predictive power of deep learning with the numerical rigor of classical solvers, this work establishes an efficient and robust paradigm for solving large-scale GEPs.

Type 1 Diabetes (T1D) management is a complex task due to many variability factors. Artificial Pancreas (AP) systems have alleviated patient burden by automating insulin delivery through advanced control algorithms. However, the effectiveness of these systems depends on accurate modeling of glucose-insulin dynamics, which traditional mathematical models often fail to capture due to their inability to adapt to patient-specific variations. This study introduces a Biological-Informed Recurrent Neural Network (BIRNN) framework to address these limitations. The BIRNN leverages a Gated Recurrent Units (GRU) architecture augmented with physics-informed loss functions that embed physiological constraints, ensuring a balance between predictive accuracy and consistency with biological principles. The framework is validated using the commercial UVA/Padova simulator, outperforming traditional linear models in glucose prediction accuracy and reconstruction of unmeasured states, even under circadian variations in insulin sensitivity. The results demonstrate the potential of BIRNN for personalized glucose regulation and future adaptive control strategies in AP systems.

Musculoskeletal disorders pose significant risks to athletes, and early risk assessment is essential for prevention. However, most existing methods are designed for controlled settings and fail to reliably assess risk in complex environments due to their reliance on a single type of data. This research introduces ViSK-GAT (Visual-Skeletal Geometric Attention Transformer), a novel multimodal deep learning framework that classifies musculoskeletal risk using both visual and skeletal coordinate-based features. A custom multimodal dataset (MusDis-Sports) was created by combining images and skeletal coordinates, with each sample labeled into eight risk categories based on the Rapid Entire Body Assessment (REBA) system. ViSK-GAT integrates two innovative modules: the Fine-Grained Attention Module (FGAM), which refines intra-modal features through self-attention before fusion, and the Multimodal Geometric Correspondence Module (MGCM), which enhances cross-modal alignment between image features and coordinates. The model achieved robust performance, with all key metrics exceeding 93%. Probability distribution error metrics also showed a low Root Mean Squared Error (RMSE) of 0.1205 and a Mean Absolute Error (MAE) of 0.0156. ViSK-GAT consistently outperformed state-of-the-art (SOTA) deep learning backbones and showed its potential to advance artificial intelligence-driven musculoskeletal risk assessment and enable timely interventions in sports.

Graph Neural Networks (GNNs) have achieved notable success in tasks such as social and transportation networks. However, recent studies have highlighted the vulnerability of GNNs to backdoor attacks, raising significant concerns about their reliability in real-world applications. Despite initial efforts to defend against specific graph backdoor attacks, existing defense methods face two main challenges: either the inability to establish a clear distinction between triggers and clean nodes, resulting in the removal of many clean nodes, or the failure to eliminate the impact of triggers, making it challenging to restore the target nodes to their pre-attack state. Through empirical analysis of various existing graph backdoor attacks, we observe that the triggers generated by these methods exhibit over-similarity in both features and structure. Based on this observation, we propose a novel graph backdoor defense method SimGuard. We first utilizes a similarity-based metric to detect triggers and then employs contrastive learning to train a backdoor detector that generates embeddings capable of separating triggers from clean nodes, thereby improving detection efficiency. Extensive experiments conducted on real-world datasets demonstrate that our proposed method effectively defends against various graph backdoor attacks while preserving performance on clean nodes. The code will be released upon acceptance.

Current training data attribution (TDA) methods treat the influence one sample has on another as static, but neural networks learn in distinct stages that exhibit changing patterns of influence. In this work, we introduce a framework for stagewise data attribution grounded in singular learning theory. We predict that influence can change non-monotonically, including sign flips and sharp peaks at developmental transitions. We first validate these predictions analytically and empirically in a toy model, showing that dynamic shifts in influence directly map to the model's progressive learning of a semantic hierarchy. Finally, we demonstrate these phenomena at scale in language models, where token-level influence changes align with known developmental stages.

Software requirements are derived from a variety of elicitation techniques, many of which have a conversational nature, like interviews. However, evaluating whether those derived requirements faithfully reflect the stakeholders' needs remains a challenging manual task. In this paper, we formalize the task of aligning the transcript of an interview with a collection of requirements represented as user stories. We propose two heuristic metrics for alignment, called (i) requirements faithfulness: the proportion of stories supported by the transcript, and (ii) interview coverage: the proportion of transcript supported by at least one story. Then, we run experiments with large language models and embedding models that assess the ability of evaluating these metrics automatically. Experiments over four datasets show that an LLM-based solution achieves 0.86 macro-F1 on manually labeled chunk-story pairs. We also show how embedding models can be used as blockers to make the approach more scalable. This work paves the way for more research on linking conversational artifacts with requirements. The formal framework and the automated matching techniques are basic components that can be used for emerging tasks such as tracing requirements to interviews and generating requirements from conversations.

While automatic metrics drive progress in Machine Translation (MT) and Text Summarization (TS), existing metrics have been developed and validated almost exclusively for English and other high-resource languages. This narrow focus leaves Indian languages, spoken by over 1.5 billion people, largely overlooked, casting doubt on the universality of current evaluation practices. To address this gap, we introduce ITEM, a large-scale benchmark that systematically evaluates the alignment of 29 automatic metrics with human judgments across six major Indian languages, enriched with fine-grained annotations. Our extensive evaluation, covering agreement with human judgments, sensitivity to outliers, language-specific reliability, inter-metric correlations, and resilience to controlled perturbations reveals four central findings: (1) LLM-based evaluators show the strongest alignment with human judgments at both segment and system levels; (2) outliers exert a significant impact on metric-human agreement; (3) In TS, metrics are more effective at capturing content fidelity, whereas in MT, they better reflect fluency; and (4) Metrics differ in their robustness and sensitivity when subjected to diverse perturbations. Collectively, these findings offer critical guidance for advancing metric design and evaluation in Indian languages.

Given a map description through global traversal navigation instructions, an LLM can often infer the implicit spatial layout and answer user queries by providing shortest paths. However, such context-dependent querying becomes incapable as environments grow larger, motivating the need for incremental map construction that builds a complete topological graph from stepwise observations. We propose LLM-MapRepair, a framework for LLM-driven construction and map repair, designed to detect, localize, and correct structural inconsistencies in incrementally constructed navigation graphs. Our contributions include a Version Control mechanism for graph construction, an Edge Impact Score for repair prioritization, and a cleaned variant of the MANGO benchmark tailored for LLM-driven map construction and repair. We evaluate the framework on four evaluation settings: a synthetic per-component ablation (gpt-4.1, n=20 seeds per cell), a cross-vendor sweep over seven LLMs from OpenAI, Anthropic, and Google on both synthetic and TextWorld procedurally-generated text-adventure games, a repair-stage evaluation on all 42 cleaned-MANGO games with non-zero residual conflicts (534 conflicts; three vendors x three modes plus two non-LLM references), and an end-to-end natural-text deployment on Chapters 16-17 of Dream of the Red Chamber. On the DRC deployment, LLM-MapRepair achieves 94.3% node recall (+8.6 pp over direct LLM mapping) and 88.2% edge recall (+55.8 pp), using GPT-4.1; the recall improvements come with predicted node and edge counts that are roughly 4x the ground-truth counts (Table 4), reflecting the discretization-driven over-generation trade-off we discuss in the Limitations.

Prompt-based continual learning (CL) offers a parameter-efficient way to adapt large language models (LLMs) across task sequences. However, existing methods often rely on task-aware inference and maintain an expanding set of task-specific prompts, leading to (1) severe performance degradation on earlier tasks when task identifiers are unavailable for prompt selection at inference time, and (2) limited scalability as task sequence grows. We propose GRID, a unified framework designed to address these challenges. GRID incorporates an output-space-aware decoding mechanism that enhances backward transfer by leveraging representative inputs and automatic label semantic normalization, alongside a gradient-guided prompt selection strategy that compresses less informative prompts into a single aggregated representation for scalable, memory-efficient continual learning. Extensive experiments on long-sequence and negative-transfer benchmarks show that GRID improves backward transfer, achieves competitive forward transfer, and substantially reduces prompt memory across encoder-decoder and decoder-only architectures, including T5, Qwen, and LLaMA. Source code is available at https://github.com/AnushkaTi/GRID.

While large language models (LLMs) have demonstrated strong performance on factoid question answering, they are still prone to hallucination and untruthful responses, particularly when tasks demand information outside their parametric knowledge. Indeed, truthfulness requires more than accuracy -- models must also recognize uncertainty and abstain when unsure to avoid hallucinations. This presents a fundamental challenge for existing methods: approaches that optimize for accuracy often amplify hallucinations, while those that encourage abstention can become overly conservative, sacrificing correct answers. Both extremes ultimately compromise truthfulness. In this work, we present TruthRL, a general reinforcement learning (RL) framework that directly optimizes the truthfulness of LLMs. Specifically, we implement TruthRL using GRPO with a simple yet effective ternary reward that distinguishes correct answers, hallucinations, and abstentions. It incentivizes models to reduce hallucinations not only by providing correct responses, but also by enabling abstention when uncertain, thereby improving truthfulness. Extensive experiments across four knowledge-intensive benchmarks show that TruthRL significantly reduces hallucinations (e.g., 43.5% $\rightarrow$ 19.4%) and improves truthfulness (e.g., 5.3% $\rightarrow$ 37.2%), with consistent gains across various backbone models. Analysis shows that the improvement of TruthRL arises from enhanced capability of LLMs to recognize their knowledge boundary, hence avoiding being overly conservative as the baselines are.

Wildfires are among the most severe natural hazards, posing a significant threat to both humans and natural ecosystems. The growing risk of wildfires increases the demand for forecasting models that are not only accurate but also reliable. Deep Learning (DL) has shown promise in predicting wildfire danger; however, its adoption is hindered by concerns over the reliability of its predictions, some of which stem from the lack of uncertainty quantification. To address this challenge, we present an uncertainty-aware DL framework that jointly captures epistemic (model) and aleatoric (data) uncertainty to enhance short-term wildfire danger forecasting. In the next-day forecasting, our best-performing model improves the F1 Score by 2.3% and reduces the Expected Calibration Error by 2.1% compared to a deterministic baseline, enhancing both predictive skill and calibration. Our experiments confirm the reliability of the uncertainty estimates and illustrate their practical utility for decision support, including the identification of uncertainty thresholds for rejecting low-confidence predictions and the generation of well-calibrated wildfire danger maps with accompanying uncertainty layers. Extending the forecast horizon up to ten days, we observe that aleatoric uncertainty increases with time, showing greater variability in environmental conditions, while epistemic uncertainty remains stable. Finally, we show that although the two uncertainty types may be redundant in low-uncertainty cases, they provide complementary insights under more challenging conditions, underscoring the value of their joint modeling for robust wildfire danger prediction. In summary, our approach significantly improves the accuracy and reliability of wildfire danger forecasting, advancing the development of trustworthy wildfire DL systems.

Reconstructing accurate implicit surface representations from point clouds remains a challenging task, particularly when data is captured using low-quality scanning devices. These point clouds often contain substantial noise, leading to inaccurate surface reconstructions. Inspired by the Noise2Noise paradigm for 2D images, we introduce NoiseSDF2NoiseSDF, a novel method designed to extend this concept to 3D neural fields. Our approach enables learning clean neural SDFs from noisy point clouds through noisy supervision by minimizing the MSE loss between noisy SDF representations, allowing the network to implicitly denoise and refine surface estimations. We evaluate the effectiveness of NoiseSDF2NoiseSDF on benchmarks, including the ShapeNet, ABC, Famous, and Real datasets. Experimental results demonstrate that our framework significantly improves surface reconstruction quality from noisy inputs.

Despite considerable progress in the development of machine-text detectors, the ease with which machine-text can be manipulated to evade detection has led to suggestions that the problem is inherently intractable. In this work, we investigate the limits of such evasion strategies. We demonstrate that while current attacks, ranging from prompt engineering to detector-guided optimization can effectively degrade performance of standard detectors, they fail to erase the underlying stylistic "fingerprints" of machine text. We show that few-shot detectors that utilize the stylistic feature space are robust to these evasion attempts, reliably detecting samples even from models explicitly tuned to prevent detection. This raises the question: does style represent a universal defense against machine-detection attacks? We demonstrate that the answer is "no'' by introducing a novel paraphrasing approach that simultaneously optimizes for undetectability and adherence to specific human styles. We show that unlike prior methods, this attack effectively evades all considered detectors, including those that utilize writing style. However, we find that this evasion is not absolute: as the number of documents available for analysis grows, the human and machine distributions become distinguishable again. Overall, our findings suggest that reliable machine-text detection requires moving beyond single-document analysis to multi-document analysis.

Test-time scaling, primarily manifested through multi-step Chain-of-Thought (CoT) reasoning via Reinforcement Learning (RL), has emerged as a pivotal paradigm for enhancing the reasoning capabilities of Large Language Models (LLMs). However, a significant theoretical gap persists: traditional token-level analysis fails to capture the macroscopic dynamics of reasoning-level scaling. To address this, we introduce CoT-Space, a novel theoretical framework that recasts the reasoning process from a discrete token-prediction task to an optimization process within a continuous, reasoning-level semantic space. By modeling the reasoning trajectory from both noise and risk perspectives and revitalizing foundational principles from classical learning theory, we demonstrate that the observed convergence to an optimal CoT length is a natural consequence of the fundamental trade-off between underfitting and overfitting. We further utilize RL as a tool to elicit and verify these results in our experiments. Our findings provide a mechanistic explanation for the internal test-time scaling via RL, offering a principled theoretical foundation to optimize reasoning trajectories in modern LLMs.

Human gaze estimation is essential for applications such as human-computer interaction, social robotics, and assistive systems. However, achieving accurate, interpretable, and real-time performance in unconstrained environments remains challenging. Existing appearance-based methods often face trade-offs between spatial robustness, computational efficiency, and effective use of contextual information. To address this, we introduce CapStARE, a capsule-based architecture that combines a frozen ConvNeXt backbone for efficient feature extraction, capsule formation with attention-based routing for structured facial reasoning, and dual GRU decoders for lightweight sequential modeling over short-horizon observation windows. This design preserves interpretable part-whole facial relationships while improving prediction stability through local contextual consistency. Experimental results demonstrate strong performance on ETH-XGaze (3.36) and MPIIFaceGaze (2.65), while also generalizing competitively on Gaze360 (9.06), all with real-time inference (<10 ms). These findings suggest that the proposed method provides a practical and robust framework for appearance-based gaze estimation in real-world interactive environments. The related code and experimental results are publicly available at: https://github.com/toukapy/capsStare

Using diffusion transformers for media generation may require evaluating attention over extremely long sequences, with attention layers accounting for the majority of generation latency. Exploiting sparsity in attention maps offers a promising opportunity to reduce this cost. In this work, we show that attention maps in diffusion transformers exhibit significant fine-grained sparsity in video generation models. Existing sparse attention methods, however, are too coarse-grained, leaving a large fraction of redundant computation unaddressed, or incur high overheads at finer granularity. We propose FG-Attn, a novel, low-overhead fine-grained sparse attention mechanism that skips score computations at the granularity of a MxN tile, where N>=1 and M>=16, and where each block is the result of query-key dot products between M queries and N keys. FG-Attn addresses the key challenge of hardware underutilization in sparse attention kernels on GPUs, without incurring the overheads of irregular memory access and redundant operations. FG-Attn can fully supersede existing sparse attention methods and extend block sparse attention methods to finer granularities on modern GPUs. At 70% sparsity, FG-Attn is up to 2.45X faster than the state-of-art FlashInfer, and reduces attention kernel time by 14.7% on average. FG-Attn speeds up end-to-end video generation times by up to 1.40X (1.18X on average) over Flash Attention 3.

Generalist robots should be able to understand and follow user instructions. Despite providing a powerful architecture for mapping open-vocabulary language instructions to robot actions, current vision-language-action (VLA) models struggle to follow fine-grained commands. One cause for this is a lack of semantic diversity and language grounding in existing robot datasets and, specifically, a lack of fine-grained task diversity for similar observations. To address this, we present a novel method to augment existing robot datasets by leveraging vision-language models to create counterfactual labels. By augmenting existing datasets with these labels, we increase the diversity and granularity of language grounding for robot datasets, ultimately improving the language-following capabilities of VLAs. We evaluate the resulting model's ability to follow language instructions, ranging from simple object-centric commands to complex referential tasks, by conducting vision-language navigation experiments in 3 different indoor and outdoor environments. Our experiments show that counterfactual relabeling (without additional data collection) significantly improves instruction-following in VLA policies, outperforming state-of-the-art methods and doubling the success rate compared to VLAs trained on unaugmented data. We also evaluate our method for manipulation VLAs and find a similar gain in performance on tasks with distractors.

Conformal prediction (CP) is well-suited for uncertainty quantification in time series forecasting due to its distribution-free coverage guarantees. However, existing multi-step methods often struggle to balance coverage validity with efficiency: they either calibrate horizons independently, ignoring temporal correlations, or enforce strict simultaneous coverage, resulting in overly conservative intervals. In this work, we propose O2CP: Optimization-based Online Conformal Prediction, a unified framework for online conformal prediction that explicitly models multi-step error dependencies without sacrificing long-term marginal coverage guarantees. We first prove that standard online conformal updates maintain validity as long as calibration parameters remain within a defined "safe" region. Leveraging this theoretical insight, we introduce a two-layer architecture: an outer layer that defines admissible parameter sets to ensure validity, and an inner layer that performs constrained optimization to model joint error distributions and minimize horizon-wide objectives. To make this computationally feasible, we develop a lightweight sampling strategy that estimates joint distributions without requiring large calibration sets. Extensive experiments on real-world datasets, including autonomous driving, climate forecasting, and public health, demonstrate that O2CP consistently outperforms state-of-the-art baselines, achieving target coverage with significantly sharper prediction intervals and reduced regret over long horizons.

Large language models (LLMs) have created new opportunities to assist teachers and support student learning. While researchers have explored various prompt engineering approaches in educational contexts, the degree to which these approaches generalize across domains--such as science, computing, and engineering--remains underexplored. In this paper, we introduce Chain-of-Thought Prompting + Active Learning (CoTAL), an LLM-based approach to formative assessment scoring that (1) leverages Evidence-Centered Design (ECD) to align assessments and rubrics with curriculum goals, (2) applies human-in-the-loop prompt engineering to automate response scoring, and (3) incorporates chain-of-thought (CoT) prompting and teacher and student feedback to iteratively refine questions, rubrics, and LLM prompts. Our findings demonstrate that CoTAL improves GPT-4's scoring performance across domains, achieving gains of up to 38.9% over a non-prompt-engineered baseline (i.e., without labeled examples, chain-of-thought prompting, or iterative refinement). Teachers and students judge CoTAL to be effective at scoring and explaining responses, and their feedback produces valuable insights that enhance grading accuracy and explanation quality.

Autoregressive (AR) encoder-decoder models dominate high-quality multilingual ASR, but their left-to-right decoders make inference latency scale with transcript length. A natural alternative, CTC-style non-autoregressive (NAR) systems avoid this bottleneck but their conditional independence assumption sacrifices transcript-level generative modeling. Masked diffusion language models (e.g., LLaDA, MDLM) offer a competitive NAR text-generation approach. We ask whether such models can bring NAR ASR into the accuracy regime of strong AR ASR systems while removing the left-to-right bottleneck. We propose Whisfusion, which trains a dedicated masked diffusion decoder from scratch on top of frozen Whisper-large-v3 audio embeddings, denoising masked transcripts in just a few steps. We train on ~68k hours of 11-language speech with high-mask specialization to align training with the fully masked starting point of inference, and decode via Parallel Diffusion Decoding. Whisfusion surpasses Whisper-large-v3 on group-average accuracy across English, European, and CJK benchmarks, while running 4-5x faster, additionally surpassing Whisper-turbo in both accuracy and throughput. It reaches accuracy competitive with Canary and Qwen3-ASR while running 3-7x faster. These results establish masked diffusion as a Pareto-competitive non-autoregressive paradigm for high-throughput multilingual transcription. Code and model weights are available at https://github.com/taeyoun811/Whisfusion.

In time series forecasting, covariates represent external factors that influence target variables. Some covariates are observable only in the past (observed covariates, such as recorded weather data), while others are known in advance (known covariates, such as calendar events or discount schedules). Although covariates have the potential to enhance forecasting performance, most deep learning-based forecasting models struggle to address the length discrepancy between variables caused by the future portion of known covariates and fail to leverage them flexibly. Moreover, capturing dependencies between target variables and covariates is non-trivial, as models must accurately reflect the local impact of covariates while simultaneously modeling global cross-variate dependencies. To address these challenges, we propose CITRAS, a decoder-only Transformer that flexibly integrates multiple target variables, observed covariates, and known covariates. While preserving strong autoregressive modeling capabilities, CITRAS introduces two novel mechanisms within patch-wise cross-variate attention: Key-Value (KV) Shift and Attention Score Smoothing. KV Shift seamlessly incorporates the future portion of known covariates into the forecasting process by aligning them with target variables based on their concurrent dependencies. Attention Score Smoothing refines locally accurate patch-wise cross-variate dependencies into global variate-level dependencies by smoothing the historical attention scores. Experimentally, CITRAS demonstrates strong performance across a wide range of real-world datasets in both covariate-informed and multivariate settings, showcasing its versatile ability to leverage cross-variate and cross-time dependencies for improved forecasting accuracy.

Audio-visual Target Speaker Extraction (AV-TSE) aims to isolate a target speaker's voice from multi-speaker environments by leveraging visual cues as guidance. However, the performance of AV-TSE systems heavily relies on the quality of these visual cues. In extreme scenarios where visual cues are missing or severely degraded, the system may fail to accurately extract the target speaker. In contrast, humans can maintain attention on a target speaker even in the absence of explicit auxiliary information. Motivated by such human cognitive ability, we propose a novel framework called MeMo, which incorporates two adaptive memory banks to store attention-related information. MeMo is specifically designed for real-time scenarios: once initial attention is established, the system maintains attentional momentum over time, even when visual cues become unavailable. We conduct comprehensive experiments to verify the effectiveness of MeMo. Experimental results demonstrate that our proposed framework achieves SI-SNR improvements of at least 2 dB over the corresponding baseline.